Chips 2020

THE FRONTIERS COLLECTION
THE FRONTIERS COLLECTION

Series Editors:
A.C. Elitzur L. Mersini-Houghton M. Schlosshauer M.P. Silverman R. Vaas H.D. Zeh
J. Tuszynski
The books in this collection are devoted to challenging and open problems at the forefront of
modern science, including related philosophical debates. In contrast to typical research
monographs, however, they strive to present their topics in a manner accessible also
to scientifically literate non-specialists wishing to gain insight into the deeper implications
and fascinating questions involved. Taken as a whole, the series reflects the need for a
fundamental and interdisciplinary approach to modern science. Furthermore, it is intended to
encourage active scientists in all areas to ponder over important and perhaps controversial
issues beyond their own speciality. Extending from quantum physics and relativity to
entropy, consciousness and complex systems – the Frontiers Collection will inspire readers
to push back the frontiers of their own knowledge.
For a full list of published titles, please see back of book or springer.com/series 5342
Series home page – springer.com

Bernd Hoefflinger
Editor
Chips 2020
A Guide to the Future of Nanoelectronics
Editor
Bernd Hoefflinger
Leonberger Str. 5
71063 Sindelfingen
Germany
Series Editors:
Avshalom C. Elitzur
Bar-Ilan University, Unit of Interdisciplinary Studies, 52900 Ramat-Gan, Israel
email: avshalom.elitzur@weizmann.ac.il
Laura Mersini-Houghton
Dept. Physics, University of North Carolina, Chapel Hill, NC 27599-3255, USA
email: mersini@physics.unc.edu
Maximilian A. Schlosshauer
Institute for Quantum Optics and Quantum Information, Austrian Academy of Sciences,
Boltzmanngasse 3, A-1090 Vienna, Austria
email: schlosshauer@nbi.dk
Mark P. Silverman
Trinity College, Dept. Physics, Hartford CT 06106, USA
email: mark.silverman@trincoll.edu
Jack A. Tuszynski
University of Alberta, Dept. Physics, Edmonton AB T6G 1Z2, Canada
email: jtus@phys.ualberta.ca
Rüdiger Vaas
University of Giessen, Center for Philosophy and Foundations of Science, 35394 Giessen,
Germany
email: ruediger.vaas@t-online.de
H. Dieter Zeh
Gaiberger Straße 38, 69151 Waldhilsbach, Germany
email: zeh@uni-heidelberg.de
ISSN 1612-3018
ISBN 978-3-642-22399-0 e-ISBN 978-3-642-23096-7
DOI 10.1007/978-3-642-23096-7
Springer Heidelberg Dordrecht London New York
Library of Congress Control Number: 2011943001
# Springer-Verlag Berlin Heidelberg 2012

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from Springer. Violations
are liable to prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply,
even in the absence of a specific statement, that such names are exempt from the relevant protective
laws and regulations and therefore free for general use.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)

Preface
On July 30, 1959, Robert Noyce filed his Integrated Circuit patent. Its fiftieth
anniversary in 2009 became the origin of this book and its central question: After
the unparalleled progress of microchips over half a century, can this story continue,
and what will it take to make this happen?
To greet the year 2000, I had written an invited paper “Chips 2020” for the
German magazine Elektronik (with a strong vote for 3D chip integration), and it
looked attractive to check the status at half-time, in 2010. However, the central
issue of this book emerged owing to more and more signs of the imminent end of
the nano(meter) roadmap: The law that halving the transistor size every 18 months
would bring automatic performance and market growth is about to end in 2015.
When the billions of 10 nm10 nm transistors packed side-by-side on a chip are
hardly useful because of their fundamental statistical variance, we face the most
important turning point in the history of microelectronics: Declining growth in
markets and services and an energy crisis on top, because, with the chip technology
of 2010 and with the present annual doubling of video and TV on the internet, this
service alone would require the total worldwide electrical power in 2015.
Chips 2020 explains the background to the 20–10 nm transistor limits in
different applications, and it focuses on the new strategy for the sustainable growth
of a nanoelectronics ecosystem with a focus on ultra-low energy of all chip
functionalities, femtojoule electronics, enabled by 3D integration on-chip and of
chip-systems incorporating new architectures as well as new lithography and
silicon technologies.
At the critical time of this new fundamental energy orientation, I have been very
fortunate that several world leaders with their teams agreed to contribute to this
book: Greg Delagi of Texas Instruments, on intelligent mobile companions, the
new lead products; Georg Kimmich of ST-Ericsson, on 3D integration for wireless
applications; Burn Lin of TSMC, on nanolithography; Jiri Marek and Udo Gomez
of Robert Bosch GmbH, on MEMS (micro-electro-mechanical systems) for auto-
motive and consumer applications; Barry Pangrle of Mentor Graphics, on power-
efficient design; Peter Roth and colleagues at IBM, on superprocessors; Yiannos
Manoli with his team at the University of Freiburg, together with Boris Murmann of
Stanford University, on analog–digital interfaces and energy harvesting; Albrecht
v
vi Preface
Rothermel of the University of Ulm, on retinal implants for blind patients; and Ben
Spaanenburg with his co-author at Lund University, on digital neural networks.
This book would be unthinkable without their contributions, and, in this critical
situation for nanoelectronics, I am even more obliged to them for their involvement
in this project.
Despite our broad agenda, we could cover only selected topics, which, we hope,
are exemplary for the potential and challenges for 2020 chips.
My sincere thanks go to Claus Ascheron at Springer in Heidelberg, who con-
stantly pursued the idea for this book and who finally convinced me in 2009 to
realize it. I thank the team at Springer Science+Business Media for their confidence
in this publication and for its professional production. I thank Stefanie Krug for her
perfect execution of many of the illustrations and Deborah Marik for her profes-
sional editing of the manuscript.
With experience as the first MOS product manager at Siemens and of
co-founding a Technical University and establishing several microchip research
facilities in the USA and Germany, I have included educational, research, and
business aspects of the nanoelectronics ecosystem. I hope that, with this scope, this
book will be helpful to all those who have to make decisions associated with future
electronics, from students to graduates, educators, and researchers, as well as
managers, investors, and policy makers.
Sindelfingen Bernd Hoefflinger
Contents
1 Introduction: Towards Sustainable 2020 Nanoelectronics . . . . . . . . . . . . 1

Bernd Hoefflinger
2 From Microelectronics to Nanoelectronics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

Bernd Hoefflinger
3 The Future of Eight Chip Technologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

Bernd Hoefflinger
4 Analog-Digital Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Matthias Keller, Boris Murmann, and Yiannos Manoli
5 Interconnects, Transmitters, and Receivers . . . . . . . . . . . . . . . . . . . . . . . . . . 131

Bernd Hoefflinger
6 Requirements and Markets for Nanoelectronics . . . . . . . . . . . . . . . . . . . . . 141

Bernd Hoefflinger
7 ITRS: The International Technology Roadmap

for Semiconductors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Bernd Hoefflinger
8 Nanolithography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Burn J. Lin
9 Power-Efficient Design Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

Barry Pangrle
10 Superprocessors and Supercomputers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215

Peter Hans Roth, Christian Jacobi, and Kai Weber
vii
viii Contents
11 Towards Terabit Memories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

Bernd Hoefflinger
12 3D Integration for Wireless Multimedia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251

Georg Kimmich
13 Technology for the Next-Generation-Mobile User Experience . . . . . . 273

Greg Delagi
14 MEMS (Micro-Electro-Mechanical Systems) for Automotive

and Consumer Electronics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
Jiri Marek and Udo-Martin Gómez
15 Vision Sensors and Cameras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315

Bernd Hoefflinger
16 Digital Neural Networks for New Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331

Lambert Spaanenburg and Suleyman Malki
17 Retinal Implants for Blind Patients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367

Albrecht Rothermel
18 Silicon Brains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383

Bernd Hoefflinger
19 Energy Harvesting and Chip Autonomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393

Yiannos Manoli, Thorsten Hehn, Daniel Hoffmann, Matthias Kuhl,
Niklas Lotze, Dominic Maurath, Christian Moranz, Daniel Rossbach,
and Dirk Spreemann
20 The Energy Crisis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421

Bernd Hoefflinger
21 The Extreme-Technology Industry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429

Bernd Hoefflinger
22 Education and Research for the Age of Nanoelectronics . . . . . . . . . . . . 439

Bernd Hoefflinger
23 2020 World with Chips . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451

Bernd Hoefflinger
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
Titles in this Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479

Contributors
Greg Delagi Texas Instruments, 12,500 TI Boulevard, MS 8723, Dallas, TX,

75243, USA, t-wright@ti.com
Dr. Udo-Martin Gómez Robert Bosch GmbH, Automotive Electronics (AE/

NE4), Postfach 1,342, 72703 Reutlingen, Germany, Udo-Martin.Gomez@de.
bosch.com
Thorsten Hehn IMTEK Microelectronics, University of Freiburg, Georges-

Koehler-Allee 102, 79106 Freiburg, Germany
Prof. Bernd Hoefflinger Leonberger Strasse 5, 71063 Sindelfingen, Germany,

bhoefflinger@t-online.de
Dr. Daniel Hoffmann HSG-IMIT, Institute of Micromachining and Information

Technology, Wilhelm-Schickard-Strasse 10, 78052 Villingen-Schwenningen,
Germany
Dr. Christian Jacobi IBM Systems & Technology Group, Technology Develop-
ment, Schoenaicher Strasse 220, 71032 Boeblingen, Germany, cjacobi@de.ibm.com
Dr. Matthias Keller IMTEK Microelectronics, University of Freiburg, Georges-

Koehler-Allee 102, 79110 Freiburg, Germany, mkeller@imtek.de
Georg Kimmich ST-Ericsson 12, rue Jules Horovitz, B.P. 217, 38019 Grenoble
Cedex, France, georg.kimmich@stericsson.com
Matthias Kuhl IMTEK Microelectronics, University of Freiburg, Georges-

ix
x Contributors
Dr. Burn J. Lin TSMC, 153 Kuang Fu Road, Sec. 1, Lane 89, 1st Fl, Hsinchu,
Taiwan, R.O.C. 300, burnlin@tsmc.com
Niklas Lotze HSG-IMIT, Institute of Micromachining and Information Technology,

Wilhelm-Schickard-Strasse 10, 78052 Villingen-Schwenningen, Germany
Suleiman Malki Department of Electrical & Information Technology,

Lund University, P.O. Box 118, Lund 2210, Germany
Prof. Yiannos Manoli Fritz Huettinger Chair of Microelectronics, University

of Freiburg, IMTEK, Georges-Koehler-Allee 102, 79110 Freiburg, Germany,
manoli@imtek.de
Dr. Jiri Marek Robert Bosch GmbH, Automotive Electronics, Postfach 1,342,
72703 Reutlingen, Germany, Jiri.Marek@de.bosch.com
Dominic Maurath IMTEK Microelectronics, University of Freiburg, Georges-

Koehler-Allee 102, 79110 Freiburg, Germany, maurath@imtek.de
Christian Moranz IMTEK Microelectronics, University of Freiburg, Georges-

Prof. Boris Murmann Stanford University, 420 Via Palu Mall, Allen-208,
Stanford, CA, 94305-40070, USA, murmann@stanford.edu
Dr. Barry Pangrle Mentor Graphics, 1,001 Ridder Park Drive, San Jose, CA,
95131-2314, USA, barry_pangrle@mentor.com
Daniel Rossbach HSG-IMIT, Institute of Micromachining and Information

Technology, Wilhelm-Schickard-Strasse 10, 78052 Villingen-Schwenningen,
Germany
Dr. Peter Hans Roth IBM Systems & Technology Group, Technology Develop-
ment, Schoenaicher Strasse 220, 71032 Boeblingen, Germany, peharo@de.ibm.com
Prof. Albrecht Rothermel University of Ulm, Institute of Microelectronics,

Albert-Einstein-Allee 43, 89081 Ulm, Germany, info@albrecht-rothermel.de
Prof. Lambert Spaanenburg Dept. of Electrical & Information Technology, Lund

University, PO Box 11822100 Lund, Sweden, Lambert.Spaanenburg@eit.lth.se
Dirk Spreemann HSG-IMIT, Institute of Micromachining and Information Tech-

nology, Wilhelm-Schickard-Strasse 10, 78052 Villingen-Schwenningen, Germany
Kai Weber System z Core Verification Lead, IBM Systems & Technology Group,
Technology Development, Schoenaicher Strasse 220, 71032 Boeblingen, Germany,
Kai.Weber@de.ibm.com
Biographies of Authors
Greg Delagi has been senior vice president and general manager of TI’s Wireless
Terminals Business since 2007, where he leads TI’s worldwide development of
semiconductors for mobile phones. Greg Delagi started his career with Texas
Instruments in 1984 with the Materials and Controls Business. He later joined
TI’s Semiconductor Group. Delagi was responsible for several of the company’s
business units such as DSL, Cable, VoIP and Wireless LAN, creating systems-on-
chip solutions. As vice president and manager of TI’s Digital Signal Processing
(DSP) business, he guided the company to its world leadership with a 62% share
of the world market in 2006. He is a graduate of Nichols College in Dudley, MA,
where he earned a BSBA in 1984.
xi
xii Biographies of Authors
Udo Gómez has been Director of Engineering since 2006 for Advanced Sensor
Concepts at Bosch, Automotive Electronics Division, in Reutlingen, Germany,
responsible for MEMS sensor predevelopment activities. He also serves as chief
expert for MEMS sensors. From 2003 until 2005, he was responsible for the
development of a novel automotive inertial sensor cluster platform for active safety
applications. Dr. Gómez studied physics at the University of Stuttgart, Germany. In
1997, he received his Ph.D. from the University of Stuttgart for his work on molecular
electronics. In January 1998, he joined the California Institute of Technology as a
post-doctoral fellow. In 1999, he started his work at Bosch Corporate Research.
Bernd Hoefflinger started his career as an assistant professor in the Department of

Electrical Engineering at Cornell University, Ithaca, NY, USA. Returning to Ger-
many, he served as the first MOS product manager at Siemens, Munich. With that
background, he became a co-founder of the University of Dortmund, Germany, and
later head of the Electrical Engineering Departments at the University of Minnesota
and at Purdue University, IN, USA. In 1985, he became the director of the newly
established Institute for Microelectronics Stuttgart (IMS CHIPS), a public contract
research and manufacturing institute. In 1993, IMS CHIPS became the world’s first
ISO 9000 certified research institute. He launched rapid prototyping with electron-
beam lithography in 1989. He established the institute as a leader in high-dynamic-
range CMOS imagers and video cameras from 1993 onwards. Among the develop-
ments in CMOS photosensors was the chip design and manufacturing for the first
sub-retinal implants in humans in Europe in 2005. He retired in 2006.
Biographies of Authors xiii
Christian Jacobi received both his M.S. and Ph.D. degrees in computer science
from Saarland University, Germany, in 1999 and 2002, respectively. He joined
IBM Research and Development in 2002, working on floating-point implementa-
tion for various IBM processors. For IBM’s System z10 mainframe, he worked on
the L1.5 cache unit. From 2007 to 2010, Dr. Jacobi was on international assignment
to Poughkeepsie, NY, where he worked on the Load-Store-Unit for the IBM
zEnterprise 196 mainframe. Dr. Jacobi is now responsible for the cache design of
IBM’s mainframe processors, and for future architecture extensions.
Matthias Keller was born in Saarlouis, Germany, in 1976. He received his

Diploma degree in electrical engineering from the University of Saarland, Saar-
brücken, Germany, in 2003 and a Dr.-Ing. degree (summa cum laude) from the
University of Freiburg, Germany, in 2010. From 2003 to 2009, he was a research
assistant at the Fritz Huettinger Chair of Microelectronics at the University of
Freiburg, Germany, in the field of analog CMOS integrated circuit design with an
emphasis on continuous-time delta-sigma A/D converters. In 2009, he was awarded
a tenured position as a research associate (“Akademischer Rat”). His research
interests are analog integrated circuits based on CMOS technology, in particular
delta-sigma A/D converters and patch-clamp readout circuits.
xiv Biographies of Authors
Georg Kimmich is product manager at ST-Ericsson in Grenoble, France. He is

responsible for the product definition of digital baseband system-on-chip products
within the 3G Multimedia and Platform Division. His focus is on the technology
roadmap definition for chip packaging. Georg Kimmich graduated in computer
engineering at HFU, Furtwangen, Germany in 1990. He held several design and
management positions in the design automation and ASIC design domain with
Thomson Multimedia and STMicroelectronics in Germany, France and the US. In
2006, he joined the application processor division of STMicroelectronics, where he
was responsible for the system-on-chip development team of the Nomadik applica-
tion processor family.
Burn J. Lin became a vice president at TSMC, Ltd., in 2011. He joined TSMC
in 2000 as a senior director. Prior to that, he founded Linnovation, Inc. in 1992.
Earlier he held various technical and managerial positions at IBM after joining IBM
in 1970. He has been extending the limit of optical lithography for close to four
decades. He pioneered many lithography techniques, among them deep-UV lithog-
raphy starting from 1975, multi-layer resist from 1979, Exposure-Defocus method-
ology from 1980, k1 reduction from 1987, attenuated phase-shifting mask from
1991, 193 nm immersion lithography from 2002, and polarization-dependent stray
light from 2004. He is currently working on cost-effective optical lithography and
multiple-electron-beam mask-less lithography for the 20 nm node and beyond.
Biographies of Authors xv
Dr. Lin is the editor-in-chief of the Journal of Micro/nanolithography, MEMS, and

MOEMS, life member of the US National Academy of Engineering, IEEE Life
Fellow, and SPIE Fellow. He has received numerous awards, among them the 1st
Semi IC Outstanding Achievement Award in 2010, the 2009 IEEE Cledo Brunetti
Award, the 2009 Benjamin G. Lammé Meritorious Achievement Medal, and in
2004 the 1st SPIE Frits Zernike award. Throughout his career, he has received
numerous TSMC and IBM Awards. He has written one book and three book
chapters, published 121 articles, mostly first-authored, and holds 66 US patents.
Suleyman Malki received, in October 2008, a Ph.D. degree from the Depart-
ment of Electrical and Information Technology at Lund University for work on the
topic “Reconfigurable and Parallel Computing Structures in Structured Silicon”.
Before that he received a master’s degree in computer science and engineering from
Lund University. His research has focused mainly on the implementation and
verification of highly intelligent systems on reconfigurable chips (FPGA). Currently
he is devoting most of his time to his recent biological and industrial offspring.
Yiannos Manoli was born in Famagusta, Cyprus. As a Fulbright scholar, he

received a B.A. degree (summa cum laude) in physics and mathematics from
Lawrence University in Appleton, WI, in 1978 and a M.S. degree in electrical
engineering and computer science from the University of California, Berkeley, in
1980. He obtained a Dr.-Ing. degree in electrical engineering from the University of
xvi Biographies of Authors
Duisburg, Germany, in 1987. From 1980 to 1984, he was a Research Assistant at the
University of Dortmund, Germany. In 1985, he joined the newly founded Fraunho-
fer Institute of Microelectronic Circuits and Systems, Duisburg, Germany. From
1996 to 2001, he held the Chair of Microelectronics with the Department of
Electrical Engineering, University of Saarbrücken, Germany. In 2001, he joined
the Department of Microsystems Engineering (IMTEK) of the University of Frei-
burg, Germany, where he established the Chair of Microelectronics. Since 2005, he
has additionally served as one of the three directors at the “Institute of Microma-
chining and Information Technology” (HSG-IMIT) in Villingen-Schwenningen,
Germany. He spent sabbaticals with Motorola (now Freescale) in Phoenix, AZ, and
with Intel, Santa Clara, CA. Prof. Manoli has received numerous best paper and
teaching awards, and he has served on the committees of ISSCC, ESSCIRC, IEDM
and ICCD. He was Program Chair (2001) and General Chair (2002) of the IEEE
International Conference on Computer Design (ICCD). He is on the Senior Edito-
rial Board of the IEEE Journal on Emerging and Selected Topics in Circuits and
Systems and on the Editorial Board of the Journal of Low Power Electronics. He
served as guest editor of Transactions on VLSI in 2002 and Journal of Solid-State
Circuits in 2011.
Jiri Marek has been Senior Vice President of Engineering Sensors at Bosch,
Automotive Electronics Division, since 2003, responsible for the MEMS activities
at Bosch. He started his work at Bosch in 1986. From 1990 until 1999, he was
responsible for the Sensor Technology Center. In 1999, he became Vice President
Engineering of Restraint Systems and Sensors. Dr. Marek studied Electrical Engi-
neering at the University of Stuttgart, Germany, and Stanford University, USA. In
1983, he received his Ph.D. from the University of Stuttgart for his work at the Max
Planck Institute Stuttgart on the analysis of grain boundaries in large-grain poly-
crystalline solar cells. After a post-doctoral fellowship with IBM Research, San
José, CA, he was a development engineer with Hewlett-Packard, Optical Commu-
nication Division.
Biographies of Authors xvii
Boris Murmann is an Associate Professor in the Department of Electrical

Engineering, Stanford University, CA, USA. He received his Dipl.-Ing. (FH)
degree in communications engineering from Fachhochschule Dieburg, Germany,
in 1994, a M.S. degree in electrical engineering from Santa Clara University, CA, in
1999, and his Ph.D. degree in electrical engineering from the University of Cali-
fornia, Berkeley, in 2003. From 1994 to 1997, he was with Neutron Mikrolektronik
GmbH, Hanau, Germany, where he developed low-power and smart-power ASICs
in automotive CMOS technology. Dr. Murmann’s research interests are in the area
of mixed-signal integrated circuit design, with special emphasis on data converters
and sensor interfaces. In 2008, he was a co-recipient of the Best Student Paper
Award at the VLSI Circuit Symposium and the recipient of the Best Invited Paper
Award at the Custom Integrated Circuits Conference (CICC). In 2009, he received
the Agilent Early Career Professor Award. Dr. Murmann is a member of the
International Solid-State-Circuits Conference (ISSCC) program committee, an
associate editor of the IEEE Journal of Solid-State Circuits, and a Distinguished
Lecturer of the IEEE Solid-State Circuits Society.
Barry Pangrle is a Solutions Architect for Low Power in the Engineered

Solutions Group at Mentor Graphics Corporation. He has a B.S. in computer
engineering and a Ph.D. in computer science, both from the University of Illinois
at Urbana-Champaign. He has been a faculty member at University of California,
Santa Barbara and Penn State University, where he taught courses in computer
xviii Biographies of Authors
architecture and VLSI design while performing research in high-level design

automation. Barry has previously worked at Synopsys, initially on high-level
design tools and then later as an R&D director for power optimization and analysis
tools. He was the Director of Design Methodology for a fabless start-up company
and has also worked at a couple of privately held EDA companies, where he
focused on design automation tools for low power/energy designs. He has published
over 25 reviewed works in high level design automation and low power design and
served as a Technical Program Co-chair for the 2008 ACM/IEEE International
Symposium on Low Power Electronics Design (ISLPED). He was also actively
involved with the technical program committees for ISLPED and DAC for 2009
and was one of the General Co-Chairs for ISLPED 2010.
Peter H. Roth received his Dipl.-Ing. degree in electrical engineering and his Dr.-
Ing. degree from the University of Stuttgart, Germany, in 1979 and 1985, respectively.
In 1985, he joined the IBM Germany Research and Development Lab in Boeblingen,
starting in the department of VLSI logic-chip development. Since 1987 he has been
leading the VLSI test and characterization team of the Boeblingen lab. Later, Dr. Roth
led several development projects in the area of IBM’s Mainframe and Power micro-
processors. He also was heavily involved in the development of gaming products such
as the Cell processor for the Sony PlayStation. Today, Dr. Roth is responsible for the
hardware strategy for the IBM Germany Research and Development Laboratory.
Biographies of Authors xix
Albrecht Rothermel received his Dipl.-Ing. degree in electrical engineering

from the University of Dortmund, and his Dr.-Ing. (Ph.D.) degree from the Univer-
sity of Duisburg, both Germany, in 1984 and 1989, respectively. From 1985 to
1989, he was with the Fraunhofer Institute of Microelectronic Circuits and Systems,
Duisburg, Germany. From 1990 to 1993, he was with Thomson Consumer Elec-
tronics (Thomson Multimedia), Corporate Research, Villingen-Schwenningen,
Germany. As manager of the IC design laboratory, he was involved in analog and
mixed circuit design for audio and video. Since 1994, he has been with the Institute
of Microelectronics, University of Ulm, Germany, as a Professor of Electrical
Engineering. Dr. Rothermel was a guest scientist at Thomson Multimedia in
Indianapolis, USA (1997), at the Edith-Cowan University in Perth, Western Aus-
tralia (2003), and at the Shandong University in Jinan, P.R. China (2006). He has
published more than 100 papers, book chapters, and patents and received numerous
best-paper awards. He was an Associate Editor of the IEEE Journal of Solid-State
Circuits, and he is a member of the program committees of ISSCC, ESSCIRC, and
ICCE. He is a senior member and distinguished lecturer of the IEEE.
Lambert Spaanenburg received his master’s degree in electrical engineering

from Delft University and his doctorate in technical sciences from Twente Univer-
sity, both in The Netherlands. He started his academic journey at Twente University
in the field of VLSI design, eventually serving as CTO of ICN, the commercial
spin-off of the ESPRIT Nelsis project. At IMS in Stuttgart, he co-created the neural
xx Biographies of Authors
control for the Daimler OSCAR 1992 prototype, currently an upcoming standard
for car safety measures. He became a full professor at the University of Groningen,
The Netherlands. Further research of his group in neural image processing led in
2002 to Dacolian, which held a 60% market share for license-plate recognition
before it merged into Q-Free ASA. Overall, Dr. Spaanenburg has produced more
than 200 conference papers, 20 reviewed journal articles and seven books or
chapters. He has served annually on several program committees and journal
boards. He has been involved in many technology transfer projects, including
eight university spin-offs. Currently a guest at Lund University, he is exploiting
further research on embedded sensory networks with cloud connectivity under the
umbrella of RaviteQ AB in Sweden.
Kai Weber received his Dipl.-Ing.(BA) in information technology from the

University of Collaborative Education, Stuttgart, and a bachelor degree in computer
science from the Open University, UK, in 2003. He joined IBM Research and
Development in 2003, working on formal verification of floating-point units and
leading the floating-point unit verification teams for IBM’s POWER6 and z10
processors. Mr. Weber was the processor verification team leader for the IBM
zEnterprise 196 mainframe and continues in this role for future generations of
IBM System z processors. In addition Mr. Weber is participating in two joint
research projects with the University of Tuebingen and Saarland University eval-
uating hybrid systems and business analytic workloads.
Acronyms and Abbreviations
AAEE American Association for Engineering Education

AC Alternating current
ACM Association for Computing Machinery
ADC Analog–digital converter
AI Artificial intelligence
ALE Atomic-layer epitaxy
ALU Arithmetic logic unit
AMD Age-related macula degeneration
ANN Artificial neural network
AP Action potential
APS Active-pixel sensor
APSM Advanced porous silicon membrane
AQL Acceptable quality level
ARPA Advanced Research Projects Agency
ASIC Application-specific integrated circuit
ASIP Algorithm-specific integrated processor
ASP Application-specific processor
ASS Application-specific system
AVG Available voltage gain
AVS Adaptive voltage scaling
BAN Body-area network
BAW Bulk acoustic wave
BEOL Back-end of line
BI Backside illumination
BiCMOS Bipolar CMOS
BIST Built-in self-test
BMI Brain–machine interface
BST Boundary-scan test
C4 Controlled-collapse chip connect
CAD Computer-aided design
CAM Content-addressable memory
CARE Concerted Action for Robotics in Europe
xxi
xxii Acronyms and Abbreviations
CAT Computer-aided test

CCD Charge-coupled device
CDMA Code-division multiple access
CE Continuing education
CFC Chip-integrated fuel cell
CIFB Cascade of integrators in feedback
CIFF Cascade of integrators in feed-forward
CIS Charge-integration sensor
CISC Complex-instruction-set computer
CMOS Complementary metal–oxide–semiconductor
CNN Cellular neural network
CNT Carbon nanotube
COO Cost of ownership
CORDIC Coordinate rotation digital computer
CPU Central processing unit
CSF Contrast-sensitivity function
CT-CNN Continuous-time CNN
DA Design automation
DAC Design-Automation Conference
DAC Digital–analog converter
DARPA Defense Advanced Research Projects Agency
DBB Digital baseband
dBm Power on a log scale relative to 1 mW
DC Direct current
DCF Digital cancelation filter
DDR2 Dual data rate RAM
DEM Dynamic element matching
DFT Design for test
DIBL Drain-induced barrier lowering
DIGILOG Digital logarithmic
DLP Digital light processing
DMA Direct memory access
DN Digital number
DNN Digital neural network
DPG Digital pattern generator
DPT Double-patterning technology liquid-immersion lithography
DRAM Dynamic random-access memory
DRIE Deep reactive-ion etching
DSP Digital signal processor
DT-CNN Discrete-time CNN
DTL Diode–transistor logic
D-TLB Data-cache translation lookaside buffer
DVFS Dynamic voltage and frequency scaling
EBL Electron-beam lithography
Acronyms and Abbreviations xxiii
ECC Error-correcting code

ECG Electrocardiogram
ECL Emitter-coupled logic
EDA Electronic design automation
EEG Electroencephalography
EITO European Information Technology Organization
ELO Epitaxial lateral overgrowth
ELTRAN Epitaxial-layer transfer
EMI Electromagnetic interference
ENOB Effective number of bits
EOT Equivalent oxide thickness
ERC Engineering Research Center
ERD Emerging research devices
ERM Emerging research materials
ESD Electrostatic discharge
ESL Electronic-system level
ESP Electronic safety package
ESP Electronic Stability Program
ESPRIT European Strategic Programme for Research in Information
Technology
EUV Extreme ultraviolet
EWS Electrical wafer sort
FACETS Fast Analog Computing with Emergent Transient States
FC Fuel cell
FD SOI Fully depleted silicon-on-inulator
FED Future Electron Devices
FEOL Front-end of line
FeRAM Ferroelectric random-access memory
FET Field-effect transistor
F2 Square of minimum feature size
FFT Fast Fourier transform
FIFO First-in-first-out
FIPOS Full isolation by porous silicon
FIR Finite impulse response
FIT Failure in 107 h
FLOP Floating-point operation
FMEA Failure mode and effect analysis
FOM Figure-of-merit
FPAA Field-programmable analog array
FPGA Field-programmable gate array
fps frames per second
FR Floating-point register
FSM Finite-state machine
FSR Full signal range
xxiv Acronyms and Abbreviations
GAPU Global analogic programming unit

GBP Gain–bandwidth product
GFC Glucose fuel cell
GIPS Giga instructions per second
GOPS Giga operations per second
GPS Global Positioning System
GPU Graphics processing unit
GR General-purpose register
GSM Global System for Mobile Communication
HAC Hardware accelerator
HD High definition or high density
HDL Hardware description language
HDR High dynamic range
HDRC High-Dynamic-Range CMOS
HDTV High-definition television
HIPERLOGIC High-performance logic
HKMG High-k metal gate
HLS High-level synthesis
HVM High-volume manufacturing
IBL Ion-beam lithography
IC Integrated circuit
ICT Information and communication technology
IEDM International Electron Devices Meeting
IEEE Institute of Electrical and Electronics Engineers
IFFT Inverse fast Fourier transform
I2L Integrated injection logic
IO Input–output
IOMMU I/O memory mapping unit
IP-core Intellectual property core
IS Instruction set
ISO International Organization for Standardization
ISSCC International Solid-State Circuits Conference
IT Information technology
I-TLB Instruction-cache translation lookaside buffer
ITRS International Technology Roadmap for Semiconductors
ITS Intelligent transportation systems
JEDEC Joint Electron Device Engineering Council
JESSI Joint European Submicron Silicon Initiative
JIT compiler Just-in-time compiler
JND Joust noticeable difference
JPEG Joint Photography Expert Group
KGD Known-good die
LCA Life-cycle analysis
LDO Low-voltage dropout regulator
Acronyms and Abbreviations xxv
LOG Localized epitaxial overgrowth

LPC Linear predictive coding
LPDDR Low-power DDR
LSB Least significant bit
LSI Large-scale integration
LTE Long-term evolution
LUT Look-up table
LVDL Low-voltage differential logic
MASH Multi-stage noise shaping
MBE Molecular-beam epitaxy
MCC Manchester carry chain
MCM Multi-chip module
MCU Multi-core processing units
MEBL Multiple-electron-beam lithography
MEMS Micro-electro-mechanical system
MIPS Mega instructions per second
MITI Ministry of International Trade and Industry (Japan)
MLC Multi-level per cell
MMI Man–machine interface
MMI Machine–machine interface
MOPS Million operations per second
MOS Metal–oxide–semiconductor
MOSIS MOS IC implementation system
MPEG Motion-Picture Expert Group
MPU Microprocessor unit
MRAM Magnetic random-access memory
NA Numerical aperture
NGL Next-generation lithography
NHTSA National Highway Traffic Safety Administration
NIR Near infrared
NM Noise margin
NMOS Negative channel-charge MOS
NoC Network on chip
NTF Noise transfer function
NV Nonvolatile
OCT Optical-coherence tomography
ODM Original-device manufacturer
OECD Organisation for Economic Co-operation and Development
OECF Optoelectronic conversion function
OFDM Orthogonal frequency-division multiplexing
OpenCL Open Compute Language – Standard established by the Khronos
Group for platform independent description of highly parallel
computations
ORTC Overall roadmap technology characteristics
xxvi Acronyms and Abbreviations
OSAT Out-sourced assembly and test

OSCI Open SystemC Initiative
OSR Oversampling ratio
PAE Power-added efficiency
PC Personal computer
PCB Printed-circuit board
PCM Phase-change memory
PCS Personal Communications Service
PDN Power distribution network
PD SOI partially depleted silicon-on-inulator
PE Processing element
PEM Polymer electrolyte membrane
PLL Phase-locked loop
PMOS Positive channel-charge MOS
PMU Power management unit
POP Package-on-package
PPA Power, performance, and area
ppm Parts per million
PVDF Polyvinylidenefluoride
PZT Lead zirconate titanate
QCIF Quarter common intermediate format
QMS Quality-management system
QXGA Quad extended graphics array
RAM Random-access memory
R&D Research and development
REBL Reflective electron-beam lithography
RET Resolution-enhancement technique
RF Radio frequency
RFID Radio frequency identification
RISC Reduced-instruction-set computer
ROI Region of interest
ROM Read-only memory
ROW Rest-of-world
RP Retinitis pigmentosa
RRAM Resistive random-access memory
RTL Register-transfer level
Rx Receive
SAR Successive approximation register
SBT Strontium bismuth tantalite
SC Switch capacitor
SD Standard definition
SDR Single data rate
SECE Synchronous electric charge extraction
SEG Selective epitaxial growth
Acronyms and Abbreviations xxvii
SET Single-electron transistor

SIA Semiconductor industry association
SIMOX Silicon isolation by implanting oxygen
SiO2 Silicon dioxide
SLD System-level design
SMASH Sturdy multi-stage noise shaping
SMP Symmetric multi-processing
SMT Simultaneous multithreading
SNDR Signal-to-noise-and-distortion ratio
SNM Static noise margin
SNR Signal-to-noise ratio
SNT Silicon nanotube
SO Switch operational amplifier
SOC System-on-chip
SOI Silicon-on-insulator
SPAD Single-photon avalanche diode
SPI Serial peripheral interface
SQNR Signal-to-quantization-noise ratio
SRAM Static random-access memory
SRC Semiconductor Research Corporation
STF Signal transfer function
STL Subthreshold leakage
STT Spin–torque transfer
SyNAPSE Systems of Neuromorphic Adaptive Plastic Scalable Electronics
TAM Total available market
TFC Thin-film-on-CMOS
TLM Transfer-level modeling
TOPS Tera operations per second
TQM Total quality management
TSV Through-silicon via
TTL Transistor–transistor logic
Tx Transmit
UGBW Unity-gain bandwidth
UL Underwriters Laboratories
ULP Ultralow power
UM Universal model (CNN)
UMTS Universal Mobile Telecommunications System
VCO Voltage-controlled oscillator
VDE Verband Deutscher Elektrotechniker
VDI Verband Deutscher Ingenieure
VDC Vehicle dynamics control
VDS Vehicle dynamics system
VGA Video Graphics Array
VHDL VLSI (originally VHSIC) hardware description language
xxviii Acronyms and Abbreviations
VHSIC Very-high-speed integrated circuit

VLSI Very-large-scale integration
VRD Virtual retinal display
VSoC Vision system on chip
wph Wafers per hour
WSI Wafer-scale integration
WSTS World Semiconductor Trade Statistics
XML Extensible Markup Language
Chapter 1
Introduction: Towards Sustainable 2020
Nanoelectronics
Bernd Hoefflinger
Abstract Faced with the immanent end of the nanometer roadmap at 10 nm, and
with an electronics energy crisis, we have to engineer the largest strategy change in
the 50-years history of microelectronics, renamed to nanoelectronics in 2000 with
the first chips containing 100-nm transistors.
Accepting the 10 nm-limit, the new strategy for the future growth of chip
functionalities and markets has to deliver, within a decade, another 1,000
improvement in the energy per processing operation as well as in the energy per
bit of memory and of communication. As a team from industry and from research,
we present expectations, requirements and possible solutions for this challenging
energy scenario of femto- and atto-Joule electronics.
The introduction outlines the book’s structure, which aims to describe the
innovation eco-system needed for optimum-energy, sustainable nanoelectronics.
For the benefit of the reader, chapters are grouped together into interest areas like
transistors and circuits, technology, products and markets, radical innovations, as
well as business and policy issues.
1.1 From Nanoelectronics to Femtoelectronics
In the year 2000, the first microchips were produced with gate lengths <100 nm,
and microelectronics received the new label nanoelectronics. The drive in the
industry along the nano-roadmap towards shorter transistors continued in order
to build faster processors and to pack more memory bits on each chip. At the
same time, the research community had widespread programs running on
B. Hoefflinger (*)
Leonberger Strasse 5, 71063 Sindelfingen, Germany
e-mail: bhoefflinger@t-online.de
B. Hoefflinger (ed.), CHIPS 2020, The Frontiers Collection, 1

DOI 10.1007/978-3-642-23096-7_1, # Springer-Verlag Berlin Heidelberg 2012
2 B. Hoefflinger
quantum–nanometer-scale phenomena with considerable optimism for near-term

practical success. By the year 2010, the planned 32 nm milestone (node) had been
reached, including most of the expected chip specifications, but with a twist since
2005: the 3D integration of many thin chips on top of each other, connected through
TSVs (through-silicon vias), to enhance the progress on the roadmap. But the year
2010 was also marked by the rapidly growing consensus that the end of the
roadmap is near at 15 nm (2016) or 10 nm (2018?) at best, and that none of the new
quantum-nano devices will have any economic impact before 2025–2030. This
poses the serious question: Will the progression of chips come to a standstill, and
with it the world’s driving technology (information and communication)? The
answer is: Not necessarily, if we accept the nanometer-limit and, at the same
time, exchange the nanometer priority for a femtojoule priority: Energy per func-
tion, often called the power efficiency, is the new yardstick.
Between 2010 and 2020, the energy per chip function, such as processing,
memory, or communication, has to be reduced by a factor of 1,000 if the
nanoelectronics market is going to have enough to offer to six billion potential
global customers. Remarkably, many of these chip functions run at picojoule (pJ)
levels in 2010. The challenge is now to achieve femtojoule (fJ) levels. Therefore we
set our focus on moving from nano(meter) to femto(joule) electronics.
The task is challenging indeed, because the nanometer roadmap only offers a
final contribution of threefold at best, which may enter as 3 ¼ 9 in an energy
figure-of-merit (FOM), leaving us with another factor of >100 in needed
improvements within a decade. One approach to identifying a future strategy is
to consider how the remarkable advances in the 100 years of electronics and,
particularly, in the past 50 years of integrated circuits were achieved and which
repertory of R&D results of the past 30 years could be put to the test in product
developments of the present decade. This is the theme of Chap. 2, “From Micro-
electronics to Nanoelectronics”.
Our path to the 2020 goals will point out critical electronic functions, which are
most challenging and intriguing, since we reach and go below the energy per
operation in the synapses of natural brains: The typical energy per operation of a
neuron’s synapse is 10 fJ.
We pose five fundamental questions in this book, and we give answers, all of
which point out tough energy requirements on future nanochips so that we con-
clude: Sustainable nanoelectronics has to be femto(joule) electronics.
Here are the five questions:
– Why do we hit the end of the nano-roadmap?
– Which femtoelectronic solutions can we find for critical functions?
– What are the requirements for new chip products and how can we meet them?
– Which radical femtoelectronic solutions should we seek for intelligent
computing?
– What are the challenges for the femtoelectronics ecosystem of education,
research, and business?
– How does the 2020 world benefit from femtoelectronic chips?
1 Introduction: Towards Sustainable 2020 Nanoelectronics 3
1.2 Why Do We Hit the End of the Nano-roadmap?
Because microelectronics has advanced so successfully over 50 years with a linear

strategy of scaling down transistor sizes, it comes as a surprise that this mode of
progression will hit its limits by 2015. We deal with these problems in Chaps. 3
(10 nm transistors), 7 (ITRS, the nano-roadmap), and 8 (nanolithography).
In Chap. 3, we evaluate the future of eight chip technologies, revealing that their
development towards a fruitful future took between 25 and more than 35 years, an
example being the present emphasis on the 3D integration of chips, which was
taken on as a major research topic 30 years ago, only to be put aside in the late
1980s because the 2D scaling strategy did the job well and less disruptively.
Transistors with gate lengths of <10 nm on the scaling roadmap were built in the
lab and published before 2000. Why are their fundamental problems being consid-
ered only now? In Sect. 3.2, we identify two fundamental problems, which are
present no matter how precisely we can process these 10 nm transistors:
1. The atomic variance (only ~5 doping atoms in the channel) makes the spread of
transistor thresholds larger than the allowed supply voltage, so not all of them
can be turned on or off.
Gain = 1
800 8
Switching Energy
Switching Energy (eV)
Supply Voltage (mV)
600 x 6
400 4
x
x
200 2
x
x
Supply Voltage x
0 0
0 10 20 30 40
Gate Length (nm)
Fig. 1.1 Minimum supply voltage and switching energy of a CMOS inverter as a function of gate
length for an equivalent gate-oxide thickness (EOT) of 1 nm
4 B. Hoefflinger
2. The available voltage gain (because of drain-induced barrier lowering) becomes

<1, and without amplification, circuits lose their signals in noise (at least in all
electronics operating in natural environments).
Because of these fundamental limits, we actually find that the most important
basic amplifier in modern chip electronics, the CMOS (complementary metal–
oxide–semicondiuctor) inverter, has its minimum energy per operation at a gate
length of 20 nm (Fig. 1.1), because, moving towards 10 nm, we have to raise the
supply voltage in order to compensate for the facts listed above. Nevertheless, the
switching energy of this fundamental amplifier is just 5 eV ¼ 1018 J, which means
that we have to move just 25 electrons through a potential difference of 200 mV.
A 10 nm transistor and, for that matter, even a 20 nm transistor, performs this by
moving just one electron at a time through the channel at speeds (Fig. 1.2) depending
on the instantaneous voltages. That is, practically all minimum transistors with gate
lengths <20 nm are single-electron transistors. We present these fundamental
details at the outset of this book in order to show that, at the end of the nano-
roadmap, we are putting (at least digital) information on single electrons.
Our emphasis when discussing the International Technology Roadmap for
Semiconductors (ITRS) in Chap. 7 is the projection for many of the other components
frequency of occurrence
1 GHz 10 GHz 100 GHz 1 THz 10 THz
100 µA
current
10 µA
1 µA
100 nA
10 nA
1 nA
103 104 105 106 107
velocity (cm/s)
Fig. 1.2 Current carried by a
single electron at a given
1 ns 100 ps 10 ps 1 ps 0.1 ps
velocity over a 10 nm
distance transit time
on nanochips, such as capacitors and multilevel interconnects, which are important in

an assessment of the processing, memory, and communication functions on a chip.
Any of the aggressive projections for down-sizing the features on future chips
assume that we can image these features onto the Si wafers, and everyone agrees
that nanolithography is the absolute key technology in realizing these incredibly
complex structures effectively for many successive layers per wafer and millions
of wafers per year. Burn Lin, as vice president of lithography at the world’s
largest contract manufacturer of wafers, TSMC, recognized as one of the three
semiconductor companies at the forefront of technology, addresses both physical
and economic issues in Chap. 8. He analyzes the two major contending techniques,
EUV (extreme ultraviolet) and MEB (multiple-electron-beam direct write), and he
tells us that EUV systems impose extreme requirements on the primary laser
source, on the precision of mirrors and masks, and that the overall energy efficiency
from plug to wafer is so low that the necessary primary electrical power might be
bigger than the total power available for a gigafab (giant semiconductor fabrication
plant). On the other hand, an e-beam system, attractive because it is maskless,
would need 130,000 parallel, steered beams in order to be competitive, and such a
system has not been built as of 2010 owing to a lack of product development. These
enormous problems are a third, practical reason besides the fundamental ones listed
above for a 10 nm limit on chip features.
1.3 Which Femtoelectronic Solutions Can We Find

for Critical Functions?
Regarding the question of femtoelectronic solutions, we take the standard position

that innovations in technology drive progress (technology push), but we focus this
fundamental force on the energy question, combined with operating speed and chip
area (footprint/function), in order to meet the great expectations made of future
chips. The topics covered and the chapters in which they can be found are listed in
Table 1.1.
Table 1.1 Femtoelectronic Topic Chapter or section

topics in this book
Logic and computing 3, 10
Analog/digital interfaces 4
Interconnects and transceivers 5
Memories 11
3D integration 3.5, 12
MEMS sensors 13
Vision sensors 14
Retinal implants 17
Power-efficient design 9
Energy harvesting and chip autonomy 19
6 B. Hoefflinger
Fig. 1.3 The cross-coupled

inverter pair (quad) as a
differential signal Quad
regenerator, the fundamental
building block of ultra-low-
voltage digital
femtoelectronics
We start out with digital logic in Sect. 3.3, with the conclusion from the
preceding section that the variance of 10 nm transistors forces the application of
a new type of robust regenerator, the cross-coupled inverter pair, offering a
differential, high-speed, full-swing regenerator with minimum dynamic energy.
This block of four properly connected transistors (two NMOS and two PMOS)
becomes the fundamental circuit unit to be optimized regarding chip footprint and
process complexity, presented as a device-level 3D-integrated process in Sects. 3.4
and 3.5. The message here is that, in 10 nm CMOS, the target of making a transistor
falls short of the necessary fundamental building block, which is the four-transistor
cross-coupled inverter pair with minimum footprint, interconnects, and process
steps (Fig. 1.3).
Out of the multitude of applications of the differential regenerator in digital
logic, we select the toughest component, the high-speed, minimum-energy, and
minimum-footprint multiplier, the largest macro in arithmetic and, especially, in all
signal-processing units (images and video). As an example of possible development
directions, we describe the HIPERLOGIC multiplier in Sect. 3.6, which has the
potential to perform 600 M 16 16 bit multiplications with a power-efficiency of
one TOps mW1 ¼ 1 fJ, which would exceed the 2010 state-of-the-art by a factor
of 1,000 (Fig. 1.4).
On the higher levels of digital chip functionality, high-performance processors
and their future are described by Peter Roth, Christian Jacobi, and Kai Weber of
IBM in Chap. 10. Owing to the clock speed limits of <10 GHz towards the end of
the nano-roadmap, multicore processors have become a must, and they discuss the
associated challenges for operating systems and application development, needing
new levels of hardware–software co-design.
This complex design scenario is analyzed systematically by Barry Pangrle of
Mentor Graphics in Chap. 9, with a focus on power efficiency on all levels: system,
chip, process technology, test, and packaging. He comes to the conclusion that “it
won’t be the technology that’s the limit but the cost of implementation.”
Processors are the much-quoted hearts of advanced integrated chips, but much
more functionality is integrated today, namely true systems-on-a chip (SOCs).
These SOCs incorporate
– Sensors, which sense the environment and mostly produce analog output signals;
– Analog–digital interfaces converting analog sensor or receiver outputs to digital
signals for processing;
10 pJ
Energy/operation
or
107 eV Energy / b
1 pJ
Ch Ma
ip in
- C stre
hip am
100 fJ Int Mu
erc ltip
on
ne lier
cts
1 Synapse
10 fJ 6 • 104 eV
HIPE DRAM
RLO
GIC
Mult DiffDRAM
iplie
104 eV r 8 • 103 eV
1 fJ 6 • 103 eV
FEMTOELECTRONICS
100 aJ
SRA
10 aJ M
10 eV 7 eV
1 aJ
2010 2020
Year
Fig. 1.4 The power-efficiency of digital multipliers and high-speed memory cells advancing to
femtojoule and attojoule levels
– Receivers, which receive radio signals or signals on wires as well as optical

fibers;
– Transmitters, which send output signals on lines or into space;
– High-speed interconnects on-chip or chip-to-chip;
– Memories, which store digital data and instructions for processing these;
– Energy sources on-chip to produce or/and harvest a part or all of the chip energy
necessary for its operation;
– 3D integration as the key technology to merge these functionalities.
The state-of-the-art and the future of the highly sophisticated disciplines treating
these specialties are presented in dedicated chapters.
Boris Murmann of Stanford University with Yiannos Manoli and Matthias
Keller of the University of Freiburg, Germany, are recognized leaders in
8 B. Hoefflinger
analog–digital converters, and they show in Chap. 4 that these converters have
proceeded along their own, remarkable roadmap of energy per conversion and that
there is great potential for future improvements in energy and speed using
advanced, high-density CMOS technology nodes for digital-processing-based iter-
ation and correction schemes at low voltages for precision conversion rather than
traditional high-voltage, operational-amplifier-based converters.
Interconnects, transmitters, and receivers are studied in Chap. 5, and we find that
these communication functions are the most difficult to reduce in their energy
requirements. Even today, they make up ~1/3 of the total energy of chip systems,
and even with a rigorous transition to low-voltage differential signaling as well as
with 3D integration for shorter communication distances, the energy per bit will
only be reduced by <100 times in a decade (Fig. 1.4).
Memory energy is responsible for about another 1/3 of a chip-system’s energy,
and we will see in Chap. 11 that different memory types need different strategies for
their advancement. The high-speed static CMOS random-access memory (SRAM) is
best prepared for down-sizing both in footprint and in dynamic energy because of its
differential active cell and differential sense amplifiers. It will also benefit fully from
the 3D device-level integration of the fundamental four-transistor differential regen-
erator, so that it can reach a density of 16 Gbit/cm2 by 2020 with a dynamic energy
per bit of just 7 eV (Fig. 1.4). The density of a dynamic RAM (DRAM) is 5 times
higher in 2010. However, DRAM can only be scaled to 22 nm and 13 Gbit/cm2.
A differential DRAM is proposed, which would enable a 2D density of 50 Gbit/cm2
by 2020 and a read energy of 8,000 eV, eight times better than the best DRAM.
Figure 1.4 shows the development potential of the power efficiency of digital
multipliers (their energy per operation) and SRAM as well as DRAM (dynamic
energy per bit) over the decade 2010–2020. The remarkable advances are detailed
in the corresponding chapters. All these functions reach or surpass the
femtoelectronic mark or that of a neuron’s synapse at 10 fJ ¼ 6 104 eV per
operation (per bit).
Multilevel flash memories, the large-scale leaders among nonvolatile memories,
have made impressive progress since their first realization 1985. They are heading
towards a minimum of 500 eV/bit, but again with a scaling limit at 22 nm (3 bits per
cell) or 16 nm (2 bits per cell). These nonvolatile memories have seen the largest
development effort on 3D integration into chip stacks of up to 128 chips, producing
densities of many terabits per square centimeter. These solid-state drives (SSDs) are
displacing magnetic-disk memories. Among the alternative styles of nonvolatile
RAMs, phase-change (PCM) and resistive (RRAM) memories are progressing
towards 20 Gbit/cm2. More aggressive tokens for a nonvolatile bit, for example
the spin of an electron, continue to be burdened by energy- and space-hungry write-
and-read transistor circuitry. Single-electron wells with Coulomb confinement may
become the elements for a quantum-CCD (charge-coupled device) memory.
Regarding the variety of sensors, Jiri Marek and Udo-Martin Gómez of Robert
Bosch GmbH, Germany, world leaders in micro-electro-mechanical systems
(MEMS) for automotive applications, present, in Chap. 14, a comprehensive and
quantitative overview of how much it takes in creative and extreme quality-
conscious development to put >1 billion safety-critical electromechanical sensing

systems on the road. Tiny 3D mechanical structures combined with calibration and
processing chips are applied for the realization of new sensing systems as well as
for optimizing energy and manufacturing costs. These sensing systems are now
expanding into many, particularly portable, consumer products.
The other very important, highly integrated sensor systems are those for imaging
and video. These are discussed in Chap. 15 with respect to selected topics such as
sensitivity and pixel dark current, resolution and lens aperture, 3D integration or
backside illumination for high sensitivity, eye-like high dynamic range, and 3D
imaging. The performance of vision sensors is the first level in the complex and
endlessly challenging system of intelligent vision and performance- and bandwidth-
efficient video.
Very demanding vision specialties are retinal implants for blind people. Albrecht
Rothermel, the chip specialist in one of the world-leading consortia for such
implants, gives a balanced overview of the global status in 2010 in Chap. 17. He
describes in a holistic spirit all aspects, including chip design, physiology, surgery,
and training with patients; he further outlines what we can expect by 2020.
Implants are the exemplary type of chips for which energy is the dominant issue.
How can we possibly produce energy on-chip or harvest energy on the chip or from
its environment? One of the leading institutes in the field of energy harvesting is
IMTEK, the Institute for Microsystems Technology in Freiburg, in cooperation with
IMIT, Germany. Yiannos Manoli and his team present an exceptional tutorial,
assessing the state-of-the art and highlighting future developments of such sources
as motion, vibration, temperature differences, light, biofuel, and fuel cells. These
sources and transducers operate in ultralow-voltage systems, and the authors show in
Fig. 19.17 that, as of 2010, minimum supply voltages have been realized at L ¼ 130
nm, while voltages had to be higher at shorter gate lengths, results confirming – as in
our introductory Fig. 1.1 – that, energy-wise, shorter is no longer better.
1.4 What Are the Requirements for New Chip Products

and How Can We Meet Them?
After the assessment of Si technology and circuit capabilities and limits, we reflect
these against the requirements in the major market categories for nanoelectronics
(Table 1.2).
Table 1.2 Areas of Topic Chapter

applications and the chapters
in which they are treated Overview and markets 6
Superprocessors and supercomputers 10
Wireless and mobile companions 12, 13
Automotive applications 14
Medical implants 17
10 B. Hoefflinger
The overview in Chap. 6 tells us that the semiconductor industry, with aggregate
growth rates of 4–7% per year, advancing toward >450 billion US$ by 2020,
continues to be the technology driver for the information and communication
economy. Communication, consumer, and automotive chip markets show growth
rates above average. The perspective of
– Six billion mobile-phone subscribers,
– A computer tablet per student with free access to the world’s libraries, and
– A carebot (personal robot) per family (in Asia)
explains our emphasis on wireless mobile. Georg Kimmich, the system-on-chip
product manager at ST Ericsson, describes, in Chap. 12, the remarkable progress of
the 3D integration of heterogeneous chips, processors, memories, and transceivers,
in close interaction with wide-bandwidth architectures, toward high-quality, cost-
effective multimedia. In Chap. 13 Greg Delagi, Senior Vice-President for Wireless
Systems at Texas Instruments, takes on the full perspective of the personal mobile
companion of 2020:
– 3D imaging and display,
– Gesture interface,
– Integrated projection,
– Object and face recognition,
– Context awareness,
– Internet of things
– Brain–machine interface,
– Body-area network connection.
His chapter is the most explicit catalog of requirements on functional through-
put, transceiver sensitivity, power, and bandwidth, and, throughout, a 1,000 times
less energy per function, ultralow-voltage circuitry being the first thing to be
promoted.
On the other side of the energy equation, he stresses the needed progress on
energy harvesting. Considering the magnitude of the challenges, he asks for a
“strong collaboration in research and development from universities, government
agencies, and corporations” (see Sect. 1.6 and Chap. 22).
1.5 Which Radical Femtoelectronic Solutions Should

We Seek for Intelligent Computing?
As alternatives to traditional computing architectures with processor, instruction-

and data-memory, artificial neural networks (ANNs) saw about a dozen years of
intense research between 1988 and 2000, as often, a bit too early to have a broad,
disruptive impact. Systems with a large number of inputs and with tasks of making
decisions or recognizing patterns on the basis of rules, learning, and knowledge can
be realized with these networks at speeds and energy levels unmatched by other
architectures. Their final success is more likely now, because integrated electronic
neurons are becoming effective on all counts at Si technology nodes <100 nm and
because their fundamental 3D architecture can finally be realized with the large-
scale introduction of 3D chip integration. Digital neural networks and silicon brains
are treated in Chaps. 16 and 18, respectively.
Ben Spaanenburg of Lund University, Sweden, has had a distinguished career in
high-level chip design and cellular neural networks (CNNs). He and his coworker
Suleiman Malki present a concise tutorial on working CNN chips and chip systems,
both analog and digital. They describe new digital neuroprocessors for vision-in-
the-loop new media tasks, from software, through scalable architecture, to digital-
neuron design.
These CNNs are based on the extensively studied multilayer perceptron, which
has become a quasi standard. By contrast, really radical research programs were
launched in 2008 on silicon brains; on the one hand because of the availability of
ultralarge-scale nano-CMOS and 3D chip integration, but actually with the inten-
tion of building general-purpose, biomorphic chip systems with the brain’s com-
plexity of 1011 neurons and 1014 synapses (Chap. 18).
1.6 What Are the Challenges for the Femtoelectronics

Ecosystem of Education, Research, and Business?
The broad front along which chips can be advanced raises questions concerning all
kinds of resources: the energy crisis (Chap. 20), the extreme-technology industry
(Chap. 21), and education and research (Chap. 22).
The globally installed electric power in 2010 is ~2 TW (2 1012 W). Informa-
tion and communication technology (ICT), operated by electronic chips, is
estimated to need >20% of the world’s electric power. In Chap. 20, we look just
at the needs of data centers, those server farms with 36 million servers in 2010, the
processing backbone of the internet. They require ~36 GW, 10% of the total electric
power installed in Europe. These data centers need and are expected to increase
their performance 1,000-fold within a decade. This is obviously incompatible with
the electric power available, unless this performance increase is achieved with,
hopefully, a 1,000-fold improvement in power efficiency per function, which is
synonymous with a 1,000-fold reduction in energy per function, the magic factor
throughout this book. We also note in Chap. 20 that energy in a gigafab makes up
>80% of the cost of a fully processed Si wafer. This leads us to Chap. 21 on the chip
industry as “the extreme-technology industry”, marked by an investment rate
>15% as well as an R&D rate >15% of revenues, twice as high as the R&D rate
of the top pharmaceuticals company in the Organisation for Economic Co-operation
and Development (OECD) list of high-tech industries.
Naturally, these unique rates of progress can only be achieved with a unique
interest in and unique investments in highly skilled manpower at all levels from
kindergarten to retirement (Chap. 22). The concerns about the ecosystem of
12 B. Hoefflinger
education, research, industry, government, and the public are global. However, no
wave, like the one triggered by Sputnik in the 1960s or that in the 1980s is in sight as
of 2010.
After the world financial crisis of 2008–2009, with limited global financial and
energy resources, a clear, requirements-driven research strategy should be
established. For chips, as consistently stated throughout this book, the strategy should
be sustainable nanoelectronics towards femtojoule electronics, with the key
applications educational tools, health and care, communication, and safe mobility.
1.7 How Does the 2020 World Benefit

from Femtoelectronic Chips?
The refocusing from a nanometer to a femtojoule strategy for nanoelectronics

makes possible functionally powerful, energy-conscious chip systems serving the
global base of six billion potential customers. As stated in Chap. 23, many of these
pervasive chip innovations support health and safety in public, in private, and on the
shop floor, so they should be accompanied, from the outset, by qualification and
certification programs. Again, meeting the concerns of twenty-first century people,
minimum-energy nanoelectronics can attract broad interest in the public, well
beyond just the nanometer experts.
Chapter 2
From Microelectronics to Nanoelectronics
Bernd Hoefflinger
Abstract We highlight key events in over 100 years of electronic amplifiers and
their incorporation in computers and communication in order to appreciate the
electron as man’s most powerful token of information. We recognize that it has
taken about 25 years or almost a generation for inventions to make it into new
products, and that, within these periods, it still took major campaigns, like the
Sputnik effect or what we shall call 10 programs, to achieve major technology
steps. From Lilienfeld’s invention 1926 of the solid-state field-effect triode to its
realization 1959 in Kahng’s MOS field-effect transistor, it took 33 years, and this
pivotal year also saw the first planar integrated silicon circuit as patented by Noyce.
This birth of the integrated microchip launched the unparalleled exponential growth
of microelectronics with many great milestones. Among these, we point out the 3D
integration of CMOS transistors by Gibbons in 1979 and the related Japanese
program on Future Electron Devices (FED). The 3D domain has finally arrived as
a broad development since 2005. Consecutively, we mark the neural networks on-
chip of 1989 by Mead and others, now, 20 years later, a major project by DARPA.
We highlight cooperatives like SRC and SEMATECH, their impact on progress and
more recent nanoelectronic milestones until 2010.
2.1 1906: The Vacuum-Tube Amplifier
At the beginning of the twentieth century, the phenomenon of electricity (the charge
and force of electrons) had received over 100 years of scientific and practical
attention, and signals had been transmitted by electromagnetic waves, but their
detection was as yet very limited, because signal levels were small and buried in
noise. This changed forever when the vacuum-tube amplifier was invented in 1906
B. Hoefflinger (*)

14 B. Hoefflinger
D
f a
B
3 –
+
M 2 4 1
I2 C′ – +
A
C′′
Fig. 2.1 The vacuum-triode amplifier after De Forest 1907. In this field-effect triode, the
electrodes are the heated cathode on the left, which we would call the source today, the grid
in the center, which would be the gate, and the anode on the right, which we would call the drain
(# USPTO)
by Robert von Lieben in Austria [1] and Lee De Forest in the USA [2]. Its
predecessor was the vacuum-discharge diode, a two-terminal device consisting of
a heated cathode electrode emitting thermionic electrons, which are then collected
through a high electric field by another electrode, the anode, biased at a high voltage
against the cathode. This two-terminal device acts as a rectifier, offering a large
conductance in the described case of the anode being at a higher potential than the
cathode, and zero conductance in the reverse case of the anode being at a potential
lower than the cathode. The invention was the insertion of a potential barrier in the
path of the electrons by placing a metal grid inside the tube and biasing it at a low
potential with respect to the cathode (Fig. 2.1). The resulting electric field between
cathode and anode would literally turn the electrons around. Fewer or no electrons
at all would arrive at the anode, and the conductance between cathode and anode
would be much smaller. A variation of the grid potential would produce an
analogous variation (modulation) of the cathode–anode conductance. This three-
terminal device, the vacuum triode, consisting of cathode, anode, and control grid,
became the first electronic amplifier: It had a certain voltage gain AV, because the
grid–cathode input control voltage could be made much smaller than the
cathode–anode voltage, and it had infinite current gain AI at low rates of input
2 From Microelectronics to Nanoelectronics 15
Fig. 2.2 Equivalent circuit of i0

an ideal amplifier (voltage-
controlled current amplifier)
u1 gMu1 u0
changes, because there was no current flow in the input between cathode and grid,
while large currents and large current changes were effected in the output circuit
between cathode and anode. As a consequence, the power gain AVAI approaches
infinity.
It is worthwhile drawing the abstraction of this amplifier as a circuit diagram
(Fig. 2.2), because inventors have by now spent over 100 years improving this
amplifier, and they will spend another 100, even if the signal is not electrons. We
see that the input port is an open circuit, and the output port is represented by a
current source gmVin in parallel with an output resistance Rout.
The inventors of the vacuum-tube amplifiers were tinkerers. They based their
patent applications on effects observed with their devices and achieved useful
products within just a few years (1912).
10: Long-range radio (World War I)
These amplifiers launched the radio age, and they triggered the speculative
research on building controlled-conductance amplifying devices, which would
replace the bulky vacuum tubes with their light-bulb-like lifetime problems.
2.2 1926: The Three-Electrode Semiconductor Amplifier
The best solid-state analogue to the vacuum tube would be a crystal bar whose
conductance could be varied over orders of magnitude by a control electrode. This
is what the Austro-Hungarian–American physicist Julius Lilienfeld proposed in his
1926 patent application “Method and apparatus for controlling electric currents”
[3]. He proposed copper sulfide as the semiconducting material and a capacitive
control electrode (Fig. 2.3). This is literally a parallel-plate capacitor, in which the
field from the control electrode would have an effect on the conductance along the
semiconducting plate.
He did not report any practical results. However, since the discovery of the
rectifying characteristics of lead sulfide by K.F. Braun in 1874, semiconductors had
received widespread attention. However, it was not until 1938 that Rudolf Hilsch
and Richard Pohl published a paper, “Control of electron currents with a three-
electrode crystal and a model of a blocking layer” [4], based on results obtained
with potassium bromide. Shockley wrote, in his article for the issue of the IEEE
Transactions on Electron Devices commemorating the bicentennial of the United
16 B. Hoefflinger
11
15 15 13 15 12
+ –
14
16
16
+ 10
18
–
17 21
22
Fig. 2.3 The field-effect triode proposed by Lilienfeld in 1926 (# USPTO)
States in 1976 [5], that he had this idea in December 1939: “It has occurred to me
today that an amplifier using semi conductors rather than vacuum is in principle
possible”. Research continued during World War II on the semiconductor amplifier
[6], but a critical effort began only after the war. As we shall see, it was not until
1959 that the Lilienfeld concept was finally reduced to practice.
2.3 1947: The Transistor
One possible launch date of the Age of Microelectronics is certainly the invention
of the transistor in 1947 [5]. Shockley himself described the events leading to the
point-contact transistor as the creative failure mechanism, because the invention
resulted from the failure to achieve the original goal, namely a field-effect transistor
(FET) with an insulated gate in the style of the Lilienfeld patent. Nevertheless, this
failure, implemented as Ge or Si bipolar junction transistors, dominated microelec-
tronics into the 1980s, when it was finally overtaken by integrated circuits based on
insulated-gate FETs, the realization of the Lilienfeld concept.
2.4 1959: The MOS Transistor and the Integrated Circuit
Shockley described the first working FET in 1952, which used a reverse-biased pn
junction as the control gate, and junction FETs (JFETs) were then used in
amplifiers, where a high input impedance was required. In fact, when I was charged
in 1967 at Cornell University with converting the junior-year lab from vacuum
tubes to transistors, I replaced the vacuum triodes in the General Radio bridges by
junction FETs, and one of my students wrote in his lab report: “The field-effect
transistor thinks that it is a tube”.
Fig. 2.4 The MOS transistor 27

in Kahng’s patent filed in V
– +
1960 (# USPTO)
28
Vf L
21 22
24
–
23 25 19
13 +
p
18
16 17
15 n 14
12 11
The insulated-gate FET presented the sticky problem of large amounts of charge
at the dielectric–semiconductor interface masking any control effect by the gate
electrode. It was the use of thermally grown SiO2, the native oxide of Si, as the
dielectric on Si as the semiconductor that produced the breakthrough in 1959 [7]. It
enabled Atalla and Kahng to make the first working FET with oxide as the insulator
and a metal gate (Fig. 2.4) [8]. Its name, MOSFET, metal-oxide field-effect
transistor, was soon abbreviated to MOS transistor. It rarely appeared on the market
as a discrete, separately packaged device with three terminals, because it was easily
damaged by external voltages causing a breakdown of the insulator. MOS
transistors really developed their full potential as soon as many of them were
connected to perform functions inside an integrated circuit.
The coincidence with the invention of the integrated circuit at the same time can
be linked to the victory of Si over Ge as the favored semiconductor material and to
the rapid evolution of thermally grown SiO2 as the insulator perfectly compatible
with Si as the semiconductor. At Fairchild Semiconductor, founded in 1956 by
Robert Noyce and the so-called Treacherous Seven, who had left Shockley Tran-
sistor, the Swiss-born Jean Hoerni made his historic notebook entry on December 1,
1957, that SiO2 should be used to passivate the surface edges of p-n junctions, and
he illustrated that in his later patent application [9] in May 1959 as shown in
Fig. 2.5. The critical metallurgical junction between the bulk crystal and the
diffused region is fully shielded from the outside, and the surface is a plane oxide
insulator.
This planar structure led his boss Noyce to integrate various devices, such as
diodes and transistors, avoiding their interference by isolating them from each other
with reverse-biased pn junctions and by connecting their electrodes with thin Al
lines etched from a thin Al film evaporated on the oxide surface (Fig. 2.6). This
most famous integrated-circuit patent [10], filed by Noyce on July 30, 1959, is so
explicit, as shown in Fig. 2.6, that even today, 50 years later and at dimensions
18 B. Hoefflinger
13 12
14
12 19
16
17 18
16
21 23
24 22
Fig. 2.5 The planar manufacturing process according to Hoerni (# USPTO)
27 27 24
14 19 25
28 17 15 31 32 24 23 21 29
16 30 18 20
12
N P N P N
N 22
N
P
11
13
Fig. 2.6 The planar integrated circuit according to Noyce (patent filed 1959) [10] (# USPTO)
1,000 times smaller, the basic method of integrating devices side-by-side in Si is

still the same, resulting in a two-dimensional arrangement of transistors and other
devices, today at the scale of billions of these on the same chip.
The events of 1956–1959, pivotal for microelectronics, have been covered in
great detail, for example in [6 and 11], and we will not elaborate here on the parallel
patent by Kilby. It did not have any technical impact because of its Ge mesa
technology and soldered flying wires as interconnects.
Vo Vi
102
100 GATE 81a
GATE
+V –V
87a 96 101 94 93a
SOURCE 87 SOURCE 95
DRAIN DRAIN
82 93 81
P P N N
86 88 P
84 85 91 92 90
N
83
Fig. 2.7 Cross section through the chip surface showing a PMOS and an NMOS transistor side-
by-side and isolated from each other, after the Wanlass patent filed in 1963 [12] (# USPTO)
Fig. 2.8 The complementary +V

MOS inverter in the Wanlass
patent. Note that the gates are 37 50
connected and that the PMOS 41 30
drain is connected to the
NMOS drain, making this 54
pair an important 51 56
55 36
fundamental functional unit
(# USPTO) Vo
Vi 20 52
53
25 10
21
–V
While the Noyce patent showed a bipolar transistor, others in his company
concentrated on MOS and came up in 1963 with the ultimate solution for integrated
circuits: complementary MOS (CMOS) integrated circuits. Frank Wanlass filed this
famous patent in 1963 [12], and he presented a paper in the same year with C.T. Sah
[13]. Nothing explains the power of this invention better than the figures in the
patent (Figs. 2.7 and 2.8). The first one shows a PMOS transistor on the left with
p-type doped source and drain, conducting positive carriers (holes) under the
control of its gate, and an NMOS transistor on the right with its n-type source and
drain conducting electrons under the control of its gate. The transistors are isolated
from each other by a reverse-biased p-n junction consisting of a p-well and an n-
type substrate. The construction and functionality are listed in Table 2.1.
The complementary transistor pair as connected in Fig. 2.8 is the world’s most
perfect inverting amplifier for the restoration and propagation of digital signals. It
establishes a perfect HI and a perfect LOW at the output Vo (no. 56). There is zero
current flow, which means zero power consumption except during a signal transi-
tion, when the PMOS would provide current to charge the output to HI and the
20 B. Hoefflinger
Table 2.1 Functionality of complementary MOS transistors

Charge carriers Threshold ON voltage OFF voltage Drain–source voltage
NMOS neg pos pos neg pos
PMOS pos neg neg pos neg
Fig. 2.9 Cross section of a CMOS technology structure on an insulator as the substrate. The Lego
blocks indicate materials and aspect ratios. Yellow: n-Type Si (doped with donor-type atoms such
as As and P), Blue: p-Type Si (doped with acceptor-type atoms such as B), Red: Metal (or as yet
mostly highly As-doped polycrystalline Si), White: SiO2 insulator, Black: Interconnect metal
NMOS would discharge the output to LOW. As an amplifier, it has an operating

region with infinite voltage gain so that it is the best ever embodiment of the ideal
amplifier as shown in Fig. 2.2.
Fairchild did not pursue this ingenious invention, and Wanlass eventually left to
join the dedicated MOS companies General Instruments and AMI. His boss Noyce
did not embrace CMOS until 1980, almost 20 years later. The real CMOS pushers
were at RCA, where a pervasive CMOS culture was established through the first
textbook on FETs [14]. They also went one step further towards the ideal CMOS
structure by building the transistor layer in a thin Si film on insulator (later named
SOI), Fig. 2.9.
This obviously optimizes the vertical isolation, it allows a higher transistor
density, and it minimizes parasitic capacitances. In the mid-1960s at RCA, they
chose sapphire as the insulating substrate, which was expensive and did not match
with the Si lattice so that this approach was only used for space chips because of its
immunity to radiation. As of about 2000, 35 years later, the SOI structure finally
gained ground, based on massive process development over the last 20 years.
Back at Fairchild, Gordon Moore saw by 1965 the potential for doubling the
number of transistors per chip with every new generation of planar IC technology,
and in 1965 he had already sufficient data to show that this would happen every
18 months [15], as shown in Fig. 2.10, a reproduction of his 1975 update on this
famous curve [16].
Another powerful and lasting driving force for integrating more and more
functionality on a single chip is the minimum power and superior reliability
achievable in the planar process. We highlight this here with the observation that,
if we succeed in doubling the number of transistors or digital gates composed by
these on a single piece of silicon, the number of contacts to the outside world will
only increase by 40%, or the square root of 2. This is the famous Rent’s rule.
Because contacts mean cost and a reduction of reliability (lifetime), this means
extra pressure on increasing the chip area beyond the force provided by Moore’s
1M
BIPOLAR LOGIC
BIPOLAR ARRAYS
MOS LOGIC
64 K MOS ARRAYS
COMPONENTS PER CHIP
4K
256
16
1
60 65 70 75 80
YEAR
Fig. 2.10 Components per chip vs. year after Gordon Moore, 1965 and 1975 [16] (# 1975 IEEE)
law. These forces led to a new manufacturing paradigm, namely that of producing
chips with a manufacturing percentage yield smaller than the high nineties custom-
ary in classical manufacturing. This was a fundamental debate in the 1960s between
the classical camp of perfect-scale or right-scale integration and the advocates of
large-scale integration, who won and went through very-large-scale integration
(VLSI) to the giga-scale or giant scale integration (GSI) of today.
As an important historical coincidence, the Sputnik shock provided the perfect
scenario for a 10 push.
10: The computer on a silicon wafer (1966)
This enormous endeavor was launched through the US Air Force Office by
giving three contracts to Philco-Microelectronics, RCA, and Texas Instruments.
The speaker for the companies was Richard Petritz [17], then with TI and later a co-
founder of MOSTEK. This computer never flew, but these projects created design
automation, floor planning for yield and testability, automated test and the neces-
sary discretionary wiring for each wafer, the birth of direct electron-beam writing
on wafer, the best technology for rapid prototyping then, today, and probably also in
2020 (see chap. 8).
Another important side-effect of this project was that MOS circuits, which were
discredited as being slow and not suitable for computing as compared with bipolar
22 B. Hoefflinger
circuits, gained momentum because of their density, their topological regularity

(making them so suitable for design automation), and their much lower power
density and overall power. Through this project, the USA advanced toward putting
processing and memory functions with higher complexity on a larger scale on a
single chip, eventually leading to the first microprocessor [6].
When, in the internal Japanese competition for the electronic desktop calculator,
Sharp was looking for a competent source for their custom-designed calculator
chips, they finally gave a much publicized contract in 1969 to the Autonetics
division of Rockwell, because at home in Japan this capability did not exist. This
shame became the origin of a big national R&D program funded and coordinated by
MITI, the Ministry for International Trade and Industry:
10: The Joint Very-Large-Scale Integration (VLSI) Laboratory in Japan (1972–1980)
This initiative caught the rest of the world unprepared: the USA was deeply
entrenched in Vietnam, and the collapse of the dollar caused a long global reces-
sion. The VLSI Lab in Japan had researchers from all the large semiconductor
companies united under one roof. The result was the biggest boost, on a relative
scale, in microelectronic history for new, large-scale equipment and manufacturing.
Memories were identified as lead devices, and by 1980 NEC had become the
world’s largest semiconductor company and the Japanese dominated the global
memory market. Moore’s law and the scaling law provided simple yardsticks.
The scaling law in its 1974 version became the persistent driving force for
microelectronics. It takes us back to the fundamental structure of the FET, as
shown schematically in Fig. 2.11.
It had been noted in 1962 (see the history of the 1974 milestone of The Silicon
Engine [6]), that the FET was uniquely suited to miniaturization, by shrinking its
lateral dimensions L and W. To first order, the transistor area would be halved if L
and W shrank by a scaling factor of 0.7, providing a simple model for Moore’s law.
The scaling law declared that, in order to observe the maximum gate-field and
source-drain-field limits, the gate and drain–source voltages would have to be
tI Insulator
Gate
SiO2 Source Channel Drain SiO2
SiO2 L SiO2
Body
Fig. 2.11 Cross section of the MOS transistor annotated for scaling
reduced by the same amount, as would the insulator thickness tI, in order to
maintain the aspect ratio of the transistor.
The zero-order model for the drain current in the MOS transistor:
channel charge
current ¼ ; (2.1)
transit time
WL
channel charge ¼ capacitance voltage / field L; (2.2)
tI
transit time / field L; (2.3)
assumes the simple result in the constant-field case
WL
current / ; (2.4)
tI
suggesting a constant maximum current per micrometer for a given technology.

Many in-depth publications followed, analyzing the potential and limits of scaling-
down transistors, and we will discuss some in Sect. 3.2, but the one in 1974 by
Dennard et al. [18] triggered a strategy that is very much alive today, more than
30 years later, as we shall see in the International Technology Roadmap for
Semiconductors (ITRS) in Chap. 7.
Ever since the 1960s, we have seen an extreme rate of technology improvements
marked by the fact that the leaders rapidly increased their R&D budgets dispropor-
tionately to well over 20% of their total revenue. We have selected eight
technologies to analyze their history and future in Chap. 3. However, for the
grand picture, which innovations stand out since 1959 and 1963? The scaling
principle took it for granted that devices sat side-by-side in a 2D arrangement and
progress would come from making them smaller. Obviously there was the potential
of going into the third dimension.
2.5 1979: The Third Dimension
As we saw in Fig. 2.8, the complementary MOS transistor pair shares the input gate
contact. This pair was the target of Gibbons and Lee at Stanford, and they produced
the CMOS hamburger in 1979 [19]. In Fig. 2.12, we see the cross section of their
PMOS/NMOS transistor pair.
This can be wired into the basic inverting amplifier of Fig. 2.8 on a footprint of
close to one transistor, which would double the circuit density. Their invention is
indicative of the process innovation at the end of the 1970s: The upper Si film, in
which their NMOS transistor was formed, was deposited as a second polycrystalline-
silicon film after the first film for the gate. That upper layer was then recrystallized
with the energy from a laser. Recrystallizing Si deposited on oxide was one way to
24 B. Hoefflinger
Fig. 2.12 3D CMOS

transistor pair after Gibbons
and Lee (1980) [19] (# 1980
IEEE)
Fig. 2.13 Cross section with three transistor layers, after Kataoka (1986) [20] (# 1986 IEEE)
obtain multiple layers of silicon on top of each other separated by insulating layers of
SiO2. High-energy implantation of oxygen and proper high-temperature annealing
would produce buried layers of SiO2, thereby realizing a Si–SiO2–Si sequence of
films potentially suitable for multiple device layers. In fact, after the heavily product-
oriented VLSI Laboratory, the Japanese launched another, more fundamental effort:
10: The program on Future Electron Devices, FED (1980)
Here we will only focus on the part 3D integration and recapitulate some results,
because it took until about 2005, roughly 25 years, for this strategic direction to
take center-stage. In Fig. 2.13, we see the cross section of a Si chip surface with two
crystallized Si films, allowing three transistor layers on top of each other [20]. 3D
integration became part of the research efforts that evolved to be the greatest global
renaissance in microelectronics history.
10: VHSIC, SRC in the USA, ESPRIT in Europe (1981)
At the end of the 1970s, the Vietnam decade and a long global recession,
entrepreneurial and intellectual forces stepped forward – and were heard: Simon
Ramo, Ray Stata, and Robert Noyce in the USA, François Mitterand and Jean-
Jacques Servan-Schreiber in Europe were quoted everywhere in association with
new, substantial initiatives. The strategic program on very-high-speed integrated
circuits (VHSIC) made possible a set of very advanced, highest quality CMOS
manufacturing lines, certified by the RADC (Rome Air Development Center in up-
state New York) as qualified manufacturing lines (QMLs), which also provided
foundry services to several totally new or refurbished university departments.
Hardware, software, and tools for automated chip design were donated on a large
scale to US universities. What started as a summer course, given by Carver Mead of
Caltech and Lynn Conway of Xerox Palo Alto Research Center (Xerox PARC) at
MIT, VLSI Design, became a bible for tens of thousands of students and industrial
engineers [21]. The Defense Advanced Research Projects Agency (DARPA)
launched the foundry service MOSIS (MOS Implementation Service) for
universities under the leadership of Paul Losleben and a special university equip-
ment program of hundreds of millions of dollars, incredible dimensions those days.
The VHSIC Program produced the first standard hardware-description language
VHDL (originally VHSIC Hardware Description Language), which is still the
standard today. A uniquely successful university–industry research cooperative,
the SRC (Semiconductor Research Cooperative), was founded in 1982 under the
chairmanship of Robert Noyce, which quickly grew to a capacity of 500 Ph.D.
students working on all aspects of microelectronics together with assignees from
industry. A leader from industry, Eric Bloch from IBM, became president of the
National Science Foundation (NSF). He created the university Engineering
Research Centers (ERCs) on the premise that at least three faculties and more
than 15 professors would work together with over 50 Ph.D. students in such a
center. The program started with ten centers and now has over 50 in the USA.
The European Commission in Brussels, until then focused on coal, steel, and
agriculture, prepared a technology program. As a historical note, as founders and
leaders of two of the existing university pilot lines for large-scale integrated circuits
in Europe, in Leuven, Belgium, and Dortmund, Germany, Roger Van Overstraeten
and I were invited in 1979 to advise Giulio Grata, the responsible person in
Brussels, on the elements of a European Microelectronics Research Program,
which was launched eventually as ESPRIT (European Strategic Programme for
Research in Information Technology).
The 1980s were marked by numerous megaprojects, multi-hundred million
dollar investments per new facility, many with public money, one typical target
being the first 1 Mbit DRAM memory chip. The communist German Democratic
Republic went bankrupt on its megaproject, because they had to pay outrageous
26 B. Hoefflinger
amounts of money to acquire US equipment and computers through dark channels in

the communist bloc. New regions and states competed for microelectronics industry
investments, and today’s map of chip manufacturing centers was pretty much
established then. That period was also the origin of the global-foundry business
model exemplified by TSMC (Taiwan Semiconductor Manufacturing Co.)
established by Morris Chang after a distinguished career at Texas Instruments (TI).
One special result of the VHSIC program was high-speed electron-beam direct-
write-on-wafer lithography for rapid prototyping and small-volume custom
integrated circuits. The variable-shape beam, vector-scan systems Perkin-Elmer
AEBLE 150, used by ES2 (European Silicon Structures) in France, and Hitachi
HL700, used in Japan, Taiwan, and in my institute in Germany, established
electron-beam lithography as a viable technology to help the scaling-driven indus-
try to realize one new technology generation after the other. Line widths of 400 nm
or less were considered the limit of optical lithography so that sizeable centers
for X-ray lithography were set up at large synchrotrons in the USA at Brookhaven
and in Germany at BESSY in Berlin and COSY in Karlsruhe. Wafer shuttles
were planned to transport wafers there in volume, plans which never materialized
because optical lithography with deep-UV laser sources was able to do the job.
With a capability of millions of transistors on a single chip and supported by a
foundry service for up-to-date prototyping, the research community was
encouraged to take on big questions, such as information processing not with the
von Neumann architecture of processor, program memory, and data memory, but
closer to nature, also named biomorphic.
2.6 1989: Neural Networks on Chips
The integration of large numbers of electronic functions on a single microchip was

exploited from the early 1960s for array-type tasks, one particular direction being
imager arrays made up of optical sensor elements and circuits for reading out the
signals from the picture elements (pixels) [14]. The Japanese FED Program pro-
duced a 3D technology in which a Si photodiode layer was fabricated on top of two
CMOS layers [22], where the CMOS layers would be used for reading and
processing the pixel signals (Fig. 2.14).
This can be viewed as an early embodiment of a layered Si retina. A specific
retina was then proposed by Mead and Mahowald (Fig. 2.15) in Mead’s book
Analog VLSI and Neural Systems [23]. This book again became a bible and the
opener for worldwide activity on building neural networks on Si microchips.
Figure 2.15 shows an analog implementation with resistors as synapses. Even in
such a rigid setup, powerful functions such as the detection of shapes could be
performed. Many other researchers chose general, programmable digital two-
layer perceptrons. An example is shown in Fig. 2.16 [24], which can help to
explain why these networks have significant potential for intelligent information
processing.
Fig. 2.14 Micrograph of an array of phototransistors with vias to the underlying CMOS layers for
read-out electronics (1986) [22] (# 1986 IEEE)
Fig. 2.15 The analog neural network of Mead and Mahowald for vision tasks such as shape
detection consisting of active pixels and resistors as synapses (1989) [23] (# Springer 1989)
For the neural controller for automatic steering of a vehicle, driving data
obtained on a 2 km section of a normal road with a human driver in the car were
enough to train the neurocontroller. This is just one example of the prolific research
that was launched through the investments in the 1980s.
The 1990s were marked by the worldwide re-engineering of large companies,
enhanced by structural consequences after the end of the Cold War. This could have
28 B. Hoefflinger
fyaw (tk) CROAD(tk) cROAD(tk) yOFF(tk) yOFF(tk)

1 2 3 4 5
6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
10
21 i
cROAD (tk) = cROAD (tk–i) ∗ 1–
10
i=1
û (tk + 1) 10
i
yOFF (tk) = yOFF (tk–i) ∗ 1–
10
i=1
Fig. 2.16 Trainable neural controller with 21 neurons for automatic steering of a vehicle (1993)
[24] (# 1993 IEEE)
hit the semiconductor industry much worse if there had not been two large-scale
effects offsetting these problems:
• The force of Moore’s law and of the scaling law.
• The tremendous push of the Tigers Korea, Taiwan, and Singapore, with just
about 80 million people taking on the world of chip manufacturing, and
dominating it today.
The effect of both of these factors has been the refocusing of the traditional
players and an unparalleled strategic and global alliance of all major players.
10: International SEMATECH and the Roadmap (1995)
Originally, SEMATECH started in 1988 as a cooperation of the US semicon-

ductor and equipment industry, including government funding, in order to
strengthen the US position in the field. In Europe, the JESSI (Joint European
Submicron Silicon) project was initiated under the umbrella of EUREKA, the
less bureaucratic agency that manages European cooperative strategy, while the
funding runs on a per-country basis, which means that national governments fund
their constituency. The focus on manufacturing, which relies much on global
resources and partnering, soon let non-US companies join so that International
SEMATECH was formed in 1995, first concentrating on developing the capability
to manufacture on the basis of newly available 300 mm wafers. The International
Technology Roadmap for Semiconductors (ITRS) was started with worldwide
Fig. 2.17 The ITRS, shown here for the minimum dimension: 1999 issue
expert committees, and it became the single most powerful document defining the
global strategy of the semiconductor industry. SEMATECH has expanded its
charter since 2000 by operating large cooperative R&D facilities in Austin, TX,
and Albany, NY, including major funding by the respective states. The Europeans
concentrated their most significant efforts in Grenoble, France, and Leuven,
Belgium, and IMEC (the Inter-University Microelectronics Center) in Leuven,
founded by Roger Van Overstraeten in 1983, is today the world’s largest indepen-
dent R&D center for nanoelectronics. No other industry has developed such a joint
global strategy and infrastructure to master the progress manifested by a cumulative
R&D budget above 20% of revenue. It is interesting to overlay an early Roadmap
for the minimum required lateral dimension on chips with more recent data
(Fig. 2.17). It shows that the rate of progress on the nanometer scale has been
pushed repeatedly beyond earlier targets [25].
Towards 100 nm, the limits of scaling down MOS transistors were seen as being
near, triggering speculative research on that limit and how to go beyond it. A very
short channel of well under 100 nm between the metallurgical source and drain
junctions was producing a situation in which electrons would tunnel through these
junctions and might get trapped in the channel island if the electrostatic “binding”
energy exerted by the gate were large compared with the kinetic energy kBT at the
environmental temperature T (kB is the Boltzmann constant). Obviously, cooling
would help to create this situation, as well as a very small channel island, where this
electron would be locked up and only released for certain gate voltages and
source–drain voltages. This tiny island containing a single electron (or none) was
called a quantum dot, and it became very attractive as a memory element. A three-
terminal field-effect triode (Fig. 2.18) consisting of a gate, a quantum dot and two
30 B. Hoefflinger
Fig. 2.18 Concept of a SET

(1987) (# Wiley-VCH)
tunnel junctions was proposed as a single-electron transistor (SET) in 1987 [26].

The first experimental SETs were reported a few years later, and, clearly, one
electron per bit of information would be a major quantum step of progress.
For comparison, at the time, a memory transistor stored a ONE as 50,000
electrons on its floating gate, and a ZERO was about 5,000 electrons in a non-
volatile memory. The programming pulse had to place these amounts of charge on
the floating gate with narrow margins, and the read-out amplifier had to distinguish
this difference. Research began on techniques to place three distinguishable
amounts of charge on the floating gate, which would establish three threshold
voltages and to refine the readout so that it could differentiate between these
voltages [27]. Figure 2.19 shows measured threshold-voltage distributions. There
was no overlap, and the means were about 1 V apart. Only 10,000 electrons were
needed now to distinguish the bits [27]. This capability provided a true quantum
jump in bit density. This storage capability development advanced with remarkable
speed to about 250 electrons in 2010 as the length of a memory transistor decreased
to 34 nm.
2.7 2000: The Age of Nanoelectronics Begins
The Y2K effect in microelectronics was that volume chip production reached the
100 nm lithography level, and compatible overall processing was achieved. Total
capital expenditure for a factory based on 300 mm wafers exceeded US$ 1 billion.
Fig. 2.19 Distribution of multiple threshold voltages achieved in 1995 [27] (# 1995 IEEE Press –
Wiley)
10: Next-Generation Lithography, NGL (2000)
The performance of optical lithography was seen to be at its limits, and this, in
fact can be considered as another attribute of the new age of nanoelectronics,
namely that, the generation of these sub-100 nm lateral structures and zillions of
these on a wafer would require a lithography beyond short-wavelength refractive-
optics-based patterning. The largest project in the history of International
SEMATECH now became NGL (next-generation lithography), aimed at providing
a non-optical alternative for anything smaller than 45 nm to be available for
32 B. Hoefflinger
prototyping in 2005. The contenders were ion-beam lithography (IBL) and extreme
ultraviolet (EUV) lithography.
IBL was favored in Europe. A prototype was built in Vienna, Austria, based on a
hydrogen-ion beam going through silicon stencil masks produced in Stuttgart,
Germany. It was completed and demonstrated 45 nm capability in 2004. However,
the international SEMATECH lithography experts group decided to support solely
the EUV project, because IBL was assessed to not have enough throughput and less
potential for down-scaling. For EUV lithography, a sufficiently effective and
powerful source for 13 nm radiation, as well as the reflective optics and masks,
had not become available by 2010, and EUV lithography has been rescheduled for
introduction in 2012.
On the evolutionary path, optical lithography survived once more as it did
against X-rays in the 1980s. At the level of deep UV (DUV), 193 nm, a major
breakthrough was achieved by immersing the lens into a liquid on top of the silicon
wafer. The liquid has a refractive index several times higher than air or vacuum
significantly increasing the effective aperture. This immersion lithography saved
the progress on the roadmap for one or two generations out to 22 nm. Thereafter, the
future beyond optical is an exciting topic to be covered in Chap. 9.
It is natural that the scaling law cannot be applied linearly. Although lithog-
raphy provided almost the factor 0.7 per generation, the transistor size could not
follow for reasons of manufacturing tolerance, and its switching speed did not
follow for physical reasons. Therefore, the industry as of 2000 had to embrace a
more diversified and sophisticated strategy to produce miniature, low-power,
high-performance chip-size products. It finally embraced the third dimension,
20 years after the FED Program in Japan, which we discussed in Sect. 2.6. Of
course, this renaissance now occurred at much reduced dimensions, both lateral
and vertical. While in the mid-1980s it was a touchy art to drill a hole (from now
on called a via) through a 200 mm thick silicon/silicon dioxide substrate, wafers
were now thinned to 50 mm or less, and 20 years of additional process develop-
ment had produced a formidable repertory for filling and etching fine structures
with high aspect ratios of height vs. diameter. Test structures on an extremely
large scale emerged since 2006 with wafers stacked and fused on top of each
other and thin W or Cu vias going through at high density to form highly
parallel interconnects between the wafer planes. One early example is shown
in Fig. 2.20 [28].
The mobilization in this direction of technology has been remarkable. Stacked
wafers with through-silicon vias (TSVs) obviously provided a quantum jump in bit
and transistor density per unit area. Enthusiastic announcements were made by
industry leaders proclaiming new laws of progress beyond Moore. And it is true
that, besides the gain in density, interconnect lengths are much reduced, partly
solving the problem of exploding wiring lengths in 2D designs. The added
manufacturing cost is accrued at dimensions that are more relaxed, and pre-testing
each wafer plane before stacking provides more options for handling the testability
of the ever more complex units. Stacking and fusing processor planes and memory
planes offers a divide-and-conquer strategy for the diverging roadmaps for
Fig. 2.20 Cross sections

through a 3D memory
produced by stacking eight
wafers with TSVs (2006) [28]
(# 2006 IEEE)
processor and memory technologies and for the partitioning of the total system. We
will come back to the significance of this type of 3D evolution in Chap. 3.
The first decade of the new millennium has seen further tremendous progress in
CMOS technology in nanoelectronics, best described by the recent consensus that
NMOS and PMOS transistors have reasonable and classical or conventional
characteristics down to channel lengths of 5 nm, so that the worldwide design
know-how and routines can be leveraged for new and improved products to an
extent that is only limited by our creative and engineering capacity to develop these.
For the grand long-term picture, we address two recent achievements that have
great potential or that indicate the direction in which we might perform speculative
research on how to achieve entirely new levels of electronic functionalities.
2.8 2007: Graphene and the Memristor
The ultimately thin conducting film to which one could apply control, would have a
thickness of one atomic layer. It would be a 2D crystal. It is not surprising that, in
the context of widespread carbon research, this 2D crystal was eventually realized,
observed, and characterized in carbon, where it is called graphene (Fig. 2.21). In
2007, Geim and Novoselov pulled off this single-atom-thick film of carbon from
graphite with Scotch tape and transferred it to a SiO2 layer on top of silicon [29].
The graphene layer fit on the oxide layer so well that the measurements confirmed
theories on 2D carbon crystals going back to 1947, and high electron mobilities
were observed. The already large carbon research community converged and
expanded on graphene. A high-frequency transistor and an inverter with comple-
mentary transistors were reported soon after. The film deposition techniques for
producing these graphene layers appear to be compatible with large-scale Si
manufacturing so that graphene has high potential for future nanoelectronics.
34 B. Hoefflinger
Fig. 2.21 Real graphene

single-atom layer with the C
atoms clearly visible (From
Wikipedia)
Fig. 2.22 Micrograph of 17

memristors (J.J. Yang, HP
Labs)
Another recent achievement resulting from nanometer-scale electronics research

is a new two-terminal device, which has an analog memory of its past with high
endurance. It was reported in 2007 by Williams and members of his Laboratory for
Information and Quantum Systems at Hewlett-Packard [30]. The device consists of
two titanium dioxide layers connected to wires (Fig. 2.22). As the researchers
characterized their devices, they arrived at a model that corresponded to the
memristor, a two-terminal device postulated and named by Leon Chua in 1971
[31] on theoretical grounds. The memristor would complement the other three
devices resistor, capacitor, and inductor.
A two-terminal device that could be programmed or taught on the go would be

very powerful in systems with distributed memory. For example, the resistor
synapses in Mead’s retina (Fig. 2.15) could be replaced by these intelligent resistors
to build very powerful neural networks for future silicon brains.
With graphene as a new material and the memristor as a new device, we
conclude our grand overview of the technological arsenal that has been developed
over more than 60 years and which forms the basis of our 2020 perspective in the
following chapter.
References
1. von Lieben R.: Kathodenstrahlenrelais. German Patent No. 179807. Issued 4 March 1906
2. De Forest L.: Device for amplifying feeble electrical currents. US Patent No. 841387, filed
25 Oct 1906. Issued 15 Jan 1907
3. Lilienfeld J.E.: Method and apparatus for controlling electric currents. US Patent No. 1745175,
filed 8 Oct 1926. Issued 18 Jan 1930
4. Hilsch R, Pohl R.W.: Steuerung von Elektronenstr€ omen mit einem Dreielektrodenkristall und
ein Modell einer Sperrschicht [Control of electron currents with a three-electrode crystal and a
model of a blocking layer]. Z. Phys. 111, 399 (1938)
5. Shockley W.: The path to the conception of the junction transistor. IEEE Trans. Electron Dev.
23, 597 (1976)
6. See “The Silicon Engine” at www.computerhistory.org/semiconductor/. Accessed Feb 2011
7. Atalla M.M.: Stabilisation of silicon surfaces by thermally grown oxides. Bell Syst. Tech. J.
38, 749 (1959)
8. Kahng D.: Electric field controlled semiconductor device. US Patent No. 3102230, filed
31 May 1960. Issued 27 Aug 1963
9. Hoerni J.A.: Method of manufacturing semiconductor devices. US Patent No. 3025589, filed
1 May 1959. Issued 20 March 1962
10. Noyce R.N.: Semiconductor device-and-lead structure. US Patent No. 2981877, filed 30 July
1959. Issued 25 April 1961
11. Saxena A.N.: Invention of Integrated Circuits – Untold Important Facts. World Scientific,
Singapore (2009)
12. Wanlass F.M.: Low stand-by power complementary field effect circuitry. US Patent No.
3356858, filed 18 June 1963. Issued 5 Dec 1967
13. Wanlass F.M, Sah C.T.: Nanowatt logic using field-effect metal-oxide semiconductor triodes.
IEEE ISSCC (International Solid-State Circuits Conference) 1963, Dig. Tech. Papers,
pp. 32–33
14. Wallmark J.T, Johnson H (eds.): Field-Effect Transistors. Prentice-Hall, Englewood Cliffs
(1966)
15. Moore G.: Cramming more components onto integrated circuits. Electron Mag. 38(8),
114–117 (1965)
16. Moore G.: Progress in digital integrated electronics. IEEE IEDM (International Electron
Devices Meeting) 1975, Tech. Dig., pp. 11–13
17. Petritz R.L.: Current status of large-scale integration technology. In: Proceedings of AFIPS
Fall Joint Computer Conference, Vyssotsky, Nov 1967, pp. 65–85
18. Dennard R.H, Gaensslen F.H, Yu H.N, Rideout V.L, Bassous E, LeBlanc A.R.: Design of ion-
implanted MOSFET’s with very small physical dimensions. IEEE J. Solid-State Circuits 9,
256 (1974)
19. Gibbons J.F, Lee K.F.: One-gate-wide CMOS inverter on laser-recrystallised polysilicon.
IEEE Electron Dev. Lett. 1, 117 (1980)
36 B. Hoefflinger
20. Kataoka S.: Three-dimensional integrated sensors. IEEE IEDM (International Electron
Devices Meeting) 1986, Dig. Tech. Papers, pp. 361–364
21. Mead C, Conway L.: Introduction to VLSI Systems. Addison-Wesley, Reading (1979)
22. Senda K, et al.: Smear-less SOI image sensor. IEEE IEDM (International Electron Devices
Meeting) 1986, Dig. Tech. Papers, pp. 369–372
23. Mead C, Ismail M.: Analog VLSI Implementation of Neural Systems, ISBN 978-0-7923-9040-
4, Springer (1989).
24. Neusser S, Nijhuis J, Spaanenburg L, Hoefflinger B.: Neurocontrol for lateral vehicle guid-
ance. IEEE Micro. 13(1), 57 (1993)
25. www.ITRS.net/. Accessed Feb 2011
26. Likharev K.K.: IEEE Trans. Magn. 23, 1142 (1987)
27. Bauer M, et al.: A multilevel-cell 32 Mb flash memory. IEEE ISSCC (International Solid-State
Circuits Conference), Dig. Tech. Papers, 1995, pp. 132–133
28. Lee K et al.: Conference on 3D Architectures for Semiconductor Integration and Packaging,
San Francisco, Oct–Nov 2006
29. Geim A.K, Novoselov K.S.: The rise of graphene. Nat. Mater. 6, 183 (2007)
30. Wang Q, Shang D.S, Wu Z.H, Chen L.D, Li X.M.: “Positive” and “negative” electric-pulse-
induced reversible resistance switching effect in Pr0.7Ca0.3MnO3 films. Appl. Phys. A 86,
357 (2007)
31. Chua L.O.: Memristor – the missing circuit element. IEEE Trans. Circuit Theory 18,
507 (1971)
Chapter 3
The Future of Eight Chip Technologies
Bernd Hoefflinger
Abstract We select eight silicon chip technologies, which play significant roles in
the decade 2010–2020 for the development of high-performance, low-energy chips
2020 and beyond. In the spirit of the 25-years rule, all of these technologies have
been demonstrated, and some, in fact, are very mature and yet are worth to be
revisited at the nano-scale.
The bipolar transistor remains superior in transconductance and bandwidth, and
the complementary cross-coupled nano-pair can become the best ultra-low-energy
signal-regenerator.
MOS and CMOS circuits continue to be the most effective solutions for giant-
scale integration in a silicon-on-insulator technology. However, the end of progress
with just scaling down transistor dimensions is near, and this is not a matter of
technology capability, but one of atomic variance in 10 nm transistors. Once only
~6 doping atoms are responsible for their threshold and voltage gain, 95% of these
transistors would have between 1 and 9 such atoms. Their threshold would vary
more than their supply voltage. We show that, at these dimensions, not a transistor,
but the cross-coupled pair is CMOS inverters is the elementary and necessary signal
regenerator at the heart of ultra-low-voltage differential logic. This assembly of
four transistors is the perfect target for 3D integration at the transistor-level on-chip,
and selective Si epitaxy is shown as an exemplary solution, including self-assembly
eliminating certain lithography steps. This optimized 4T building block enables a
6T SRAM memory cell scalable to a 2020 density competitive with a DRAM cell,
which cannot be scaled because of capacitor size and transistor leakage. The 4T
block is also the key accelerator in the differential logic HIPERLOGIC, exemplified
by an n-bit by n-bit multiplier, which also illustrates the importance of new circuit
architectures. DIGILOG, a most-significant-bit-first multiplier has a complexity
O(3n) versus O(n2/2) in commonly used Booth multipliers. With HIPERLOGIC, a
B. Hoefflinger (*)

38 B. Hoefflinger
16 16 bit multiplier is projected to require only 1fJ per multiplication in 2020,

about 1,000-times less than the status in 2010 and 10-times less than a synapse in
the human brain. Therefore, processing has the potential to become 1,000-times
more energy efficient within a decade.
The 3D integration of chips has become another key contributor to improve the
overall energy efficiency of systems-of-chips incorporating sensing, transceiving,
and computing.
In the previous chapter, we highlighted those technologies that significantly

advanced microelectronics to the present level of nanoelectronics. In this chapter,
we focus on eight chip technologies, the history of their evolution, and their future
role and potential. We highlight their S-curves of innovation rate and market
penetration with four milestones:
– First prototype
– First product
– First killer product
– Market dominance
Our eight candidates are:
– Bipolar transistors
– Metal–oxide–semiconductor integrated circuits (MOS ICs, PMOS or NMOS)
– Complementary MOS (CMOS) and bipolar CMOS (BiCMOS) ICs
– Silicon-on-insulator (SOI) CMOS ICs
– 3D CMOS ICs
– Dynamic and differential MOS logic
– Chip stacks
– The single-electron transistor and other concepts
The evolution of all eight technologies and their future significance are displayed
in Fig. 3.1, and for each one, we will comment on their past and present role as well as
describe the challenges and opportunities looking at 2020 and beyond.
1. The bipolar transistor, first prototyped in 1950, was the transistor type of choice
for the first integrated circuits in 1960. Its killer products, the operational
amplifier and transistor–transistor logic (TTL), brought an IC market share for
this type of IC of over 95% in 1970. It had lost its lead by 1980, and bipolar ICs
are today, on the basis of heterojunction bipolar transistors (HBTs) with maxi-
mum frequencies of 600 GHz, a high-value niche specialty for terahertz (THz)
transceivers and photonic modulators.
2. MOS field-effect-transistor ICs appeared as prototypes in 1960, earned some
limited recognition as controllers and random-access memories (RAMs) by 1968,
and began their March to the top in 1972 with microprocessors and dynamic RAMs
(DRAMs) as killer products, which led MOS ICs to pass bipolar ICs in 1980.
3. CMOS (complementary MOS) ICs were first presented in 1965 and became
a success for those western and Japanese companies specializing in watch
3 The Future of Eight Chip Technologies 39
1950 1960 1970 1980 1990 2000 2010 2020
Bipolar Transistor
PMOS/NMOS-ICs
CMOS and BICMOS-ICs
SOICMOS-ICs
3DCMOS-ICs
Chip-Stacks
Single-Electron ICs
Dynamic and
Differential
MOS Logic
First Realisation
First Product
First Killer Product
Market Dominance
Fig. 3.1 The life cycles of eight chip technologies
circuitry and high-speed static RAM (SRAM). Killer products such as these
SRAMs, high-speed, low-power microprocessors, and application-specific,
mixed-signal circuits and gate arrays forced all semiconductor companies by
1981 to quickly establish a CMOS capability, often enhanced by on-chip
CMOS-bipolar drivers and amplifiers, a specialty that had been proposed in
1969. As early as 1985, CMOS had taken over the lead from single-transistor-
type MOS ICs, and it has kept this lead ever since.
4. A fundamental improvement on CMOS, namely putting the complementary
NMOS and PMOS transistors into a thin Si film on an insulator (SOI), had
been realized as early as 1964 on sapphire and began to gain ground on silicon
dioxide as of 1978, but it took until 2000 for major players to introduce volume
products, and it may take until 2012 for SOI-CMOS to become the technology
40 B. Hoefflinger
type with the leading market share. It is the only type that is scalable to transistor
channels <10 nm long and will be the workhorse for the 2020s.
5. 3D CMOS technology, folding PMOS and NMOS transistors on top of each
other with the gate sandwiched in between, was first demonstrated in 1979 and
was brought to maturity with selective epitaxial overgrowth (SEG) by 1990, but
it may take until about 2012 for its net reduction of processing steps and of area
and wire savings to catch on. As an exemplary goal, a 3D six-transistor SRAM
cell has been presented for the 10 nm node, requiring just ten electrons per bit
and providing a density of 10 Gbit/cm2.
6. Dynamic MOS logic was first realized in 1965 to exploit the fundamental synergy
of charge storage, switching, and amplification, that is, signal restoration,
provided in MOS ICs. It helped NMOS ICs to offer speed, to save power, and
to survive for a long time against CMOS. Standard CMOS required two comple-
mentary transistors per logic input variable. In order to improve the transistor
efficiency, various styles of dynamic CMOS logic were invented, where logic
functions are mapped on arrays of NMOS switches, and precharge, sampling, and
drivers are built in CMOS (examples are Domino and cascode voltage switch
logic, CVSL). The persistent drive towards lower voltages V (to save energy CV2)
was advanced significantly with differential CMOS logic such as HIPERLOGIC
in 1999, and this type of logic is seen as a must for robust logic, when the
fundamental variance of transistor parameters such as their “threshold” becomes
as large as the supply voltage for transistor lengths <20 nm.
7. We date today’s drive towards 3D integration of chip technologies back to 1968,
when IBM introduced large-scale flip-chip ball bonding to ceramic carriers as
the C4 (controlled-collapse chip connect) process. It was introduced industry-
wide 25 years later. The next significant contributions came from the Japanese
Future Electron Devices (FED) program, which promoted wafer thinning and
contacts going fully through the wafer, so-called through-silicon vias (TSVs), in
the late 1980s. It took until 2005 for these techniques to be finally adopted
mainstream, and they have now become a massive industry drive to keep pace
with market requirements and to benefit from the economic quantum jump that
TSV chip stacking provides for memory density.
8. The single-electron transistor (SET) was first proposed in 1987 as a nanometer-
size MOS transistor, in which the source and drain PN junctions have been
replaced by tunnel junctions such that the transistor channel becomes a quantum
box that can hold just one electron. In the search for the most advanced
nonvolatile memory, this would take us to the impressive milestone of just one
electron per bit. Among the read-out options for such a memory, one may be
reminded of the charge-coupled device (CCD) serial memories in the 1970s.
However, no practical prototype had been reported by 2010, so circuits of this
type will only become products in the 2020s, a horizon where they will have to
compete with practical nano-CMOS RAMs needing just a few electrons per bit.
Overall, we see that it took major innovations 25–35 years to make it onto the
market. With a horizon of 2020, we see that all candidate technologies should have
been implemented in convincing prototypes by now and that it may pay off to
revisit innovations of 25 years ago that appeared then too early for their time.
3.1 The Bipolar Transistor
Our starting line for the bipolar transistor is the Noyce patent [1] and the planar
Si bipolar transistor shown there (Fig. 3.2). It displays a junction-isolated NPN
transistor, for which we would have to imagine a top contact to the deep collector to
make it suitable for integration.
The first product milestone in 1960 was planar bipolar ICs, both analog and
digital, and the progress on Moore’s curve started with this technology. Digital
logic exploded with three families: diode–transistor logic (DTL), transistor–transistor
logic (TTL), and emitter-coupled logic (ECL). With well-organized customer
support and multi-volume textbooks, Texas Instruments (TI) and Motorola
dominated the market with TTL and ECL products. Digital products from
controllers to mainframes were the high-volume killer-apps around 1970, and TI
quickly became the world’s largest semiconductor manufacturer. The progress in
bipolar technology, as in any electronic device, depended on the critical transit time
of the electrons, here the time to traverse the neutral base of width WB, and on the
parasitics, in rank order: collector resistance, collector capacitance, base resistance,
and collector–base Miller capacitance. Decades of hard work have gone into this,
and a status of the mid-1980s is shown in Fig. 3.3 with the buried collector layer and
lateral oxide isolation.
With its emitter-last process flow, bipolar has been a fundamental headache for
integration. So many circuits are common-emitter type, which means that the
emitter is connected to ground, and since there is no better ground than the silicon
bulk crystal, the emitter should be first. It is worth remembering that IBM and
Philips once chose this disruptive route with integrated injection logic (I2L), a low-
power logic intended to compete with CMOS. But the deep emitter had a small
emitter efficiency, and switching was slow because of a deep, thick base. At that
time, bipolar technology had already lost the battle for large-scale integration
against MOS ICs, which surpassed the bipolar IC revenue in 1980.
Fig. 3.2 Cross section of a bipolar transistor after Noyce (1959) [1]. # USPTO
42 B. Hoefflinger
Fig. 3.3 NPN bipolar

transistor with buried
collector and oxide
isolation
In the standard bipolar process flow, a lateral PNP transistor is also available and
used in analog circuits. Because its base width WB is determined by lithography, its
switching speed is low, and a complementary bipolar technology did not evolve.
However, in the vertical NPN transistor, the critical base width is determined by the
difference between the depth of the base diffusion and that of the emitter diffusion,
and this configuration has been tailored continuously towards higher frequencies.
The maximum intrinsic frequency fi is the inverse of the base transit time tB:
2DnB
fi ¼ : (3.1)
WB2
The diffusion constant DnB of the electrons in the base is related to their mobility
m via the Einstein relation
DnB ¼ mn kB T=q: (3.2)
A high electron mobility and a short base have been the target for decades, and
heterojunction bipolar transistors are the optimal solution. In these transistors, the
base is a SiGe compound, in which the electron mobility is two to three times that in
Si. We can describe the state of the art as m ¼ 1,000 cm2 V1 s1 and a neutral,
effective base width of 50 nm, so that we obtain an intrinsic frequency limit of
2 THz.
Parasitic capacitances and resistances result in maximum reported frequencies of
over 500 GHz [2]. These have been achieved with 0.13 mm technologies, for which
the typical parameters are:
Emitter area: 0.13 mm 0.13 mm,
Max. current (HBT): 1.3 mA,
Collector area: 5 0.13 mm 8 0.13 mm,
Collector capacitance: 2.8 fF.
This results in a minimum intrinsic switching time of 2.8 fF/1.3 mA ¼ 2.15 ps,
which does not yet consider the collector series resistance and load resistance as
well as capacitance and inductance.
3.1.1 Scaling Laws for Bipolar Transistors
Scaling laws or rules for bipolar transistors are not so well known. However, there
is a body of modeling relations mostly derived from the Gummel–Poon transistor
model [3], and in the intrinsic frequency limit, (3.1), we have the first driving force
towards higher speed, namely reducing the base width WB. If we pursue this, we
have to observe a number of scaling rules in order to maintain the transistor
performance as measured by its current, voltage, and power gain.
The forward current gain
DnB NE LE
bF ¼ (3.3)
DpE NB WB
is determined first by the ratio of the diffusion constants for electrons and holes in
Si, which is 3. Then we see the product of emitter dopant (typically As) concentra-
tion and diffusion length of the minorities (holes). This product is practically
a materials constant for a Si technology, because the larger the majority doping
the smaller the minority diffusion length. For reference, we take this product as
1014/cm2. Now we see that, in order to maintain current gain, we have to raise the
base doping level as we shorten the base width. There are more reasons to watch the
number of acceptors in the neutral base NBWB, which is called the Gummel number.
The intrinsic voltage gain of the transistor in the common-emitter mode, deter-
mined by the output conductance, is characterized by the ratio of collector current
to Early voltage VA (an extrapolated voltage at which the collector current for
a given base current would be zero)
VA ¼ qNB WB WC =eSi ; (3.4)
where eSi is the dielectric constant of silicon. We obtain the available voltage
gain as
IC VA VA
AV ¼ gm =gCE ¼ ¼ ; where Vt ¼ kB T=q: (3.5)
Vt IC Vt
High voltage gain requires a large Gummel number and a large collector
depletion width, which means a high-resistivity collector. This is associated with
a large collector series resistance, causing slower speed. We also notice that the
larger Gummel number means a lower current gain (3.3) as a direct trade-off
situation. In fact, the available power gain appears as the figure of merit of a Si
bipolar technology:
q DnB
AP ¼ bF AV ¼ NE LE WC =Vt : (3.6)
eSi DpE
44 B. Hoefflinger
We see that the Gummel number has disappeared, and all other quantities are
material or basic technology parameters, the collector depletion depth being the
only variable with a choice. Inserting numbers, we find
1:6 1019
AP ¼ 3 1014 3 104 =0:025 ¼ 6 106 (3.7)
1012
as the figure of merit for our reference technology with a 3 mm collector depletion
typical for a 5 V transistor. Voltage scaling towards a shorter collector depletion
would mean a direct reduction of available voltage and power gain.
Another important technology and design parameter critical for scaling is the
current density range for high transconductance and current gain. It is determined
by the knee current. At that current level, the collector current density is so high that
the density of minority electrons entering the base reaches the acceptor density NB
in the base, and the transconductance drops by 50% (Fig. 3.4). This critical collector
current density is reached as the peak electron density rises to NB:
NB
JK ¼ qDnB : (3.8)
WB
At current densities above this level, the transconductance drops to gmK ¼

IC/2Vt, which occurs for scaled-down emitter areas at fairly small currents so that
such a transistor loses its edge as a high-speed load driver. Even for a heterojunction
transistor, where we may reach a diffusion constant of 25 cm2 s1, an effective
acceptor level of 1019 cm3, and a base width of 50 nm, the knee current density
log IC, IB
IK
Knee current
√IK IS IC
IB
IC
IS
βF UEB
Fig. 3.4 Semilog plots of base and collector currents as well as current gain showing the base-
current knee at low levels due to recombination and the collector-current knee at high levels due to
strong injection
occurs at 8 106 A cm2, which for our 0.13 mm 0.13 mm emitter means a
current of 1.36 mA.
Nevertheless, SiGe bipolar transistors offer terahertz performance and are the
active devices filling the terahertz gap. They provide the needed performance in the
specialty areas of broadband transceivers, millimeter radar and terahertz imaging.
As to the compatibility with mainstream IC processing: They require extra
processing steps and thicker epitaxial Si films at extra cost. More cost-effective,
high-speed CMOS technologies are infringing on the territory held by bipolar
today. We will address that again in Chap. 5 on interconnects and transceivers.
3.2 MOS Integrated Circuits
Our first milestone for MOS integrated circuits is the realization of an MOS
transistor by Kahng and Atalla in 1959 [4, 5]. The first products integrating
hundreds of these into digital circuit functions on a single chip appeared in the
mid-1960s, and the first killer products were the microprocessor and the random-
access memory (RAM). Among the many incremental improvements of the MOS
technologies, an early outstanding one was the replacement of the Al gate late in the
process by a deposited polycrystalline gate earlier in the process, where it served as
a mask for the diffusion of the source and drain regions, thereby self-aligning these
electrodes with the gate. Because this same polysilicon layer could also serve as a
first interconnect layer for local wiring, only four masks and a minimum number of
processing steps were needed for this technology. We illustrate this process with
Lego blocks in Fig. 3.5 on the basis of a more recent SOI structure.
The LegoMOS model presented here serves very well to demonstrate the
technology properties and constraints:
– The 2 2 brick marks the smallest feature possible with lithography.
– The 1 brick marks a thin layer, whose thickness is controlled by thin-film
growth or deposition.
– Each button will stand for roughly one acceptor (yellow or red) or one donor
impurity (blue) in our ultimate NMOS nanotransistor.
The starting material is a Si wafer, which has a buried-oxide insulating layer
(white) and a thin Si film with P-type doping on top (blue). The intended transistor
sites (active areas), of which we just show one, are masked against the thermal
oxidation of the rest of the surface, which is thereby converted to Si dioxide
(Fig. 3.5, top). The masking is then removed from the Si islands, and the thin
gate dielectric is grown on the wafer. The polycrystalline Si film is deposited next
(red), and the gates are defined by lithography and etched. These gate structures
(red) serve as masks for the implantation of N-type dopants (As), which produce the
source and drain electrodes (yellow) of the transistor (Fig. 3.5, middle).
An intermediate oxide is deposited next. On this, the contact holes are defined by
lithography and etched. The following vapor deposition of Al will also fill the
46 B. Hoefflinger
Fig. 3.5 Three steps in the manufacturing of a basic MOS integrated circuit. Top: Active area
(blue). Middle: Poly gate (red) and source/drain (yellow). Bottom: Contacts (black)
contact holes (black, Fig. 3.5, bottom), and this layer is structured by lithography
and etching into thin-film interconnect lines. The resulting topography, where the
n-type Si (yellow) allows interconnects as well as the poly layer and the Al layer, is
sufficient to wire any circuit function.
This four-mask process, active areas, gates, contacts, metal, tells the MOS
success story of low cost and high production yield, and it led to the race for
miniaturization: With higher-resolution lithography and etching, transistors could
be made smaller and faster and be packed more densely. In about 1970, a second
metal interconnect layer, separated by another deposited insulating layer, was
added. This required two extra masks, one for defining contact holes, called vias,
between the first and second metal layers, and a second mask for etching the second
metal layer. Today more than six layers of metal are used. Table 3.1 and Fig. 3.6
present a grand overview of the evolution of the critical dimensions of MOS ICs,
including those of the International Technology Roadmap for Semiconductors
(ITRS), which has many more details to be covered in a later chapter.
We continue here by analyzing the key performance criteria of MOS transistors:
switching speed and energy. We consider the ideal cross section of an NMOS
transistor on top of an insulating layer as shown in Fig. 3.7. As a parallel-plate
capacitor, it has a capacitance
eI
C0 ¼ LW; (3.9)
tI
Table 3.1 Evolution of the critical dimensions of MOS ICs

Year 1970 1980 1990 2000 2010 2020
L [nm] 20,000 2,000 500 120 32 8
tI [nm] 120 15 10 5.0 2.5 1.0
L/tI 167 133 50 24 13 8
Gain eeSiI LtI 2 54 42 15 6 2 0.7
100
(mm)
10
167 Gate Length L (mm)

1
133
10–1
ratio L / tI = 50
24
10–2
13
Insulator
Thickness tI (mm) 8
10–3
1970 1980 1990 2000 2010 2020
Year
Fig. 3.6 Critical dimensions of MOS ICs 1970–2020

48 B. Hoefflinger
Gate
++ + +
Oxide
+ – – +
–
+ – – +
Source – + Drain
+ – –
+ – – – +
+ – – +
+ – – – +
Fig. 3.7 Cross section of an NMOS transistor including charges
with the dielectric constant of the insulator eI, its thickness tI, the length L between
the metallurgical junctions of source and drain, and its width W. If we apply
a positive voltage VG, we find Q0 positive charges on the gate and a negative charge
Q0 in the silicon:
Q0 ¼ C0 VG :
With a voltage VDS applied between drain and source, we establish an electric
field E ¼ VDS/L pointing in the negative x direction. If all the negative charges were
mobile electrons, they would now move towards the drain with a certain velocity v.
Within the channel transit time L/v, the entire charge Q0 would be collected at the
drain, so that we would obtain the drain current
ID ¼ Q0 v=L ¼ C0 VG v=L: (3.10)
This is only correct for very small drain–source voltages. We have to consider
that the voltage difference between gate and channel at the drain end is only
VGD ¼ VGSVDS, so that the charge density decreases as we move along the channel
Q00 ðx ¼ 0Þ ¼ C00 VGS ;

Q00 ðx ¼ LÞ ¼ C00 ðVGS VDS Þ:
We introduce the average charge
hQi ¼ ½Q00 ð0Þ þ Q00 ðLÞLW=2;

eI (3.11)
hQi ¼ WLðVGS VDS =2Þ:
tI
Under the force of a moderate electric field E, mobile charges in semiconductors

move with a velocity proportional to this field v ¼ m0 E, where the mobility m is
about 250 cm2 V1 s1 for electrons in a Si MOS transistor. We take an effective
transit time for the charge hQi as

1=t ¼ v=L ¼ m0 VDS L2 ;
so that we obtain a more accurate drain current as
hQi eI 1
ID ¼ ¼ m0 ðVGS VDS =2ÞVDS : (3.12)
t tI L
This is the drain current according to the so-called Shockley gradual-channel

approximation. It reaches a maximum current as the drain voltage reaches the gate
voltage:
VDS ! VGS ;
m0 eI W VGS
2 (3.13)
ID0 ¼ :
tI L 2
It was already evident in 1970 that this current was larger than what was
measured in NMOS transistors. The reason is carrier-velocity saturation. The
maximum electron velocity in most semiconductors at room temperature is
vL ¼ 107 cm s1 ;
vL ¼ m0 EC ; (3.14)
m0
v¼ E:
1 þ E=EC
For a mobility of 250 cm2 V1 s1, the critical field is 40,000 V cm1. Here the
mobility is already reduced by 50%, and actual fields are much higher. Not only is
the maximum current reduced, but so is the drain voltage at which it is reached. For
a modeling of this effect, see [6] and [7]. We settle here for an optimistic estimate of
the maximum current. The most aggressive assumption is that all electrons in the
channel travel at the velocity limit:
ID1 ¼ Q0 vL =L;
eI (3.15)
ID1 ¼ WvL VGS :
tI
This optimistic extreme has a number of significant implications. The channel
length does not appear in this equation, removing one basic argument for scaling.
The outer limit for the transconductance would be
eI 0:4 1012
G01 ¼ vL ¼ 107 104 ¼ 4 mS mm1
tI 107
and the maximum current

0
ID1 max ¼ eI vL EImax ¼ 0:4 1012 107 5 106 104 ¼ 2 mA mm1 :
50 B. Hoefflinger
This simple assumption overestimates the velocity on the source side of the
channel. We now consider this and still simplify the situation by letting the field at
the drain end approach infinity in order to transport the vanishing charge there at
pinch-off. The maximum current becomes [7]
IDS max ¼ k ID1 ;

rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
EC L VGS (3.16)
k ¼1 1þ2 1 :
VGS EC L
This correction factor k is shown in Fig. 3.8.

So far, we have assumed that the NMOS transistor starts to conduct current as
soon as the gate voltage is larger than zero, in other words that its threshold voltage
VT is zero. However, for most practical applications, a positive threshold voltage is
required so that our model drain current only starts to flow when the gate voltage
exceeds this threshold. We introduce this excess voltage as the effective gate
voltage:
VGeff ¼ VGS VT :
In Tables 3.2 and 3.3 we assess the maximum channel charge, the maximum
current per micrometer width and the maximum current for a minimum transistor
with W/L ¼ 2 for all the past milestones and for the 2020 target, and we illustrate
these in Fig. 3.9.
1.0
IDS∞
VL C′0 W VG
0.5
0 2 4 6 8 10
μ0 VG V
= G
vL L Ec L
Fig. 3.8 Correction factor k for the maximum available current [7]
Table 3.2 Transistor parameters and NMOS transistor performance 1970–2020

Year 1970 1980 1990 2000 2010 2020
Length [nm] 20,000 2,000 500 120 32 8
Thickness [nm] 120 15 10 5.0 2.5 1.0
Supply voltage [V] 20 5.0 3.3 1.5 0.5 0.4
Threshold [V] 4.0 1.5 1.0 0.7 0.2 0.1
Table 3.3 Maximum current per micrometer width and intrinsic switching energy 1970–2010
Year 1970 1980 1990 2000 2010 2020
Channel electrons 2.7 107 5 105 3 104 1.2 103 60 15
Current [mA mm–1] 80 200 300 270 240 190
Switching energy [eV] 7 108 3 106 1.2 105 1870 50 3
Our assessment of the intrinsic maximum current and transconductance shows

that scaled-down Si NMOS transistors approach a limit dictated by the dielectric
constant of the gate insulator and the maximum electron velocity fairly independent
of the channel length, contrary to the decades up to 2000, where they improved
almost in proportion to the reduction of the channel length.
The minimum intrinsic switching time is
tS ¼ C0 VDD =IDS max ;

L VDD
tS ¼ ;
k vL VDD VT
offering a linear improvement as L is reduced. The intrinsic bandwidth is
k vL
f3dB ¼ ;
2pL
which, for k ¼ 0.6, VT ¼ 0 and a 10 nm channel, results in 200 GHz. The evolution
of bandwidth and switching times of NMOS transistors from 1970 with projections
for 2020 can be found in Fig. 3.10 and Table 3.4.
After this generic assessment of maximum available current at a maximum gate
voltage given by the breakdown field of the dielectric, typically 5 MV cm1, we
now take a closer look at the transistor channel. Figure 3.7 already tells us that
there is a more complex balance of charges in the transistor channel, in our case the
dominant normally-off NMOS transistor with a positive threshold voltage. This
requires an acceptor doping density NC [cm3], which we recognize as fixed
negative charges in the channel. On its source and drain side, we see the PN
junction depletion layers against source and drain set up as pairs of positively
charged donors in the source and drain layer and negatively charged acceptors in
the channel. In equilibrium (drain–source voltage zero), these depletion layers set
up potential barriers each 0.7 V high. The width of each depletion layer is
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 0:7 eSi
WD ¼ ;
qNC
which is 3 nm wide for a doping density of 1020/cm3. For a metallurgical channel

length of 8 nm, this leaves us with an effective channel length of 2 nm. In this
region, for a Si film 2 nm thick and a channel width of 10 nm, we find just
52 B. Hoefflinger
Fig. 3.9 Maximum channel 108

charge, current per
micrometer width, and 107
intrinsic switching energy
106
1970–2020
105
Max. Channel
Electrons
104
103
102
10
1
1970 1980 1990 2000 2010 2020
Year
400
300
Max. Current
mm Width
(mA / mm)
200
100
0
1970 1980 1990 2000 2010 2020
Year
Gate Length
20 μm 2 μm 0.5 μm 120 nm 32 nm 8 nm
109
108
107
Switching Energy (eV)
106
105
104
103
102
10
100 kBT
1
1970 1980 1990 2000 2010 2020
Year
1 THz 1 ns
100 GHz 100 ps
Switching Time
3dB Bandwidth
10 GHz 10 ps
Switching
Time
Bandwidth
1 GHz 1 ps
100 MHz 0.1 ps

1970 1980 1990 2000 2010 2020
Year
Fig. 3.10 Bandwidth and switching times of NMOS transistors 1970–2020
Table 3.4 Bandwidth and switching times of NMOS transistors 1970–2020

Year 1970 1980 1990 2000 2010 2020
Bandwidth [GHz] 0.7 4.0 10 40 100 200
Switching time [ps] 250 60 17 5 1.6 0.8
four acceptors. We show in Fig. 3.7 that it takes an equal number of charges on the
gate to compensate these acceptors, and the necessary gate voltage would be
the threshold voltage. Any higher voltage would now induce free electrons in the
channel. This would be the electrostatic model, which we have considered so far.
However, there are more forces present than we have identified as yet. As soon as
we apply any voltage between the gate and the source, we lower the potential
barrier between source and channel below the equilibrium of 0.7 V. As a conse-
quence, the electric field in the source depletion layer, which confined electrons to
the source so far, is reduced, and electrons can traverse the barrier and will be
collected at the drain under any applied drain bias. This current component is a
diffusion current, as in bipolar transistors controlled by the gate as an insulated
base. It is alternatively called leakage or subthreshold current. We will call it the
barrier-control mode of transistor operation, because not only the gate but also the
drain has an effect on this barrier.
In Fig.3.11, we show a short-channel NMOS transistor in which the potential
barrier FB is controlled by applied voltages VGVS and VDVS. Impurities are
54 B. Hoefflinger
Fig. 3.11 2D field and potential model for a short-channel transistor under barrier control
engineered in the channel such that there is a density N1 near the surface and N2 in a
buried layer. The height of the potential barrier can be modeled as [6]
FB ¼ AðVGS VT Þ þ BVDS CVDS ðVGS VT Þ: (3.17)
As we expect, the threshold voltage here is dependent on the drain voltage:

BVDS
VT ¼ VT0 :
A CVDS
We introduce
C0 þ Cd
n¼ ;
C0
rffiffiffiffiffiffi (3.18)
eSi tI N1
¼ ;
eI L N2
and we obtain simple and powerful expressions for the barrier dependence on gate
and drain voltage:
1 2
A¼ ;
n

B¼ ; (3.19)
n
4eSi
C¼ :
qN2 L2 n
At the threshold voltage, FB ¼ 0, and we define a maximum subthreshold

current Ic such that
ID1 ¼ IC expðFB =Vt Þ (3.20)
is the subthreshold or barrier-controlled current. The log of this current is plotted in

Fig. 3.12 vs the input gate voltage and the drain voltage. As we see, both voltages
exert control over the current, and a quantitative assessment of these effects
becomes essential.
In order to connect this current to the drift current (3.16) above the threshold
voltage, we set
1
IC ¼ IDS max ðVGeff ¼ Vt Þ;
e
and we obtain a continuous current from subthreshold diffusion type to above-

threshold drift type as [8]

IDSmax FB
Itot ¼ IC ln exp þ exp (3.21)
IC Vt
for FB < 0 and noting that ln(1 + x) ¼ x for x < 1.

56 B. Hoefflinger
VD6
VD5
VG5
VD4
VG4
IgID VD3 IgID VG3
VD2
VD1 VG2
VG1
VG1 VG2 VG3 VG4 VG5 VD1 VD2 VD3 VD4 VD5 VD6
VGS VDS
Fig. 3.12 The drain current of a short-channel MOS transistor in its subthreshold or barrier-
control mode
We now find that, for transistors shorter than 30 nm, the operating range of
<0.5 V means mostly the barrier-control mode. We derive the transconductance
from (3.17) and (3.20) as
gm ¼ ID1 =AVt ;
the output conductance as
gDS ¼ ID1 =BVt ;
and the available voltage gain as

rffiffiffiffiffiffi
eI L N2
AV ¼ gm =gDS ¼ A=B ¼ ð1 2Þ= ¼ 2: (3.22)
eSi tI N1
We see immediately that a high dielectric constant of the insulator together with
its scaling directly with the gate length are essential to maintain a gain >1, that the
channel should be doped inversely, which means that the deep density N2 should be
higher than N1 at the Si surface, and that we keep the depletion capacitance Cd in
(3.18) small compared to the gate capacitance.
We build a 10 nm reference NMOS nanotransistor with the following parameters:
L ¼ 10 nm
tI ¼ 3 nm
eI/eSi ¼ 1
N2/N1 ¼ 4
Cd/C0 ¼ 3/4
We obtain the following results:
Transconductance: 150 mV per decade
Output conductance: 700 mV per decade
Voltage gain: 4.7
ON/OFF current ratio: 27:1 for 0.4 V supply voltage
Max. current at 0.4 V: 20 mA, VT ¼ 0 V
Min. current at 0 V: 18 nA for VT ¼ 400 mV
Max. current at VGS – VT ¼ 25 mV: 0.5 mA
For this critical current at the top of the subthreshold region and the onset of the
drift-current region and any current smaller than this, there is on the average just
one electron in the channel: Our 10 nm NMOS nanotransistor is a single-electron
transistor.
For a comparison with published results [9], we study the characteristics in
Fig. 3.13. We recognize that these are mostly subthreshold characteristics with
transitions at the high-current sections into drift-current modes. All characteristics
show a strong dependence on the drain–source voltage as we expect from our study
10–3
Vds = 0.05, 0.10, ..., 0.40 V
10–4 Lg = 5 nm
Drain current / d (A μm–1)
10–5
10–6
10–7
10–8
10–9
–1.5 –1.0 –0.5 0.0 0.5 1.0 1.5
Gate voltage Vqs (V)
Fig. 3.13 Current–voltage characteristics of 5 nm NMOS and PMOS transistors after

Wakabayashi et al. (2006) [9]. # 2006 IEEE
58 B. Hoefflinger
of the barrier-control mode. The analysis of the characteristics shows for the two
transistor types:
NMOS Transistor:
Transconductance: 320 mV/decade
Output conductance: 320 mV/decade
Voltage gain: 1
PMOS Transistor:
Transconductance: 270 mV/decade
Output conductance: 320 mV/decade
Voltage gain: 1.18.
These are very informative and instructive results: At L ¼ 5 nm, it becomes very
difficult to build useful MOS transistors.
Our models as described by (3.18) and (3.22) tell us that a higher dielectric
constant and a thinner gate dielectric would be needed at the cost of further
reducing the operating voltage, and careful channel engineering would be needed.
On top of this squeeze, we have to realize that anywhere below 20 nm we operate
not only with serious manufacturing tolerances, but with the fundamental variance
of numerical quantities: Any quantity with a mean value N has a variance N and a
standard deviation s equal to the square root of N.
3.2.1 The Fundamental Variance of Nanotransistors
We have noticed already in Fig. 3.7 that, in the MOS transistor channel, we deal
with a very sensitive balance of charges, which becomes a matter of small numbers
at nanometer dimensions. In our reference nanotransistor, we have carefully doped
the channel with a density of 1020 acceptors/cm3 near the source and drain metal-
lurgical junctions.
In equilibrium, at this density, 36 acceptors set up the source and drain depletion
layers, each 3 nm wide. In the remaining 4 nm channel region, we place five
acceptors, corresponding to 2 1019 acceptor volume concentration, to control
the threshold voltage and the gain of the transistor: N1: one acceptor near the Si
surface; N2: four acceptors underneath in the 3 nm Si film. No matter how precisely
we have processed, these five acceptors have a fundamental statistical variance of 5
and a standard deviation s of two acceptors. In this statistical distribution, about
90% of all values lie within +2s and 2s of the mean value 5. If we settle for this
range as an acceptable design space, the low threshold voltage would be given by
one acceptor and the high threshold voltage by 9 acceptors. Our gate capacitor has a
capacitance of 8 1018 F, which requires 20 mV per elementary charge. This
results in a fundamental threshold range of our transistor of threshold voltage high
(9q) threshold voltage low (1q) ¼ 160 mV.
Can we build useful circuit functions with this? The answer is yes, if we can allow
currents per gate to vary by orders of magnitude. However, the situation gets even
tougher fundamentally when we consider the variance of the transconductance.
We have used channel engineering, Fig. 3.11 and (3.18), to optimize the
subthreshold slope (on/off current ratio)and the voltage gain, and we implemented
that in our reference transistor with N2/N1 ¼ 4. Counting acceptors, we allocated
4 deep and 1 near the surface out of our mean total of 5 ions. The spread of N1 + N2
from 1 to 9 can result in a maximum N2/N1 ¼ 8 and a minimum N2/N1 ¼ 1. Using
our model, we obtain the following ranges:
Transconductance high: 136 mV/decade, gain high: 7.2
Transconductance low: 300 mV/decade, gain low: 1.3.
The resulting range of transistor characteristics is shown in Fig. 3.14.
5 μA
VDS = 0.1 V
1 μA
VT = 0.2 V
300 mV / dec
Av = 1.3
Current
n
ea
100 nA
M
VT = 0.36 V
136 mV /dec
Av = 7.2
10 nA
1 nA
0 0.1 0.2 0.3 0.4
Gate Voltage (V)
Fig. 3.14 The 90% corridor of current vs gate-voltage characteristics at VDS ¼ 0.1 V for our
10 nm reference nanotransistor
60 B. Hoefflinger
This technology design space with a fundamentally large spread of threshold

voltages, a transconductance range of 2:1, and gains as low as 1.3 requires robust
circuit techniques, which we will cover in the next section, where we shall use
exclusively the availability of both NMOS and PMOS transistors.
We mention here that all of the above modeling equations, which we developed
for NMOS transistors, can be used to describe PMOS transistors provided we
remember that the mobility of holes is about 1/3 that of electrons normally and
that it can almost reach that of electrons in Si strained by the introduction of Ge.
Historically, PMOS ICs were first, because it was easier to engineer the
(negative) threshold voltage for the necessary normally-off transistors. In order to
benefit from the higher mobility of electrons, a lot of energy went into controlling
the channel impurities in NMOS transistors so that they could be made normally-
off. The biggest boost for this impurity control came with ion implantation [10] of
the dopants, which was demonstrated for MOS first by Mostek in 1970. NMOS
became the basis for killer products such as microprocessors and memories with
such impressive growth that MOS ICs passed bipolar in world-market share in
1980, and Nippon Electric Corporation (NEC) passed TI to become the world’s
largest semiconductor company. The investment in NMOS computer-aided design,
circuit libraries, and process technology became so large that the major players kept
their emphasis on this technology well past 1980, while the more radical comple-
mentary MOS (CMOS) technology was gaining ground steadily for reasons that we
will assess in the following section.
3.3 CMOS and BiCMOS Integrated Circuits
The CMOS process and circuit as invented in 1963 by Wanlass, [11, 12] and shown
in Fig. 2.7, derives its success from the features of the pair of complementary
transistors as illustrated in Figs. 2.8 and 3.15.
1 VDD
2
V0
3 VTP
VI V0
VTP VTN
VTN 3′
VTN
4
VTP 5 VTP
VDD1 VDD2
Fig. 3.15 Circuit diagram (right) and transfer characteristic (left) of the CMOS inverter
Table 3.5 Properties of PMOS and NMOS transistors

Charge carriers Threshold ON voltage OFF voltage Drain–source voltage
NMOS Neg Pos Pos Neg Pos
PMOS Pos Neg Neg Pos Neg
The complementary properties and parameters of NMOS and PMOS transistors

are listed in Table 3.5, and from these we obtain, somewhat simplified, the transfer
characteristic in Fig. 3.15 for the inverter. It has three sections.
1. For input voltages VI between zero and the threshold of the NMOS transistor, the
output voltage is at the positive supply voltage, because the PMOS transistor is
turned on and charges the output to the supply voltage, uncompromised by the
NMOS transistor, which is turned off below threshold.
2. For high input voltages beginning at a threshold voltage below the supply
voltage, the output voltage is zero, because now the NMOS transistor is turned
on and discharges the output to zero, while the PMOS transistor is turned off and
does not compete.
3. Between these perfect HI and LO levels, we have a transition region, which we
have simplified here with a constant slope, namely the voltage gain of the
inverter. If indices N and P denote NMOS and PMOS respectively, this gain is
gmN þ gmP
AV ¼ : (3.23)
gDSN þ gDSP
Equation 3.23 shows that the combination of the transistor types establishes gain
safely, although one type may even have an internal gain <1. This will be important
when we deal with variances at the nanometer scale.
Historically, the magic of CMOS arose from the attributes that we see in
Fig. 3.15.
– CMOS provides the perfect regeneration amplifier: Although the input may be
less than a perfect HI or LO, the inverter output provides a perfect LO or HI,
respectively.
– A CMOS circuit ideally only draws a current and dissipates energy while it is
switching. As soon as signal levels are stable, there is no current flow.
– CMOS circuits in subthreshold mode operate robustly and dissipate very little
power. Supply voltages can be lower than the sum of the threshold voltages for a
very low energy CV2.
– CMOS is tolerant to large differences in on-resistance between PMOS pull-up
and NMOS pull-down transistor networks: PMOS and NMOS transistors can
have unit size irrespective of their individual function. This allows very-large-
scale design automation.
In spite of these significant advantages over single-transistor-type PMOS or
NMOS circuits, which require tailored ratio design per gate type, and which have
marginal gain, signal swings less than the supply voltage, and permanent power
62 B. Hoefflinger
Fig. 3.16 Transistor diagram

of a CMOS/bipolar inverter
after [13]
dissipation, CMOS started out in two niches only: digital watches, immediately
controlled by Japan, and radiation-hard circuits (small-Si-volume SOI, large noise
margin).
Because channel lengths were determined by lithography and significantly
longer than the base widths of bipolar transistors, switching speeds were slow
(Fig. 3.10), NPN bipolar drivers were proposed in 1969, [13] and Fig. 3.16, to
improve the switching speed, particularly for large capacitive loads:
IL ¼ bF IDS ;
where bF is the current gain (3.3) of the bipolar transistor. However, such
technologies initially were very expensive, and it proved to be difficult to make
good NPN transistors with optimum parameters such as given by (3.3–3.6) together
with optimized MOS transistors so that progress was slow. About 10 years later,
BiCMOS drivers had become economical with all-ion-implanted technologies,
further assisted by lateral oxide isolation and, finally, by SOI since about 2000.
We should still bear in mind that the bipolar transistor first has to handle its own
collector RC load before handling the external load and that a properly designed
pair of CMOS inverters is often competitive: If we measure the load as F times that
of the internal capacitance of the first inverter with its channel width W, then the
second inverter designed with a width of WF1/2 delivers the optimum speed.
Moving back to digital CMOS logic, we highlight another unique feature: The
regular topography of CMOS, which led to design automation, gate arrays, and
sea-of-gates. A pioneering effort was made in 1974 by a group at Philips who
recognized [14] that CMOS logic could be designed very effectively with
– Unit-size PMOS and NMOS transistors,
– Regular standard placement of rows of PMOS and NMOS transistors,
– Wiring channels between these rows and
– Highly automated generation of standard logic cells, as well as

– Cell placement and routing for customized functions.
This local oxidation CMOS (LOCMOS) gate array technology of 1974 became
the prototype of the CMOS gate array strategies, which offered a quantum jump in
IC design productivity and the economic realization of large-scale application
(customer)-specific ICs (ASICs).
As wiring became the number-one cost factor compared with the cost of
transistor sites on-chip, an endless sea-of-transistors, also called a sea-of-gates,
was put on the chip with many degrees of freedom for dense wiring on top.
Another important step was the simplification and optimization of CMOS and
CMOS/bipolar manufacturing with the extensive application of ion implantation as
a precision process for a wide range of impurity profiles, a room-temperature
process with a low-thermal-budget anneal. This led to an all-ion-implanted
CMOS/bipolar process meeting all the specific requirements with a minimum of
processing steps [15]. This process, offering optimum collector-diffusion-isolated
(CDI) bipolar transistors together with CMOS needed only the following steps:
1. N-type implant for CMOS N-well and N-collectors,
2. active areas for transistors,
3. boron implant for NMOS-, PMOS-, and field threshold,
(a) Boron implant for active base (Gummel number, (3.3, 3.4),
4. Gate,
5. NMOS source/drain,
6. PMOS source/drain,
7. Contact,
8. Metal 1.
At the time, the second half of the 1970s, CMOS had also made its entry into
low-power analog circuits, and mixed-signal analog-digital circuits including
analog–digital converters and filters meant the start of digital telephony and the
dawn of mobile communication.
The CMOS-compatible realization of lateral high-voltage MOS transistors
(double-diffused MOS (DMOS)) with operating voltages of hundreds of volts
[7, 16] as well as vertical high-voltage, high-current transistors (power MOS) led
to the first systems-on-chip (SOC).
Regarding the CMOS innovation process in the 1970s, university labs with a
pilot line for prototyping new circuits, for example
– Stanford University, Palo Alto, CA: Prof. James D. Meindl and Jaques
Baudouin,
– University of California, Berkeley, CA: Prof. Paul R. Gray and Prof. David A.
Hodges,
– Katholieke Universiteit Leuven, Belgium: Prof. Roger Van Overstraeten,
– University of Dortmund, Germany: Prof. Bernd Hoefflinger played a key role.
64 B. Hoefflinger
In today’s language, at the 10- to 5-mm technology nodes, these labs were able to
prove new processes and circuit concepts, providing a critical arsenal of CMOS
technology [7], which was widely publicized and exploited in Japan in the context
of the Japanese
10: VLSI (Very-Large-Scale Integration) program.
The first low-power, high-performance microprocessors and digital signal

processors were announced by Japanese companies in 1979/1980 at about the
3 mm node, including all high-speed static random-access memories (SRAMs).
The US had missed large-scale entry into CMOS because of its emphasis on
NMOS dynamic memories (DRAMs) and high-speed (high-power) bipolar and
NMOS processors. The race to catch up started in 1981. TI, Digital Equipment
Corporation (DEC), HP, AT&T and Motorola were the first to push CMOS, and
CMOS passed NMOS in world-market share by 1985 (Fig. 3.1).
Logic-circuit styles have seen a lot of differentiation, and we will cover some
important ones in Sect. 3.6 on dynamic and differential logic. Here we now
concentrate on the future potential of complementary MOS circuitry at the nano-
scale, and we assess this guided by the fundamental variance of nano CMOS
circuits.
In Sect. 3.2.1 we assessed the fundamental variance of transistor characteristics
at the nanoscale as demonstrated by the 90% corridor of current vs gate voltage for
our 10 nm nanotransistor in Fig. 3.14. We now study the consequences of this
variance at the transistor level on our fundamental key circuit, the complementary
inverter and its transfer characteristic, which would look ideally like Fig. 3.15. We
learned there the requirements for this robust amplifier-regenerator: finite threshold
voltages and an internal voltage gain >1, (3.23). We found for our reference
transistor
– 90% range of threshold voltages: 160 mV,
– 90% range of voltage gain: 1.3–7.2.
Let us assume that we can succeed in making PMOS transistors symmetrical to
NMOS transistors, i.e., the same threshold range, and an inverter gain of 2:
gmN ¼ gmP ¼ 150mV=decade;
gDSN ¼ gDSP ¼ 300mV=decade:
The corridor of transfer characteristics would then be as shown in Fig. 3.17,

where the envelope is given on the left by a zero threshold of the NMOS and
160 mV of the PMOS, and on the right by 160 mV for the NMOS and zero for the
PMOS transistor.
The envelopes stand for those pairs where we would just achieve a full swing
from the positive to the negative supply voltage. These extreme cases are
Fig. 3.17 Voltage-transfer a

characteristics of the left 370
(black) and right (red)
branches of the cross-coupled 300
inverter pair and the two NM
stable states. (a) Nominally mV NM
symmetric. (b) Variance
causing worst-case offsets. OUT 200
NM: Noise margin IN
NM
100 NM
0
0 100 200 mV 300 370
OUT
IN
b
370
300
mV
OUT 200 NM
IN
NM
100
0
0 100 200 mV 300 370
OUT
IN
VTN1 ¼ VTN2 ¼ 0 $ VTP1 ¼ VTP2 ¼ high;
VTN1 ¼ VTN2 ¼ high $ VTP1 ¼ VTP2 ¼ 0:
As a function of the threshold variance and the average gain, the minimum
supply voltage for a full signal swing from within a Vt of the positive and negative
supply voltages, often called a rail-to-rail swing, is
VDD min ¼ 2Vt þ DVT ð1 þ 2=AV Þ: (3.24)

In our case, this supply voltage would be 370 mV.
In Fig. 3.17, we have also entered the static noise margin NM, the largest
distortion on a logical ONE or ZERO that would be relaxed by the cross-coupled
inverter pair, the quad, to a perfect ONE or ZERO. If the inverters are symmetrical
(a), this noise margin, relative to VDDmin, is
66 B. Hoefflinger
NM 1
¼ 0:5 : (3.25)
VDD min 2AV
In the symmetrical case, this noise margin would be 25% for both ONE and
ZERO. In the marginal case (b), one noise margin is 0, and the other is 50%.
The lesson of this exercise is that properly engineered symmetrical 10 nm
CMOS nano-inverters can serve as amplifying signal regenerators and drivers.
However, it becomes clear as well that the single nano-inverter will be plagued
by asymmetry both in its transfer characteristic and in its drive capability and that at
this level of miniaturization
– The cross-coupled CMOS pair (the quad) becomes the most robust and the
fastest voltage regenerator, and
– Fully differential signaling/logic can result in double the capacitance, but safe
operation at half the operating voltage means a 50% net reduction of the
switching energy CV2.
The quad is key to many practical and robust nanoelectronic functions:
– Fully differential signal regeneration,
– Differential amplification,
– Low-voltage, high-speed memory.
The quad has long been the heart of the memory cells in CMOS static random-
access memories (SRAMs). The six-transistor cell in Fig. 3.18 shows this cell with
fully differential read and write capability.
This cell is suitable for low-voltage standby because of the immunity inherent in
the storage of both the true and inverted data. How low can this standby voltage be?
The true and inverted voltage levels have to be separated by at least 2Vt ¼ 50 mV at
Fig. 3.18 Six-transistor CMOS SRAM cell for differential, nondestructive write and read
room temperature. To this we have to add the variance of the threshold voltages,
giving a standby voltage of 210 mV for our 10 nm target technology.
The voltage difference between ONE and ZERO is about 50 mV, and the
switching energy to flip the content of this cell would correspond to moving 2.5
electrons from one pair of gates to the other: The 10 nm CMOS SRAM cell in
standby is a two-electrons perbit memory for random-access, high-speed write and
nondestructive read.
The price for this effective type of memory is the standby current. It is, to first
order, the geometric mean of the maximum and minimum transistor current for a
given supply voltage:
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Istandby ¼ Imax Imin :
This amounts to 3 nA for our cell, or 3 A and 600 mW for a 1 Gbit 10 nm SRAM
in standby. We treat the trade-offs between DC and dynamic-switching power in
SRAMs further in Chap. 11.
We will discuss CMOS logic-options at the 10 nm node in Sects. 3.5 and 3.6.
Here we summarize our assessment of CMOS technology and circuits on their
move towards 10 nm:
– 10 nm CMOS transistors are dominated, like all semiconductor devices with
functional volumes <1,000 nm3, by the fundamental variance of the numbers of
charge carriers and doping atoms involved in the electronic functions.
– MOS field-effect transistors <50 nm operate no longer in the familiar (Shockley)
drift mode but in the diffusion or subthreshold mode, where the barrier for mobile
charges is controlled by both the gate and the drain, resulting in small and sensitive
intrinsic voltage gains of transistors.
– Variance and intrinsic gains marginally >1 require complementary NMOS/
PMOS circuitry, in particular robust signal regenerators and accelerators.
– Towards 10 nm, regenerators will have to contain CMOS quads, cross-coupled
complementary pairs, and fully differential signaling and logic will be needed.
– The push for 10 nm transistors means supply voltages <400 mV, heading
towards 200 mV, attractive for high-density memory and mobile systems.
Allowances for speed and signal integrity can mean diminishing returns.
In spite of all these concerns, CMOS is the only robust, large-scale, and
dominant chip technology for the next 20 years.
3.4 Silicon-on-Insulator (SOI) CMOS Technology
The ideal planar integration of CMOS integrated circuits was proposed and realized
more than 35 years ago, as we saw in Sect. 2.4, and as illustrated in Fig. 2.5. Lateral
and vertical oxide isolation of the complementary PMOS and NMOS transistors
allows the highest integration density and the smallest parasitic capacitances in
68 B. Hoefflinger
order to reduce the switching energy CV2. We follow the lengthy evolution of this
SOI technology. Although its implementation was shown virtually overnight in
1964 [17, 18] by Si epitaxy on sapphire, it acquired the label of being expensive and
difficult, and it remained a very small niche technology for military and space
applications, albeit meeting extreme reliability requirements.
It took 15 years for SOI to receive renewed attention. A lab in Japan and
the Stanford Electronics Laboratory [20] deserve the credit for launching the SOI
wave of the 1980s. Izumi and coworkers [19] realized the film sequence silicon–
insulator–silicon substrate by high-dose, high-energy implantation of oxygen
into silicon to produce a buried oxide layer, of course with a subsequent very-
high-temperature anneal to create a stoichiometric buried silicon dioxide layer and
to recrystallize the top Si film amorphized by the implant. Further milestones are
listed in Table 3.6.
Table 3.6 Selected SOI milestones

Year Milestone Reference
1964 Silicon-on-sapphire (SOS) [17, 18]
1978 SIMOX: Silicon isolation by implanting oxygen [19]
1979 Laser recrystallization of polysilicon deposited on silicon dioxide [20]
1981 FIPOS: Full isolation by porous silicon [21]
ELO: Epitaxial lateral overgrowth [22]
1985 Wafer bonding [23]
1994 ELTRAN: Epitaxial-layer transfer [24]
1995 Smart cut: H implantation, bonding, and splitting [25]
2005 SOI foil [53]
a b
Double porous Epitaxial silicon
silicon layer
Handle wafer
Bonding d
c
Seed wafer
SiO2 Nozzle
Handle wafer Handle wafer
Fig. 3.19 SOI wafer obtained by the transfer of an epitaxial Si layer (after [26]). A: Formation of a
two-level porous Si layer on the seed wafer. B: Growth of the epitaxial Si layer on the seed wafer.
C: Wafer bonding of the seed wafer to the handle (product) wafer. D: Splitting the wafers.
Polishing, etching, and possibly a smoothing epitaxy step for the handle (product) wafer.
E: Surface cleaning of the seed wafer
A more detailed description of these processes can be found in [26]. We focus

here on the quality of the Si films and on their economics for large-scale
manufacturing. As a rule, the surviving SOI technologies for ICs have a high-
quality crystalline film grown by Si epitaxy. That feature is available in epitaxial
lateral overgrowth. We cover that in the following section on 3D CMOS. Epitaxial
film is also available in processes following the ELTRAN scheme, which is shown
in Fig. 3.19.
3.5 3D CMOS Technologies and Circuits
For 50 years, integrated circuits have been 2D arrangements of transistors and other
components, such as resistors and capacitors, in a thin surface layer of a crystalline
Si wafer. Recently, 3D integration has become popular in the sense of stacking
chips on top of each other. We address that in Sect. 3.6. Here, we take a dedicated
look at the history and the future potential of stacking transistors on top of each
other with high density and functionality within the same Si substrate.
Early efforts in this direction were made with regard to bipolar transistors and
using PN junction isolation, but without success. Gibbons and Lee deserve the
credit for launching true 3D functional integration for CMOS in 1979. We marked
that as a significant milestone in Chap. 2 (Fig. 2.12, [20]) and we highlighted the
Japanese FED program that followed shortly thereafter. All these related efforts
suffered from the fact that only the bottom crystalline transistor layer offered high-
quality transistors; the higher layers consisted of poorer, recrystallized Si, produc-
ing an inferior transistor quality. The sole high-quality crystalline films on silicon
dioxide as an insulator were those obtained by seeded, selective epitaxial lateral
overgrowth (ELO) [22]. Because the lateral growth distance meant about as much
useless vertical growth and the seed meant a short to the underlying Si layer, this
technique did not receive much attention.
However, if one uses the seed as a vertical contact (a via) between devices
stacked on top of each other, it turns into an advantage rather than being
a shortcoming. This was proposed as a concept at one of the FED meetings [27]
with a highly compact full-adder as an example (Fig. 3.20). The real consequence,
namely to use as many seeds for selective epitaxial growth as there are CMOS
transistor pairs, was expressed in 1984 [28], and a practical process patent applica-
tion was filed, as illustrated in Fig. 3.21 [29].
Coordinated research between Purdue University and the Institute for Micro-
electronics Stuttgart, Germany, in the following years produced a number of
remarkable results based on localized epitaxial overgrowth for 3D CMOS:
– A top gate above the thin-film PMOS transistor provided a mask for the PMOS
source and drain, and it made this transistor a dual-gate transistor with a top and
a bottom channel in parallel. It was shown that this could more than double the
70 B. Hoefflinger
a
S
C C C
Qc
B B B B
A A A A
GND GND
Transistor Diagram
b A c
B GND
C
n+ S
AB n+ P n+
n+
n+ P n+
(A+B)C n+
n+ P n+
Qc n+
n+ P n+
Qc(A+B+C)
GND
X
X′ Cross section XX′
ABC
Qc S
Top View 2.5 × 4 ≈ 10p2
Fig. 3.20 Concept of a 3D-integrated full adder with stacked NMOS transistors [27]
PMOS transistor current to levels almost equal to the underlying NMOS transis-
tor with the same width (Fig. 3.22) [30].
– Independent operation of the two PMOS gates provided a logic OR function with
a perfect, very-low-leakage off state (Fig. 3.23)
– The PMOS OR function together with an additional NMOS transistor at the
lower level provided a two-input NAND gate with a very small footprint
(Figs. 3.24 and 3.25).
The topography of the stacked complementary transistor pair led to a primitive
cell that, beyond the inverter, offers the function of a selector (Fig. 3.26), which can
Fig. 3.21 Manufacturing of a 3D CMOS transistor pair with a common gate and common drains
[29]. Top: After formation of the bottom NMOS transistor and localized epitaxial overgrowth
(LOG) with the transistor drain as the seed. Bottom: After formation of the top thinned epitaxial-
film PMOS transistor with its drain connected to the bottom NMOS transistor. The selective
epitaxy acts here as self-assembly
Fig. 3.22 NMOS and PMOS

transistor currents in a 3D
CMOS pair with dual gate
PMOS. # 1989 IEEE
serve as a universal basic element for automatic logic synthesis with the benefit of a
very dense, highly regular physical layout [31].
A part of the 3D strategy was to provide the lower supply voltage from the Si
bulk, where a chip-size PN junction serves as a large charge store to reduce ground
bounce and to provide a ground contact through a “sinker” where needed, for
example, with the NMOS source for the inverter. Numerous 3D test circuits, both
full-custom and sea-of-gates type, were designed. The comparison of these and of
72 B. Hoefflinger
TG-BG Measured BG-TG Measured

10–3 10–3
3D CMOS LOG2 3D CMOS LOG2
10–4 10–4
10–5 10–5
10–6 10–6
10–7 10–7
ID [Amps]
ID [Amps]
100 nA 100 nA
10–8 10–8
10–9 10–9
10–10 10–10
10–11 10–11
10–12 10–12
10–13 10–13
6.0 4.0 2.0 0 –2.0 –4.0 –6.0 6.0 4.0 2.0 0 –2.0 –4.0 –6.0
VG [Volts] VG [Volts]
VB = –5.000 V To 0.000 V VB = –5.000 V To 0.000 V
Poly2 Poly1
Fig. 3.23 Currents for different poly-gate voltages on the dual-gate PMOS transistor. # 1989 IEEE
Fig. 3.24 Scanning electron microscopy (SEM) image of the cross section of a two-NAND gate.
The stacked dual-gate PMOS on top of the NMOS transistor can be seen on the right. LOCOS:
Local oxidation of silicon. # 1992 IEEE
Fig. 3.25 Layout comparison of 2D and 3D CMOS inverters and two-NANDs for equal design
rules. # VDI-Verlag 1993
Fig. 3.26 Functions of the 3D complementary pair with dual-gate PMOS transistor. # VDI-Verlag
1993
4LMin 4LMin
Metal Metal Metal
Oxid Oxid 0, 2
p n p+ p+ p+ n p 0, 1
p+ p+
Gate Oxid Gate 0, 2
n+ n p n+ p n n+ 0, 1
Oxid Oxid 0, 3
n+
Fig. 3.27 Cross section of a 3D CMOS quad in SOI, the fundamental building block for high-
performance nano-CMOS. # WEKA-Medien 2000
34 standard cells for the same design rules provided the average result [32] that 3D
CMOS with localized epitaxial overgrowth requires only 50% of the area as 2D
CMOS with the same mask count.
3D CMOS is even more effective in the merger of two complementary pairs, the
CMOS quad mentioned in the previous section as the universal core of CMOS
differential amplifiers, regenerators (Fig. 3.27), and storage elements such as
SRAM. In Fig. 3.28, we see such a memory cell with an additional LOG epitaxial
thin-film transistor level for the NMOS word-select transistors [32].
The density potential of 3D CMOS logic and its compatibility with computer-
aided design (CAD) automation were critically evaluated in the late 1990s and their
significance clearly demonstrated [33]. However, the rapid progress in scaling 2D
CMOS circuits met the density requirements until recently so that the 2D evolution
74 B. Hoefflinger
Fig. 3.28 Circuit diagram and cross section of a 3D CMOS SRAM cell with three levels of
transistors, including dual-gate PMOS. # VDI-Verlag 1993
has ruled, and disruptive 3D structures remained on the sidelines. Other alternative
introductions of the third dimension in vertical transistors, FinFETs, tri-gate
transistors, and carbon nanotubes with wrapped-around control gates have been
introduced in 2011 as 22 nm Tri-gate transistors by Intel. One can find surveys of
these, for example, in [34–36].
Will the third dimension arrive, finally and in a disruptive fashion? The answer is
• 3D stacking of chips and wafers arrived, fairly disruptively, in 2005, to meet the
density requirements of the market (more in the following section);
• 3D stacking of CMOS transistors on-chip will follow in view of the 10 nm
barrier.
What are the requirements for future 3D CMOS?
– Compatibility with mainstream Si CMOS technology
– No increase in total process complexity
– No disruption of the physical-layer circuit design style
– Large-scale vertical Si interconnect by selective growth to replace contacts and
metal
– High quality of the upper-level thin-crystalline-film transistors
There will certainly appear alternatives to the selective-epitaxy technology
presented here. It merits, however, to be reconsidered: What required growth of
10 mm Si mountains and their thinning-back by chemo-mechanical polishing
(CMP) almost 20 years ago, will now be <1 mm with a much smaller thermal
budget, and selective molecular-beam epitaxy (MBE) could be considered [37], as
well as atomic-layer epitaxy (ALE).
3.6 MOS Low-Voltage Differential Logic
MOS logic styles evolved rapidly in the 1960s and were mostly built on the
availability of one transistor type, which was initially P-type. The gates functioned
as voltage dividers, which required ratio design with narrow, long, high-resistance
load transistors and wide, short switching transistors. Designs varied largely
depending on designer skills and discipline. This situation improved when deple-
tion-type load transistors became available in about 1970, after the introduction of
ion implantation [10] allowed precise threshold control.
While this type of logic was pretty much a replica of bipolar logic with MOS
transistors, a radical alternative exploited the MOS transistor as a switch with a very
high off-resistance and a large on–off resistance ratio, and, at the same time, a high-
quality capacitor to store the information as a charge. As a consequence, various
styles of clocked MOS logic were introduced in the 1960s with distinct phases for
precharging the output nodes of a gate, followed by a sampling phase, in which the
status of the gate was evaluated. Unit-size transistors could be used, and only the
standard enhancement-type was needed. Complex series–parallel logic switching
networks could be used, generally with reduced total network capacitance and
higher speed.
In our chart of milestones (Fig. 3.1), we note 1980 as the year of killer products
based on dynamic MOS logic. Microprocessors and signal processors used high-
density, high-speed complex dynamic logic to achieve performance. The basis was
domino logic [38], followed by several enhancements to deal with speed problems
caused by the resistance of high-fan-in, series-connected transistors and the asym-
metry of pull-up and pull-down times of gates.
This led to the use of fully differential input trees, which were then connected to
differential amplifiers, which
– Restore full signal swing,
– Deliver perfect TRUE and COMPLEMENT outputs,
– Suppress common-mode asymmetry and offset noise,
– Can drive large-fan-out loads.
In some versions, the construction appeared in the name, such as DCVSL
(dynamic cascode voltage switched logic) [39].
The no.1 field for complex transmission-gate logic is adders and multipliers. In
particular, long carry chains with short overall delays have seen continuous
inventions and improvements, many dependent on (low-voltage) fully differential
signaling (LVDS) for safe operation with sufficient noise margin and fast detection
with differential amplifiers, eliminating common-mode voltages, crosstalk and
other noise, and generating full-swing complementary outputs with high speed.
One implementation of a “Manchester carry chain” is shown in Fig. 3.29. In the
center we see the cross-coupled pair, the quad, as the heart of the amplifier/
accelerator. It assures the safe and high-speed propagation of the carry signals in
spite of a supply voltage of only 0.5 V, as shown in Fig. 3.30 [40–42].
Arithmetic units are those logic circuits where density, speed, and energy in the
form of the number of operations per unit of energy (watts seconds ¼ joules) or
operations per second per watt (OPS/W) matter most. It is the show-stopper for
media-rich mobile electronics. Watching the energy CV2, we have to minimize the
circuitry (¼C) – and the voltage V by any means because of its quadratic effect. For
the extremely low voltages of 300500 mV, differential logic has become a must.
76 B. Hoefflinger
Fig. 3.29 One stage of a VDD

differential Manchester carry Cout Cout
chain (MCC) with a Quad
differential regenerator/ Propagate Propagate
accelerator output [40–42]. Cin Cin
The switches are VT ¼ 0
transistors Kill Generate
GND
Fig. 3.30 Generation of the most-significant-bit carry and sum of a 16-bit HIPERLOGIC adder in
a 100 nm SOI CMOS technology [40–42]. # 2000 IEEE
What can be done about the required circuitry? Beyond the optimization of
adders, the key problem is the multiplier. The raw estimate says that the task of
multiplying two N-bit words has a complexity of the order O(N2). A well-known
improvement on this is Booth encoding, which reduces the complexity to O(N2/2).
The result is normally accumulated to a length of 2 N bits, which is overkill in all
signal-processing applications, because, if the inputs have N bit accuracy, the
accuracy of the multiplication result is only N bit. One radical alternative, particu-
larly considering signal processing, is to add the binary logarithms and to do this by
getting the most-significant bit of the result first, [40–42]. This is the basis of the
DIGILOG (digital logarithmic) multiplier.
For our two binary numbers A and B, we determine the leading ONEs at
positions j and k, respectively, and then proceed as follows:
A ¼ 2j þ AR ;
B ¼ 2k þ BR ;
A B ¼ ð2j þ AR Þð2k þ BR Þ;
A B ¼ 2j B þ 2k AR þ AR BR :
The first part can be obtained easily by shifting and adding, producing a first
approximate result. Iterative procedures on the remainder ARBR deliver more
accurate results: After one iteration, the worst-case error is 6%, after two
iterations it is 1.6%, and the probability that it is <1% is 99.8%. Operating with
two iterations, we can estimate the complexity of this DIGILOG multiplier as
O(3 N), which means an improvement of 24/32 for an 8 8 bit multiplier and of
48/128 for a 16 16 bit multiplier. This is just to show that arithmetic units have
great potential for continuing improvement.
High-performance logic (HIPERLOGIC) based on the differential logic acceler-
ator of Fig. 3.29 and the DIGILOG multiplier can produce a 16 16 bit multiplier
performing 14 TOPS/W (14 GOPS/mW) at the 100 nm node and an operating
voltage of 500 mV, corresponding to70 fJ per operation [40–42]. Moving to 20 nm
and 300 mV, a performance of 700 GOPS/mW can be achieved for a 16 16 bit
multiplier with the technique described here or other ongoing developments.
The pressure for better arithmetic performance in mobile multimedia
applications is very high, as expressed in Chap. 13, with the demand for a 300-
fold improvement compared to 2010. The state-of-the art, although impressive,
underlines the 300-fold improvement required in that chapter. We have entered
several results from 2010 publications in Figs. 3.31 and 3.32.
The first plot, Fig. 3.31, can serve as the master graph, because it plots the speed
against the energy per operation. Certainly in mobiles, the issue is: At which energy
can you realize speed? We have entered the HIPERLOGIC data of above for
16 16 bit multipliers and some recently reported results from ultra-large-scale
media processors. These contain hundreds to thousands of arithmetic processing
elements (PEs), mostly organized as 4 4 bit adder/multipliers.
The best reported achieves 27 MOPS with 12 mW at a supply voltage of 340 mV
in a 32 nm technology, which gives a power efficiency of 2.8 TOPS/W for a 4 4
multiplier [43]. The adder inside this processor element is built in static CMOS
logic, which naturally becomes very slow at 340 mV in spite of a very advanced
technology node.
2048 four-bit arithmetic logic units (ALUs) were reported on an image processor
performing 310 GOPS/W with a 200 MHz clock at 1.0 V [44]. The progress in
multiplier performance between 2007 and 2010 is evident in [45], where
670 GOPS/W was reported. Similar efficiency levels were given in [46]. A trend
line of 100/decade has been drawn in Fig. 3.32, projecting 100 TOPS/W or
100 GOPS/mW for 4 4 multipliers in 2020. This may not be good enough
considering the unabated push towards HD and 3D video as outlined in Chaps. 6
and 13, where the demand is for a 1,000-fold improvement in power efficiency
78 B. Hoefflinger
Energy per Operation

1 pJ 100 fJ 10 fJ 1 fJ
1 GOP/s
GIC
RLO 16bx16b
HIPE 2020 proj.
(20 nm)
16bx16b
2000 proj.
100 MOP/s (100 nm)
4bx4b (32 nm)

16bx16b (0.8 mm)
2000
10 MOP/s
1 10 100 103
GOP/ s
Power Efficiency
mW
Fig. 3.31 This master chart of digital multipliers shows the operating speed vs the energy per
operation (top axis) and its inverse, the operations per second per watt, also called the power
efficiency (bottom axis)
between 2010 and 2020. The same rate is required for supercomputers (see
Fig. 6.4). Some more radical innovations like HIPERLOGIC will be needed to
meet the requirements of 2020. The HIPERLOGIC example not only offers a
quantum step of an order of magnitude, but it does so for a 16 16 multiplier,
and the PEs of 2020 will not be 4 x4 but 8 8 as a minimum. In 2020 a CMOS
16 16 bit digital multiplier will perform 600 MOPS with a power efficiency of
1 TOPS/mW (1fJ) (500 times the 2010, much shorter 4 4 bit multiplier at
2.8 GOPS/mW).
3.7 3D Chip Stacks
2005 can be seen as the year of the breakthrough for a 3D integration strategy that
has had a long history of sophisticated technology development in the shadow of
scaling down Si transistors. It is the stacking of prefabricated Si crystals at the chip
or wafer level. A number of milestones on this path can be identified as significant
advances and pointers. We have marked 1980 in Fig. 3.1, when the
10: Future Electron Devices (FED) program in Japan (1980–1990)
began to produce refined techniques for wafer thinning, for producing through-
silicon vias (TSV’s) and for wafer stacking. The
TOP/s
1 fJ 1
16bx16b mW
1 Neuron GOP/s
10 fJ 100
mW
Energy
IC
G
Power
LO
M
Operation Efficiency
EA
ER
P
TR
HI
GOP/s
INS
100 fJ 10
mW
MA
16bx16b x [3.40]
4bx4b [3.43] GOP/s
1 pJ 1
mW
[3.45]
4bx4b [3.44]
[3.45]
MOP/s
10 pJ 100
mW
MOP/s
100 pJ 100
2000 2010 mW
1990 2020
Year
Fig. 3.32 The power efficiency of multipliers: (Multiplications/s)/mW and the energy per multi-
plication 1990–2020
10: Electronic Eye (Das Elektronische Auge) program in Germany (1994–1998)
led to the realization of a state-of-the-art intelligent vision system with the biomor-
phic structure of high-dynamic-range pixels in a top Si layer being merged with an
underlying Si wafer containing first-level neural signal processing [47]. An
exploded view of this imager is shown in Fig. 3.33.
Another development has been pivotal in creating multilayer Si systems, namely
the use of Si as an active carrier for Si chips forming a Si multichip module (MCM)
as shown in Fig. 3.34 [48]. In this approach, the chips are flipped upside-down and
mounted face-down in the C4 (controlled-collapse chip connect) technique via
mbumps, patented by IBM in 1968. Not only is the Si carrier compatible in thermal
expansion, but it can be preprocessed to contain the proper blocking capacitance for
the chips, Fig. 3.34, eliminating ground bounce, but it can actually be smart,
providing electrostatic discharge (ESD) protection, temperature sensing, self-test
80 B. Hoefflinger
Fig. 3.33 Exploded view of a chip-stack intelligent-vision system from 1997. The top layer is the
back-end-of-line (BEOL) interconnect structure. The second layer shows the photoreceptor chip,
and the third layer represents the neural processing chip [47]. Source: IMS CHIPS 1998
Fig. 3.34 Flipped chips on a

Si carrier (1987) [48].
Courtesy IBM Boeblingen
[49], and energy harvesting, like photovoltaic cells. When the IBM C4 patent
expired in 1983, a new industry was formed providing bumping services and
chip-scale packaging (CSP), a value-adding industry that has grown into a critical
role today.
A further domain of chip stacking developed quickly with the 3D integration of
heterogeneous Si system functions into smart-power and smart-sensor products.
PowerMOS and insulated-gate bipolar (IGB) transistors are large-area Si devices,
as are most Si sensors for pressure, acceleration, etc., so that there is area on the
surface available to place small intelligent chips for control and communication.
To take into account this early introduction of chip stacks, we have placed a 1985
marker in Fig. 3.1 for these systems-on-chips (SoCs). Products in this domain are
now produced in volumes of hundreds of millions per year and are specially
covered in Chap. 14.
At about this time, digital, lightweight mobile phones began to appear, and the
stacking of different chips, such as transceivers, processors, memories, and
cameras, produced a high-volume, low-cost market for chip stacks as SoCs, and
these deserve a chapter of their own (Chap. 12).
The 2005 explosion of chip stacking happened because
– The pressure on more flash memory/cm2 became big enough to introduce
chip-stacking and
– Significant know-how for thinning wafers and forming TSVs had been develop-
ing in the background at the interface between the semiconductor and the
equipment industry to become quickly available.
IBM, for example, has had a continuous micropackaging research program over
decades [50], from which experts migrated to the Asian Tiger countries and set up
an industry providing stacking technology as a packaging service.
3D chip stacks developed into a multifaceted field, which can be organized into
two major areas:
– Wafer-level and chip-level stacking and
– Uniform- or heterogeneous-function stacking.
For the stacking of wafers, we have basically three options, as shown in
Fig. 3.35 [51].
This cross section shows the challenges and opportunities for progress in perfor-
mance and cost of new products. In Fig. 3.36, we show a concept involving the
bonding of wafers that contain two levels of transistors each produced by the
localized epitaxial growth presented in the previous section on 3D CMOS (e.g.,
Fig. 3.27). The bonded wafer contains four transistor layers, and further stacking is
proposed with mbumps.
Fig. 3.35 Three options for stacking the upper wafer on the bottom one. (a) Face-up; fusing the
back of the top wafer on the face of the bottom wafer requires full TSV on the top wafer. (b) Face-
down; no TSVs needed, allowing higher density and less interconnect parasitics. (c) Face-up with
oxide fusion, like wafer bonding, exploits strategies for vertical local and global routing to
maximize speed and density [51]. # IBM Technical Journals
82 B. Hoefflinger
Transistor levels Wafer levels
4
2
Wafer Bond Chip Stack
2
1
Chip-Trager
Fig. 3.36 3D stack with back-to-back bonded wafers and m bumps to a Si carrier below and
further stacks above. # WEKA Medien 2000
Uniform functions such as memories have the feature that vertical communica-
tion can be organized into a local area of each chip where the TSV signal and power
elevator can be concentrated effectively over numerous levels of chips. Therefore,
memories are candidates for integrating numerous layers of thinned wafers
(Fig. 3.37 [52]).
The state-of-the art in 2010 for stacked DRAMs and Flash memories is four to
eight wafers, and these devices certainly are the extreme test vehicles for managing
yield and cost in the case of wafer-level stacking. Built-in redundancy will support a
push towards more wafer levels for memories so that a trend curve as shown in
Fig. 3.38 is possibly leading to stacks of 64 wafers by 2020.
Fig. 3.37 Cross section of a stack of 32 wafers with TSVs [52]. # 2010 IEEE
When we talk about wafers, we mean wafers thinned to about 50 mm by grinding

and polishing, or significantly thinner (10–20 mm) Si films, manipulated as handle
wafers until they are fused to the next lower Si film and then detached from the
handle with techniques similar to the ones shown in Sect. 3.4, for example Fig. 3.19.
The Si film thickness is, of course, essential for the TSV diameter, its parasitic
resistance, capacitance, and inductance, and ultimately for the density achievable.
We see in Fig. 3.39 that the trend curve for the density of TSVs takes 15 years
for a 10-fold improvement. The TSV pitch of 40 mm in 2010 indicates a ratio of
1:1 between TSV pitch and wafer (film) thickness, so the keen interest in
manufacturing technologies for Si films <20 mm will continue.
A future of its own opens up for chip-level stacking. It presents formidable
challenges for handling these thin and miniature chips on the one hand, but it also
opens up the opportunity of literally cracking the complexity problems of large-
scale system chips containing different technologies and different functions with
the strategy of divide and conquer.
84 B. Hoefflinger
128
Number of stacked wafers

in DRAM & Flash production 64
64
32
32
16
16
8
8
1
2010 2012 2014 2016 2018 2020
Year
Fig. 3.38 The expected number of stacked wafers for DRAM or Flash products
One example for the production and handling of ultrathin chips is shown in
Fig. 3.40 [53]. A double-layer porous Si surface is produced similar to the
techniques described in Sect. 3.4 on SOI, Fig. 3.19, but now selectively and
including buried cavities in just the areas of the chips. The resulting wafer is fully
processed to implement the function of the chip, be it just a circuit or a sensor or
actuator. And the chips are tested for functionality. A deep-trench etch into the chip
boundaries predisposes the chips to be cracked, picked, and placed on an underly-
ing substrate. This is one possible technique for face-up stacking of fully fabricated
and tested chips.
We see that, in the world of stacking chips, we can partition the total system
into chips optimized for function, design, technology, testability, yield, and
manufacturing source, and pick up these individual chips as known-good die for
3D integration into a chip stack (or possibly several chip stacks on a Si carrier).
We can imagine how multifaceted the creation of new chip products for the best
functionality at optimum cost is getting and that this will continually change
the industry: the management of quality, the product reliability, and the product
liability have to be assured and clarified in a truly global spirit and scenario, which
will be both demanding and rewarding for those involved. This makes up another
call for global education and research enabling growth and progress in our world
with 3D-integrated chips.
Fig. 3.39 The density of 1011 x

transistor-level contacts,
on-chip vias, TSVs, and
m bumps
1010 x
Contacts
109 x
On-Chip Vias
108
Density (cm–2)
107
106
TSV´s
105
m bumps
104
1990 2000 2010 2020
Year
3.8 Single-Electron Transistor and Other Concepts
This is a book about nanoelectronic devices and products for everyone, and that
makes it a book about room-temperature, natural-environment nanoelectronics. We
have seen in Sect. 3.2 that our standard MOS transistor, at the end of its scaling
towards 10 nm, has become a single-electron transistor (SET; Fig. 3.41) in the sense
that we find at any given time just one electron in the channel moving across the
86 B. Hoefflinger
Fig. 3.40 Production of ultrathin-film chips, which have been fully processed and tested in a
standard wafer process. Good die are cracked, picked, and placed on the underlying substrate [53].
# 2006 IEEE
potential barrier either faster, generating more current ¼ high channel conduc-
tance, and an output ZERO or slower, generating less current ¼ small channel
conductance, and an output ONE.
As of 2010, no room-temperature circuits including signal regeneration, built out
of SETs with Coulomb confinement, have been reported. Large-scale logic
functions will probably not become competitive with ultralow-voltage differential
nano-CMOS. An SE device chain with Coulomb confinement may be considered as
a CCD-like serial memory (Chap. 23).
For 10 nm CMOS, we limited our assessment to single-gate transistor structures
for processing and cost reasons, and we saw that the atomic variance of dopants in
the channel had a large effect of 20 mV per doping ion on the threshold voltage and
therefore on the minimum allowable supply voltage, amplified by the allowance for
a worst-case voltage gain or noise margin. At this front, some improvement can be
gained by dual-gate (Figs. 3.22 and 3.23) or surround gate structures [35], if these
can be built with the EOT of 1 nm assumed in our quantitative study. The dual-gate
fundamental threshold sensitivity might improve to 10 mV/charge. The surround-
gate transistor topography should be vertical as in the following proposal for a
complementary bipolar transistor stack (Fig. 3.43). If this could be realized with an
equivalent oxide thickness EOT ¼ 1 nm and a Si cylinder of 10 nm diameter, the
sensitivity would decrease to 6.5 mV/charge.
We have to remember that all these 10 nm transistors operate in the subthresh-
old, barrier-control or carrier-diffusion mode. Since we operate in this barrier-
control mode, analogous to the bipolar transistor, and since we can no longer turn
the MOS transistor off anyway, it is worth revisiting the bipolar transistor, Sect. 3.1,
and scaling it down to the <10 nm level.
Let us assume an NPN nanotransistor with a 10 nm 10 nm emitter, a metal-
lurgical base width of 10 nm and the base doped at a level of NA ¼ 1020 cm3. This
configuration is identical to the channel of our MOS nanotransistor in Sect. 3.2.
In equilibrium, depletion widths WD ¼ 3 nm would extend from the emitter and the
1 GHz 10 GHz 100 GHz 1 THz 10 THz

100 μA
current
10 μA
1 μA
100 nA
10 nA
1 nA 3
10 104 105 106 107
velocity (cm /s)
1 ns 100 ps 10 ps 1 ps 0.1 ps
transit time
Fig. 3.41 Single-electron current across a channel of 10 nm length. For a channel cross section of
2 nm 10 nm, the current density at 10 mA would be 5 107 A/cm2, and the volume power
density would be 20 GW/cm3 for a drain voltage of 0.4 V
collector into the base, leaving a neutral base of 4 nm width. Within this neutral
base, we would find 40 acceptors, the Gummel number of this nanotransistor with a
standard deviation of s ¼ (40)1/2 ¼ 6 acceptors. Our corridor with 2 s would
give us Gummel numbers between 28 and 52 or 40 30%. Using (3.4), we find a
mean Early voltage VA ¼ 1.9 V, giving us an intrinsic voltage gain according to
(3.5) of
AV ¼ VA =Vt ¼ 1:9=0:025 ¼ 76
88 B. Hoefflinger
Fig. 3.42 The cross-coupled

differential inverter pair
Quad
at room temperature. If we can realize an emitter doping NE ¼ 1021/cm3, we should

be able to reach a current gain bF > 10, see (3.3), resulting in a power gain AP > 760
60%, if we accept that variances accumulate. The maximum useful current for this
nanotransistor, namely the knee current IK, see (3.8), where the transconductance
would be cut in half, would be Imax ¼ IK ¼ 1.2 mA, much larger than any thermal
limit.
This would make a reasonable device with several fundamental advantages over
the MOS nanotransistor:
– No problems with gate dielectrics,
– Optimum transconductance of 60 mV/decade (>3 times that of the MOS
transistor, Fig. 3.14),
– Perfect topology for vertical growth of transistor structures.
Following up on the last statement, we should pursue, as a vertical Si structure,
the pair of vertical complementary bipolar nanotransistors: (N+PNN+)(P+PNP+).
The pair would form one half of the quad, the cross-coupled pair of inverters
(Fig. 3.42), which is the necessary signal regenerator (Fig. 3.17) in low-voltage
differential logic (Fig. 3.29) and the heart of the high-speed SRAM (Fig. 3.18), with
the lowest operating voltage, nondestructive readout, and the smallest write and
read energy.
A structure for the vertical complementary bipolar nanotransistor pair is pro-
posed in Fig. 3.43. The realization of this transistor pair has associated with it the
incorporation of three local interconnect layers for optimum surround contacts to
the P and N bases and to the N+P+ output, and to provide three levels of local
interconnects, for example, to form a high-density quad. The proposed process flow
(Fig. 3.44) is
– SOI wafer with buried N+ as ground supply and heat sink
– P+-poly/salicide for P-base surround contact, INPUT, and interconnect
– Mask #1: P+-poly/salicide etch
– Deposit SiO2 and planarize
– Metal for N+P+ surround contact, OUTPUT, and interconnect
– Mask #2: Metal etch
IN Vcc OUT
p+
M M
n+ n+ n n+
p 1 6 nm
M M M
p+ M M
n+
n
p+ p+ p p+
SiO2 n+ SiO2
Diameter 10 nm
n+ GND n+
Fig. 3.43 Cross section of a vertical complementary bipolar nanotransistor pair as inverter
– N+-poly/salicide surround contact of N base, INPUT, and interconnect

– Mask #3: N+-poly/salicide etch
– Mask #4: Deep etch for transistor column
– Selectively doped Si crystal growth of complementary transistor column,
matched with the poly and metal contact rings
– Mask #5: Deep etch of contact columns
– Selective deposition of metal columns, in situ contacted with the conductor rings
and upwards-extendable as TSVs
The formation of complementary bipolar nanotransistors and three local contact
layers with just five masks is indicative of the potential of
– Device-level 3D integration
– The replacement of masks, overlay, and etch by selective growth (self-assembly)
– Vertical contact columns as TSVs
The complementary bipolar structure together with lateral NMOS-transistor
transmission gates (VT ¼ 0) and logic transistors (VT > 0) would form a
promising technology for low-voltage differential logic and high-speed SRAM
memory.
90 B. Hoefflinger
n+
M M
p+
SiO2
n+
n+ n+
M M M
p+ p+
SiO2 SiO2
n+
p+
n+ n n+
p
M
p+
M
n+ M
n
p+ p p+
SiO2 SiO2
n+
Fig. 3.44 Three steps in the formation of the complementary bipolar transistor pair. Top: After
mask #3 and planar oxide. Middle: After mask #4 and deep etch for the growth of the transistor
column. Bottom: After mask #5 and deep etch for contact columns
References
1. Noyce R.N.: Semiconductor device-and-lead structure. US Patent 2981877, filed 30 July 1959.
Issued 25 Apr 1961
2. Pfeiffer U.R, Ojefors E, Zhao Y.: A SiGe quadrature transmitter and receiver chipset for
emerging high-frequency applications. IEEE ISSCC (International Solid-State Circuits Con-
ference), Digest of Technical Papers, pp.416–417, (2010). doi: 10.1109/ISSCC.2010.5433832
3. Gray P.R, Hurst P.J, Lewis S.H, Meyer R.G.: Analysis and Design of Analog Integrated
Circuits. Wiley, New York (2001)
4. Atalla M.M.: Stabilisation of silicon surfaces by thermally grown oxides. Bell Syst. Tech. J.
38, 749 (1959)
5. Kahng D.: Electric field controlled semiconductor device. US Patent 3102230, filed 31 May
1960. Issued 27 Aug 1963
6. Liu S, Hoefflinger B, Pederson D.O.: Interactive two-dimensional design of barrier-controlled
MOS transistors. IEEE Trans. Electron Devices 27, 1550 (1980)
7. Hoefflinger B.: “New CMOS technologies” In: Carroll J.E. (ed.) Solid-state devices 1980,
Institute of Physics Conference Series, No. 57. pp. 85–139. Institute of Physics Publishing,
Bristol (1981)
8. Grotjohn T, Hoefflinger B.: A parametric short-channel MOS transistor model for subthresh-
old and strong inversion current. IEEE J. Solid-State Circuits 19, 100 (1984)
9. Wakabayashi H, et al.: Characteristics and modeling of sub-10-nm planar bulk CMOS devices
fabricated by lateral source/drain junction control. IEEE Trans. Electron Devices 53, 1961
(2006)
10. Hoefflinger B, Bigall K.D, Zimmer G, Krimmel E.F.: Siemens Forschungs- und Entwicklungs-
berichte 1(4), 361 (1972)
11. Wanlass F.M.: Low stand-by power complementary field effect circuitry. US Patent 3356858,
filed 18 June 1963. Issued 5 Dec 1967
12. Wanlass F.M, Sah C.T.: Nanowatt logic using field-effect metal-oxide semiconductor triodes.
IEEE ISSCC (International Solid-State Circuits Conference), Digest of Technical Papers,
pp. 32–33. (1963)
13. Lin H.C, Ho J.C, Iyer R, Kwong K.: Complementary MOS–bipolar transistor structure. IEEE
Trans. Electron Devices 16, 945 (1969)
14. Strachan A.J, Wagner K.: Local oxidation of silicon/CMOS: Technology/design system
for LSI in CMOS. IEEE ISSCC (International Solid-State Circuits Conference), Digest of
Technical Papers, pp. 60–61. (1974). doi: 10.1109/ISSCC.1974.1155337
15. Hoefflinger B, Schneider J, Zimmer G.: Advanced compatible LSI process for N-MOS, CMOS
and bipolar transistors. IEEE IEDM (International Electron Devices Meeting), Technical
Digest, pp. 261A–F. (1977). doi: 10.1109/IEDM.1977.189225
16. Plummer J, Meindl J.D, Maginness M.G.: An ultrasonic imaging system for real-time cardiac
imaging. IEEE ISSCC (International Solid-State Circuits Conference), Digest of Technical
Papers, pp. 162–163. (1974). doi: 10.1109/ISSCC.1974.1155347
17. Manasevit H.M, Simpson W.I.: Single-crystal silicon on a sapphire substrate. J. Appl. Phys.
35, 1349 (1964)
18. Wallmark J.T, Johnson H (eds.): Field-Effect Transistors. Prentice-Hall, Englewood Cliffs
(1966)
19. Izumi K, Doken H, Arioshi H.: CMOS devices fabricated on buried SiO2 layers formed by
oxygen implantation into silicon. Electron Lett. 14, 593 (1978)
20. Gibbons J.F, Lee K.F.: One-gate-wide CMOS inverter on laser-recrystallised polysilicon.
IEEE Electron Device Lett. 1, 117 (1980)
21. Imai K.: A new dielectric isolation method using porous silicon. Solid-State Electron 24, 159
(1981)
92 B. Hoefflinger
22. Ipri A.C, Jastrzebski L, Corboy J.F.: Device characterisation on monocrystalline silicon grown
over SiO2 by the ELO Process. IEEE IEDM (International Electron Devices Meeting),
pp. 437–440 (1982)
23. Laskey J.B, Stiffler S.R, White F.R, Abernathey J.R.: Silicon on insulator (SOI) by bonding
and etchback. IEEE IEDM (International Electron Devices Meeting), pp. 684 ff. (1985).
doi: 10.1109/IEDM.1985.191067
24. Yonehara T, Sakaguchi K, Sato N.: Epitaxial layer transfer by bond and etch back of porous
Si. Appl. Phys. Lett. 64, 2108 (1994)
25. Bruel M.: Silicon on insulator material technology. Electron Lett. 31, 1201 (1995)
26. Colinge J.P.: Silicon-on-insulator and porous silicon. In: Siffert P, Krimmel E (eds.) Silicon:
Evolution and Future of a Technology, p. 139. Springer, Berlin/Heidelberg (2004)
27. Hoefflinger B.: Circuit considerations for future 3-dimensional integrated circuits. Proceedings
of 2nd International Workshop on Future Electron Devices – SOI Technology and 3D
Integration, Shujenzi, Japan, March 1985
28. Hoefflinger B, Liu S.T, Vajdic B.: Three-dimensional CMOS design methodology. IEEE J.
Solid-State Circuits 19, 37 (1984)
29. Liu M.S, Hoefflinger B.: Three-dimensional CMOS using selective epitaxial growth. US
Patent 4686758, filed 2 Jan 1986. Issued 18 Aug 1987
30. Zingg R.P, Hoefflinger B, Neudeck G.: Stacked CMOS inverter with symmetric device
performance. IEEE IEDM (International Electron Devices Meeting), Digest of Technical
Papers, pp. 909–911 (1989)
31. Roos G, Hoefflinger B.: Complex 3D-CMOS circuits based on a triple-decker cell. IEEE J.
Solid-State Circuits 27, 1067 (1992)
32. Roos G.: Ph.D. Dissertation: Zur drei-dimensionalen Integration von CMOS-Schaltungen,
Fortschrittsberichte VDI, Reihe 9: Elektronik, Nr.168, (VDI, D€usseldorf, 1993)
33. Abou-Samra S.J, Alsa, P.A, Guyot P.A, Courtois B.: 3D CMOS SOI for high performance
computing. Proceedings of 1998 IEEE International Symposium Low-Power Electronics and
Design (ISLPED). pp. 54–58 (1998)
34. Waser R (ed.): Nanoelectronics and Information Technology. Wiley-VCH, Weinheim (2003)
35. Risch L.: Silicon nanoelectronics: the next 20 years. In: Siffert P, Krimmel E (eds.) Silicon:
Evolution and Future of a Technology. Springer, Berlin/Heidelberg (2004) (Chap. 18)
36. Wolf E.L.: Quantum Nanoelectronics. Wiley-VCH, Weinheim (2009)
37. Eisele I, Schulte J, Kaspar E.: Films by molecular-beam epitaxy. In: Siffert P, Krimmel E
(eds.) Silicon: Evolution and Future of a Technology. Springer, Berlin/Heidelberg (2004)
(Chap. 6)
38. Hossain R.: High Performance ASIC Design: Using Synthesizable Domino Logic in an ASIC
Flow. Cambridge University Press, Cambridge (2008)
39. Ruiz G.A.: Evaluation of three 32-bit CMOS adders in DCVS logic for self-timed circuits.
IEEE J. Solid-State Circuits 33, 604 (1998)
40. Grube R, Dudek V, Hoefflinger B, Schau M.: 0.5 Volt CMOS logic delivering 25 Million
16 16 multiplications/s at 400 fJ on a 100 nm T-Gate SOI technology. Best-Paper Award.
IEEE Computer Elements Workshop, Mesa, (2000)
41. Hoefflinger B, Selzer M, Warkowski F.: Digital logarithmic CMOS multiplier for very-high-
speed signal processing. IEEE Custom-Integrated Circuits Conference, Digest, pp. 16.7.1–5
(1991)
42. cordis.europa.eu > EUROPA > CORDIS > Archive: esprit ltr project 20023
43. Agarwal A, Mathew S.K, Hsu S.K, Anders M.A, Kaul H, Sheikh F, Ramanarayanan R,
Srinivasan S, Krishnamurthy R, Borkar S.: A 320 mV-to-1.2 V on-die fine-grained
reconfigurable fabric for DSP/media accelerators in 32 nm CMOS. IEEE ISSCC (International
Solid-State Circuits Conference), Digest of Technical Papers, pp. 328–329 (2010)
44. Kurafuji T, et al.: A scalable massively parallel processor for real-time image processing,
IEEE ISSCC (International Solid-State Circuits Conference), Digest of Technical Papers,
pp. 334–335 (2010)
45. Chen T.-W, Chen Y.-L, Cheng T.-Y, Tang C.-S, Tsung P.-K, Chuang T.-D, Chen L.-G,
Chien S.-Y.: A multimedia semantic analysis SOC (SASoC) with machine-learning engine.
IEEE ISSCC (International Solid-State Circuits Conference), Digest Technical Papers,
pp. 3380150339 (2010)
46. Lee S, Oh J, Kim M, Park J, Kwon J, Yoo H.-J.: A 345 mW heterogeneous many-core
processor with an intelligent inference engine for robust object recognition. IEEE ISSCC
(International Solid-State Circuits Conference), pp. 332–333 (2010)
47. Graf H.G.: Institute for Microelectronics Stuttgart, Germany
48. Si carrier, courtesy IBM Laboratories, B€oblingen, Germany
49. Werkmann H, Hoefflinger B.: Smart substrate MCM testability optimisation by means of chip
design. IEEE 6th International Conference on Multichip Modules, pp. 150–155. (1997)
50. Knickerbocker J.U. (ed.): 3D chip technology, IBM J. Res. Dev. 52(6), November (2008)
51. Koester S.J, et al.: Wafer-level 3D integration technology. IBM J. Res. Dev. 52, 583 (2008)
52. Kang U.: TSV technology and its application to DRAM. IEEE ISSCC (International Solid-
State Circuits Conference), Forum 1: Silicon 3D Integration Technology and Systems (2010)
53. Zimmermann M, et al.: Seamless ultra-thin chip fabrication and assembly process,
IEEE IEDM (International Electron Devices Meeting), Digest of Technical Papers,
pp. 1010–1012. (2006). doi: 10.1109/IEDM.2006.346787
Chapter 4
Analog-Digital Interfaces
Matthias Keller, Boris Murmann, and Yiannos Manoli
Abstract This chapter discusses trends in the area of low-power, high-performance

A/D conversion. Survey data collected over the past thirteen years are examined to
show that the conversion energy of ADCs has halved every two years while the
speed-resolution product has doubled approximately only every four years. A closer
inspection on the impact of technology scaling and developments in ADC design are
then presented to explain the observed trends. Next, opportunities in minimalistic and
digitally assisted design are reviewed for the most popular converter architectures.
Finally, trends in Delta–Sigma ADCs are analyzed in detail.
4.1 Introduction
Analog-to-digital converters (ADCs) are important building blocks in modern

electronic systems. In many cases, the efficiency and speed at which analog
information can be converted into digital signals profoundly affects a system’s
architecture and its performance. Even though modern integrated circuit technology
can provide very high conversion rates, the associated power dissipation is often
incompatible with application constraints. For instance, the high-speed ADCs
presented in [1] and [2] achieve sampling rates in excess of 20 GS/s, at power
dissipations of 1.2 W and 10 W, respectively. Operating such blocks in a handheld
application is impractical, as they would drain the device’s battery within a short
amount of time. Consequently, it is not uncommon to architect power constrained
M. Keller (*) • Y. Manoli

Department of Microsystems Engineering –IMTEK, University of Freiburg, Georges-K€ohler-
Allee 102, Freiburg 79106, Germany
e-mail: mkeller@imtek.de
B. Murmann
Stanford University, 420 Via Palou, Allen-208, Stanford, CA 94305-4070, USA
e-mail: murmann@stanford.edu

96 M. Keller et al.
applications “bottom-up,” by determining the analog/radio frequency (RF) front-

end and ADC specifications based on the available power or energy budget.
A discussion detailing such an approach for the specific example of a software-
defined radio receiver is presented in [3].
With power dissipation being among the most important concerns in mixed-
signal and RF applications, it is important to track trends and understand the relevant
trajectories. The purpose of this chapter is to review the latest developments in low-
power A/D conversion and to provide an outlook on future possibilities, thus
extending and updating the results presented in [4]. Following this introduction,
Sect. 4.2 provides survey data on ADCs published over the past 13 years
(1997–2010). These data show that contrary to common perception, extraordinary
progress has been made in lowering the conversion energy of ADCs. Among the
factors that have influenced this trend are technology scaling, and the increasing use
of simplified analog sub-circuits with digital correction. Therefore, Sect. 4.3 takes a
closer look at the impact of feature-size scaling, while Sects. 4.4 and 4.5 discuss
recent ideas in “minimalistic” and “digitally assisted” ADC architectures. Finally,
the last section presents recently developed approaches for delta-sigma ADCs in
order to face the limitations of aggressively scaled CMOS technologies.
4.2 General ADC Performance Trends
4.2.1 Survey Data and Figure of Merit Considerations
Several surveys on ADC performance are available in the literature [5–8]. In this
section, recent data from designs presented at the IEEE International Solid-State Circuits
Conference (ISSCC) and the VLSI Circuit Symposium will be reviewed. Figure 4.1
shows a scatter plot of results published at these venues over the past 13 years [9].
Figure 4.1a plots the conversion energy per Nyquist sample P/fNyquist, i.e., power
divided by the Nyquist sampling rate with the latter being equal to two times the
signal bandwidth fb, against the achieved signal-to-noise-and-distortion ratio (SNDR).
This plot purposely avoids dividing the conversion energy by the effective
number of quantization steps (2EN0B), as done in the commonly used figure of
merit [8]
P P
FOM ¼ ¼ (4.1)
fNyquist 2ENOB 2fb 2ENOB
where
SND R½dB 1:76

ENOB ¼ : (4.2)
6:02
4 Analog-Digital Interfaces 97
a
10−6
10−7
10−8
P / f nyquist (J)
10−9 [2]
10−10
10−11 Flash
Pipeline
10−12 [10] SAR
ΔΣ
Other
10−13 100 J/conv−step
10 J/conv−step
10−14
20 40 60 80 100
SNDR (dB)
b
Jitter = 0.1 psrms
[2] Jitter = 1.0 psrms
1010
BW (Hz)
108
106 [10]
104
20 30 40 50 60 70 80 90 100 110
SNDR (dB)
Fig. 4.1 ADC performance data (ISSCC and VLSI Circuit Symposium 1997–2010). Conversion
energy (a) and conversion bandwidth (b) versus SNDR
Normalizing by the number of quantization steps assumes that doubling the

precision of a converter will double its power dissipation, which finds only empiri-
cal justification [8]. Fundamentally, if a converter were purely limited by thermal
noise, its power would actually quadruple per added bit (see Sect. 4.3). Nonetheless,
since this assumption is somewhat pessimistic for real designs, it is preferable to
avoid a fixed normalization between precision and energy altogether when plotting
98 M. Keller et al.
data from a large range of architectures and resolutions. The FOM given by (4.1) is
useful mostly for comparing designs with similar resolution.
Figure 4.1a indicates that the lowest conversion energy is achieved by ADCs
with low to moderate resolution, i.e., SNDR < 60 dB. In terms of the FOM given
by (4.1), these designs lie within or near the range of 10–100 fJ/conversion-step,
included as dashed lines in Fig. 4.1a.
In addition to an ADC’s conversion energy, the available signal bandwidth is
an important parameter. Figure 4.1b plots bandwidth against SNDR for the given
data set. In this chart, the bandwidth plotted for Nyquist converters is equal to the
input frequency used to obtain the stated SNDR (this frequency is not necessarily
equal to half of the Nyquist frequency fNyquist). The first interesting observation
from Fig. 4.1b is that, across all resolutions, the converters with the highest
bandwidth achieve a performance that is approximately equivalent to an aperture
uncertainty between 0.1 and 1 psrms. The dashed lines in Fig. 4.1b represent
the performance of ideal samplers with a sinusoidal input and 0.1 and 1 psrms
sampling clock jitter, respectively. Clearly, any of the ADC designs at this
performance front rely on a significantly better clock, to allow for additional
non-idealities that tend to reduce SNDR as well. Such non-idealities include
quantization noise, thermal noise, differential nonlinearity and harmonic distor-
tion. From the data in Fig. 4.1b, it is also clear that any new design aiming to push
the speed–resolution envelope will require a sampling clock with a jitter of
0.1 psrms or better.
In order to assess the overall merit of ADCs (conversion energy and band-
width), it is interesting to compare the locations of particular design points in the
plots of Fig. 4.1a and 4.1b. For example, [2] achieves a bandwidth close to the
best designs, while showing relatively high conversion energy. The opposite is
true for [10]; this converter is the lowest energy ADC published to date, but was
designed only for moderate bandwidth. These examples confirm the intuition that
pushing a converter toward the speed limits of a given technology will sacrifice
efficiency and increase the conversion energy. To show this more generally, the
conversion energy of ADCs providing a bandwidth of 100 MHz or more is
highlighted in red in Fig. 4.1a. As can be seen, there are only two designs
that fall below the 100 fJ/conversion step line [11, 12]. The tradeoff between
energy and conversion bandwidth is hard to analyze in general terms. There
are however, architecture-specific closed form results, e.g., as presented in [13]
for a pipeline ADC.
Yet another important (but often neglected) aspect in ADC performance
comparisons is the converter’s input capacitance (or resistance) and full-scale
range. For most ADCs with a sampling-front-end, it is possible to improve the
SNDR by increasing the circuit’s input capacitance. Unfortunately, it is difficult
to construct a fair single-number figure of merit that includes the power needed
to drive the converter input. This is mostly because the actual drive energy will
strongly depend on the driver circuit implementation and its full specifications.
Therefore, an alternative approach is to calculate an approximation for the
input drive energy based on fundamental tradeoffs. Such a metric was recently
proposed in [14]
Cin ELSB2
EQ;in ¼ : (4.3)
2ENOB
In this expression, ELSB is the effective size of the quantization step, i.e., the
full-scale range (in volts) divided by the number of effective quantization steps
(2EN0B). For a converter that is limited by matching or thermal noise, the product in
the numerator of (4.3) is independent of resolution, and captures how efficiently the
capacitance is put to work. Normalizing by 2EN0B then yields a metric similar to
(4.1) in which the energy is distributed across all quantization levels.
The figure of merit defined by (4.3) is useful as a stand-alone metric to compare
ADCs with otherwise similar performance specifications in terms of their input
energy efficiency. Consider e.g. the 9–10 bit, 50 MS/s designs described in [15]
(successive-approximation-register (SAR) ADC with Cin ¼ 5.12 pF, FOM ¼ 52
fJ/conversion-step) and [16] (pipeline ADC with Cin ¼ 90 fF and FOM ¼ 119 fJ/
conversion-step). For the SAR ADC of [15], one finds EQ,in ¼ 1.1 1019 J/step,
while for the pipeline design of [16], one obtains EQ,in ¼ 2 1021 J/step. This
result indicates that the drive energy for the pipeline design is approximately two
orders of magnitude lower compared to the SAR design. Whether this is a signifi-
cant advantage depends on the particular application and system where the con-
verter will be used.
4.2.2 Trends in Power Efficiency and Speed
Using the data set discussed above, it is interesting to extract trends over time.
Figure 4.2a is a 3D representation of the conversion energy data shown in
Fig. 4.1a with the year of publication included along the y-axis. The resulting
slope in time corresponds to an average reduction in energy by a factor of two
approximately every 1.9 years. A similar 3D fit could be constructed for band-
width performance. However, such a fit would not convey interesting informa-
tion, as the majority of designs published in recent years do not attempt to
maximize bandwidth. This contrasts the situation with conversion energy,
which is subject to optimization in most modern designs. In order to extract
a trend on achievable bandwidth, Fig. 4.2b scatter-plots the speed–resolution
products of the top three designs in each year. This metric is justified by the
speed–resolution boundary observed from Fig. 4.1b, in which the straight lines
obey a constant product of fb and 2EN0B. A fit to the data in Fig. 4.2b reveals that
speed–resolution performance has doubled every 3.6 years, a rate that is signifi-
cantly lower than the improvement in conversion energy. In addition, as evident
from the data points, there is no pronounced trend as far as the top performance
100 M. Keller et al.
−6
P / fnyquist / Joule (dB)
−8
−10
−12
2010
120
2005 100
80
year 2000 60
40 SNDR (dB)
20
b
1012
fb x 2ENOB (Hz−steps)
1011
1010
1996 1998 2000 2002 2004 2006 2008 2010

year
Fig. 4.2 Trends in ADC performance. (a) 3D fit to conversion energy using a fit plane with a slope
of 0.5 /1.9 years along the time axis. (b) Fit to speed-resolution product of top three designs in
each year using a fit line with a slope of 2/3.6 years
point is concerned; designs of the early 2000s are almost on a par with some
of the work published recently. Consequently, the extracted progress rate of
speed–resolution performance should be viewed as a relatively weak and error-
prone indicator.
4.3 Impact of Technology Scaling
As discussed in Sect. 4.2.2, the power dissipation of ADCs has halved approxi-
mately every 2 years over the past 13 years. Over the same period, CMOS
technologies used to implement the surveyed ADCs have scaled from approxi-
mately 0.6 mm down to 45 nm. Today, the choice of technology in which an ADC is
implemented strongly depends on the application context. For stand-alone parts,
older technologies such as 0.18 mm CMOS are often preferred (see e.g., [17]).
In contrast, most embedded ADCs usually must be implemented in the latest
technologies used to realize large systems-on-chip (SoCs) [18]. Since the number
of designs used in embedded SoC applications clearly outweighs the number of
stand-alone parts, many ADC implementations in aggressively scaled technology
have been presented over the past several years. Therefore, the role of technology
scaling in the context of the trends summarized in the previous section is now
investigated. Broader discussions on the general impact of scaling on analog
circuits are presented in [5, 19, 20].
4.3.1 Supply Voltage and Thermal Noise Considerations
A well-known issue in designing ADCs in modern processes is the low voltage

headroom. Since device scaling requires a reduction in supply voltage (VDD), the
noise in the analog signals must be reduced proportionally to maintain the desired
signal-to-noise ratio. Since noise trades with power dissipation, this suggests that
to first order power efficiency should worsen, rather than improve, for ADCs in
modern technologies. One way to overcome supply voltage limitations is to utilize
thick-oxide I/O devices [21], which are available in most standard CMOS
processes. However, using I/O devices usually reduces speed. Indeed, a closer
inspection of the survey data considered in this chapter reveals that most published
state-of-the-art designs do not rely on thick oxide devices, and rather cope with
supply voltages around 1 V. To investigate further, it is worthwhile examining the
underlying equations that capture the tradeoff between supply voltage and energy
via thermal noise constraints. In most analog sub-circuits used to build ADCs, noise
is inversely proportional to capacitance
kT
N / ; (4.4)
C
where k is Boltzmann’s constant and T stands for absolute temperature. For the
specific case of a transconductance amplifier that operates linearly, it follows that
gm
fNyquist / : (4.5)
C
Further assuming that the signal power is proportional to (aVDD)2 and that the
circuit’s power dissipation is VDD multiplied by the transistor drain current, ID, one finds
P 1 1 1
/ kT SNR: (4.6)
fNyquist a2 VDD ðgm =ID Þ
The variable gm/ID in (4.6) is related to the gate overdrive (VGS – Vt) of the
transistors that implement the transconductance. Assuming MOS square law, gm/ID
equals 2/(VGS – Vt), while in weak inversion, gm/ID equals 1/ (nkT/q), with n ffi 1.5.
Considering the fractional swing (a) and transistor bias point gm/ID as constant, it is
clear from (4.6) that the energy in noise-limited transconductors should deteriorate
at low VDD. In addition, (4.6) indicates a very steep tradeoff between SNR and
energy; increasing the SNR by 6 dB requires a fourfold increase in P/fNyquist.
Since neither of these results correlate with the observations of Sect. 4.2, it is
instructive to examine the assumptions that lead to (4.6). The first assumption is that
the circuit is purely limited by thermal noise. This assumption holds for ADCs with
very high resolution, but typically few, if any, low-resolution converters are
severely impaired by thermal noise. To get a feeling for typical SNR values at
which today’s converters become “purely” limited by noise, it is helpful to plot the
data of Fig. 4.1a normalized to a 4 power increase per bit [22]. Figure 4.3 shows
such a plot in which the P/fNyquist values have been divided by
108
106
(P/fnyquist) / (P/fnyquist)min
104
102
FOM = 100 fJ/conv−step

100
20 30 40 50 60 70 80 90 100 110
SNDR (dB)
Fig. 4.3 Data of Fig. 4.1a normalized by (P/fNyquist)min as given by (4.7). This illustration suggests
the existence of an “SNDR corner” (bold line). Only ADCs with SNDR > 75 dB appear to be
primarily limited by thermal noise

P
¼ 8kT SNR (4.7)
fNyquist min
while assuming SNR ffi SNDR. The pre-factor of 8 in this expression follows from
the power dissipated by an ideal class-B amplifier that drives the capacitance C with
a rail-to-rail tone at fNyquist/2 [23]. Therefore, (4.7) represents a fundamental bound
on the energy required to process a charge sample at a given SNR.
The main observation from Fig. 4.3 is that the normalized data exhibit a visible
“corner” beyond which (P/fNyquist)/(P/fNyquist)min approaches a constant value. This
corner, approximately located at 75 dB, is an estimate for the SNDR at which a typical
state-of-the-art design becomes truly limited by thermal noise. Since ADCs with lower
SNDR do not achieve the same noise-limited conversion energy, it can be argued that
these designs are at least partially limited by the underlying technology. This implies
that, over time, technology scaling may have helped improve their energy significantly
as opposed to the worsening predicted by (4.6).
To investigate further, the data from Fig. 4.1a were partitioned into two distinct
sets: high resolution (SNDR > 75 dB) and low-to-moderate resolution (SNDR 75
dB). A 3D fit similar to that shown in Fig. 4.2a was then applied to each set and the
progress rates over time were extracted. For the set with SNDR > 75 dB, it was
found that the conversion energy has halved only every 4.4 years, while for SNDR
75 dB, energy halves every 1.7 years. The difference in these progress rates
confirms the above speculation. For high-resolution designs, (4.6) applies and
scaling technology over time, associated with lower supply voltages, cannot help
improve power efficiency. As observed in [5], this has led to a general trend toward
lower resolution designs: since it is more difficult to attain high SNDR at low supply
voltages, many applications are steered away from using high-resolution ADCs in
current fine-line processes. This is qualitatively confirmed in Fig. 4.4, which
10−6
10−8
P/fnyquist (J)
10−10
10−12
90 nm and below
100 fJ / conv−step
10 fJ / conv
10−14
20 40 60 80 100
SNDR (dB)
Fig. 4.4 Conversion energy for ADCs built in 90 nm and below

highlights the conversion energy data points of ADCs built in CMOS at 90 nm (and
VDD ffi 1 V) and below. As can be seen from this plot, only a limited number of
ADCs built in 90 nm and below target SNDR > 75 dB.
Lastly, an additional and more general factor to consider is the tradeoff between
the transconductor efficiency gm/ID and the transit frequency fT of the active
devices.
4.3.2 gm/Id and fT Considerations
Switched capacitor circuits based on class-A operational transconductance

amplifiers typically require transistors with fT ffi 50fs [24]. Even for speeds of
several tens of MS/s, it was necessary in older technologies to bias transistors far
into strong inversion (VGS – Vt > 200 mV) to satisfy this requirement. In more
recent technologies, very large transit frequencies are available in moderate- and
even weak-inversion. This is further illustrated in Fig. 4.5a, which compares typical
minimum-length NMOS devices in 180 nm and 90 nm CMOS.
For a fixed sampling frequency, and hence fixed fT requirement, newer
technologies deliver higher gm/ID. This tradeoff is plotted directly, without the
intermediate variable VGS Vt, in Fig. 4.5b. In order to achieve fT ¼ 30 GHz,
a 180 nm device must be biased such that gm/ID ffi 9 S/A. In 90 nm technology,
fT ¼ 30 GHz is achieved in weak inversion, at gm/ID ffi 18 S/A. From (4.6), it is
clear that this improvement can fully counteract the reduction in VDD when going to
a newer process. Note, however, that this advantage can only materialize when
the sampling speed is kept constant or at least not scaled proportional to the
fT improvement. This was also one of the observations drawn from Fig. 4.2b:
a converter that pushes the speed envelope using a new technology typically will
not simultaneously benefit from scaling in terms of conversion energy.
4.3.3 Architectural Impact
As discussed above, the high transistor speed available in new technologies can be
leveraged to improve the energy efficiency of analog blocks (such as amplifiers).
For similar reasons, it can also be argued that the high transistor speed has had
a profound impact on architectural choices and developments. The very high-speed
transistors in 90 nm CMOS and below have led to a renaissance or invention of
architectures that were deemed either slow or inefficient in the past. As illustrated in
Fig. 4.6, one example in this category is the SAR ADC [25], but also binary search
architectures fall within this category (e.g. [26]). In addition, the high integration
density of new processes makes massive time-interleaving (see e.g. [2]) with its
associated benefits a possibility (see Sect. 4.5.2).
30 150
gm /ID (S/A)
20 100
fT (GHz)
10 50
0 0
−0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 0.6
VGS−Vt (V)
b
45nm
90nm
180nm
102
fT [GHz]
101
5 10 15 20 25 30
gm / ID [S/A]
Fig. 4.5 Tradeoff between gm/ID and fT in modern CMOS technologies
4.4 Minimalistic Design
In addition to technology scaling, the trend toward “minimalistic” and “digitally

assisted” designs continues to impact the performance of ADCs. In this section,
ideas in minimalistic design will be discussed, followed by a discussion on the
importance of digitally assisted architectures in Sect. 4.5.
10
ISSCC 1976−2010
9 VLSI 1997−2009
8 ASSCC 1997−2008
number of publications CICC 1988−2008
7 ESSCIRC 2000−2008
0
1975 1980 1985 1990 1995 2000 2005 2010
year
Fig. 4.6 Statistics showing the number of publications of SAR ADCs at various conferences from
1975 to 2010
Power dissipation in the analog portion of ADCs is strongly coupled to the

complexity of the constituent sub-circuits. The goal of minimalistic design is to
reduce power and potentially increase speed by utilizing simplified analog sub-
circuits. In architectures that previously relied on precision op-amp-based signal
processing, there exists a clear trend toward simplified amplifier structures.
Examples include inverter-based delta-sigma modulators [27, 28] and various
approaches emphasizing op-amp-less implementation of pipelined ADCs
[16, 29–33]. Especially in switched capacitor circuits, eliminating class-A
op-amps can dramatically improve power efficiency. This is for two reasons.
First, operational amplifiers typically contribute more noise than simple gain stages,
as for example resistively loaded open-loop amplifiers. Secondly, the charge
transfer in class-A amplifier circuitry is inherently inefficient; the circuit draws
a constant current, while delivering on average only a small fraction of this current
to the load. In [34], it was found that the efficiency of a class-A amplifier in
a switched capacitor circuit is inversely proportional to the number of settling
time constants. For the typical case of settling for approximately ten or more time
constants, the overall efficiency, i.e., charge drawn from the supply versus charge
delivered to the load, is only a few percent.
As discussed further in [35], this inherent inefficiency of op-amps contributes to
the power overhead relative to fundamental limits. Consider for instance the
horizontal asymptote of Fig. 4.3, located at approximately 150 times the minimum
possible P/fNyquist. The factor of 150 can be explained for op-amp based circuits as
follows. First, the noise is typically given by b kT/C, where b can range from 5 to
10, depending on implementation details. Second, charge transfer using class-A
circuits, as explained above, brings a penalty of approximately 20. Third, op-amp
circuits usually do not swing rail-to-rail as assumed in (4.7); this can contribute
another factor of two. Finally, adding further power contributors beyond one
dominant op-amp easily explains a penalty factor greater than 100–400.
A promising remedy to this problem is to utilize circuits that process charge
more efficiently and at the same time contribute less thermal noise. A well-known
example of an architecture that achieves very high efficiency is the charge-based
successive approximation register (SAR) converter, see e.g. [2, 10, 25]. Such
converters have been popular in recent years, primarily because the architecture
is well-suited for leveraging the raw transistor speed of new technologies, while
being insensitive to certain scaling implications, such as reduced intrinsic gain (gm/
gds). A problem with SAR architectures is that they cannot deliver the best possible
performance when considering absolute speed, resolution, and input capacitance
simultaneously (see Sect. 4.2.1). This is one reason why relatively inefficient
architectures, such as op-amp-based pipelined ADCs, are still being used and
investigated.
In order to make pipelined architectures as power efficient as competing SAR
approaches, various ideas are being explored in research. Figure 4.7 shows an
overview of amplification concepts that all pursue the same goal: improve the
energy efficiency of the reside amplification.
In Fig. 4.7a, the traditional op-amp is replaced by a comparator [31], which shuts
off the capacitor charging current when the final signal value is reached. In
Fig. 4.7b, a “bucket brigade” pass transistor is used to move a sampled charge
packet q from a large sampling capacitor CS to a smaller load capacitor CL, thereby
achieving voltage gain without drawing a significant amount of energy from the
supply [30, 36]. Lastly, in Fig. 4.7c, the gate capacitance of a transistor is used to
acquire a charge sample (q1, q2). The transistor is then switched into a source-
a b
+
–
(large)
+
Vout Vout
–
V q (small) CL q
CL
c VDD
q1 Vout
–
q1 q2 + Vt
q2 + CL
q =0
VBIAS (floating)
Fig. 4.7 Energy-efficient switched capacitor amplification concepts. (a) Comparator-based

switched capacitor circuit [31]. (b) Bucket brigade circuit [30, 36]. (c) Dynamic source follower
amplifier [16]
follower configuration, moving all signal charges q1 + q2 to the small capacitance

from gate to drain, which also results in voltage amplification [16].
A general concern with most minimalistic design approaches is that they tend to
sacrifice robustness, e.g., in terms of power supply rejection, common mode
rejection, and temperature stability. It remains to be seen if these issues can be
handled efficiently in practice. Improving supply rejection, for instance, could
be achieved using voltage regulators. This is common practice in other areas of
mixed-signal design, for example in clock generation circuits [37, 38]. Especially
when the power of the ADC’s critical core circuitry is lowered significantly,
implementing supply regulation should be a manageable task.
A second issue with minimalistic designs is the achievable resolution and
linearity. Op-amp circuits with large loop gain help linearize transfer functions;
this feature is often removed when migrating to simplified circuits. For instance, the
amplifier scheme of Fig. 4.7c is linear only to approximately 9-bit resolution. In
cases where simplicity sacrifices precision, it is attractive to consider digital means
for recovering conversion accuracy. Digitally assisted architectures are therefore
the topic of the next section.
4.5 Digitally Assisted Architectures
Technology scaling has significantly reduced the energy per operation in CMOS
logic circuits. As explained in [39], the typical 0.7 scaling of features along with
aggressive reductions in supply voltage have led in the past to a 65% reduction in
energy per logic transition for each technology generation.
As illustrated in Fig. 4.8a, a 2-input NAND gate dissipates approximately 1.3 pJ
per logic operation in a 0.5 mm CMOS process. The same gate dissipates only 4 fJ in
a more recent 65 nm process; this amounts to a 325-fold improvement in only
12 years. The corresponding reduction in ADC conversion energy considered in
Sect. 4.2 amounts to a 64-fold reduction over the same time. This means that the
relative “cost” of digital computation in terms of energy has reduced substantially
in recent years.
To obtain a feel for how much logic can be used to “assist” a converter for the
purpose of calibration and error correction, it is interesting to express the conver-
sion energy of ADCs as a multiple of NAND-gate energy. This is illustrated in
Fig. 4.8b assuming ENAND ¼ 4 fJ and FOM ¼ 100 and 500 fJ/conversion, respec-
tively. At low resolutions, e.g., ENOB ¼ 5, a single A/D conversion consumes as
much energy as toggling approximately 1,000 logic gates. On the other hand, at
ENOB ¼ 16, several million logic gates need to switch to consume the energy of
a single A/D conversion at this level of precision.
The consequence of this observation is that in a low-resolution converter, it is
unlikely that tens of thousands of gates can be used for digital error correction in the
high-speed signal path without exceeding reasonable energy or power limits.
A large number of gates may be affordable only if the involved gates operate at
103
Logic Gate Energy ENAND (fJ)
102
101
100
0 100 200 300 400 500 600
Technology (nm)
b
107
106
(P/fnyquist) / ENAND
105
104
103
FOM =100 fJ/conv−step
102
4 6 8 10 12 14 16
ENOB (bits)
Fig. 4.8 Trends in ADC performance. (a) Typical energy per logic transition values (2-input
NAND gate) for standard VT CMOS technologies. (b) Ratio of ADC conversion energy (P/fNyquist)
and energy per logic transition (2-input NAND gate in 65 nm CMOS)
a low activity factor, e.g., if they are outside the signal path, or if they can be shared
within the system. Conversely, in high resolution ADCs, each analog operation is
very energy consuming and even a large amount of digital processing may be
accommodated in the overall power or energy budget.
In the following, a non-exhaustive discussion of opportunities for leveraging
digital logic gates in the design of ADCs is provided.
4.5.1 Oversampling
The longest-standing example of an architecture that efficiently leverages digital

signal processing abilities is the oversampling delta-sigma converter. This architecture
uses noise-shaping to push the quantization error outside the signal band [22].
Subsequent digital filtering creates a high-fidelity output signal, while the constituent
analog sub-circuits require only moderate precision. Even in fairly old technologies, it
was reasonable to justify high gate counts in the converter’s decimation filter, simply
because the conversion energy for typical high-SNDR converters is very large.
A new paradigm that might gain significance in the future is the use of significant
oversampling in traditional Nyquist converters. An example of such an ADC is
described in [40]. As was noted from Fig. 4.5b, migrating a converter with a fixed
sampling rate to technologies with higher fT can help improve power efficiency.
Ultimately, however, there is diminishing return in this trend due to the weak-
inversion “knee” of MOS devices as shown in Fig. 4.5a. The value of gm/ID no
longer improves beyond a certain minimum gate overdrive; it therefore makes no
sense to target a transistor fT below a certain value. This, in turn, implies that for
optimum power efficiency, an ADC should not be operated below a certain clock
rate. For example, consider the fT versus gm/ID plot for 45 nm technology in
Fig. 4.5b. For gm/ID > 20 S/A, fT drops sharply without a significant increase in
gm/ID. At this point, fT ¼ 50 GHz, implying that is still possible to build a switched
capacitor circuit with fclock ffi 50 GHz/50 ¼ 1 GHz.
To date, there exists only a limited number of applications for moderate- to high-
speed ADCs that require such high sampling rates, and there will clearly remain
a number of systems in the future that demand primarily good power efficiency at only
moderate speeds. A solution to this situation could be to oversample the input signal by
a large factor and to remove out-of-band noise (thermal noise, quantization noise, and
jitter) using a digital filter. Per octave of oversampling, this increases ADC resolution
by 0.5 bit. In a situation where a converter is purely limited by noise, this improvement
is in line with the fundamental thermal noise tradeoff expressed in (4.6).
4.5.2 Time-Interleaving
Time-interleaved architectures [41] exploit parallelism and trade hardware com-

plexity for an increased aggregate sampling rate. Time-interleaving can be
beneficial in several ways. First, it can help maximize the achievable sampling rate
in a given technology. Second, time-interleaving can be used to assemble a very fast
converter using sub-ADCs that do not need to operate at the limits of a given
architecture or technology. As discussed above, this can help improve energy
efficiency.
A well-known issue with time-interleaved architectures, however, is their sensi-
tivity to mismatches between the sub-converters. The most basic issue is to properly
match the offsets and gains in all channels. In addition, for the common case of
architectures that do not contain a global track-and-hold circuit, bandwidth and
clock timing mismatches are often critical [42].
Using digital signal processing techniques to address analog circuit mismatch in
time-interleaved converter arrays is an active research area [43]. Basic algorithms
that measure and remove gain and offsets in the digital domain have become
mainstream (see e.g. [44]), while techniques that address timing and bandwidth
mismatches are still evolving.
As with most digital enhancement techniques, the problem of dealing with
timing and bandwidth mismatches consists of two parts: an algorithm that
estimates the errors and a mechanism that corrects the errors. For the correction
of timing and bandwidth errors, digital methods have been refined substantially
over the years [45]. Nonetheless, the complexity and required energy of the
proposed digital filters still seem to be beyond practical bounds, even for today’s
fine-line technology. As a consequence, timing errors in practical time-
interleaved ADC are often adjusted through digitally adjustable delay lines [1,
2, 46], while the impact of bandwidth mismatches is typically minimized by
design.
As far as estimation algorithms are concerned, there is a wide and growing
variety of practical digital domain algorithms [43]. Particularly interesting are
techniques that extract the timing error information “blindly,” without applying
any test or training signals [47, 48]. It is clear that algorithms of this kind will
improve further and find their applications in practical systems.
4.5.3 Mismatch Correction
Assuming constant gate area (W L), transistor matching tends to improve in

newer technologies. In matching-limited flash ADC architectures, this trend has
been exploited in the past to improve the power efficiency by judiciously down-
sizing the constituent devices [14]. In order to scale such architectures more
aggressively, and at the same time address new sources of mismatch in nano-
scale technologies, it is desirable to aid the compensation of matching errors
through digital means.
In flash ADCs, there are several trends in this direction. As illustrated in Fig. 4.9,
the first and most transparent idea is to absorb offset voltages (Vos) in each
comparator using dedicated “trim DACs” or similar circuits that allow for a digital
threshold adjustment [49–51]. The input code for each DAC can be determined at
Analog Digital
+VOS–
Vin Encoder Dout
+
(High Speed)
–
Calibration
D/A Logic
(Low Speed)
Fig. 4.9 One slice of a flash ADC showing a comparator with trimming-DAC-based offset
calibration
start-up or whenever the converter input is in an idle condition. Alternatively, and

for improved robustness, it is also possible to adjust the trim codes continuously in
the background (during normal operation), e.g., using a chopper-based approach
[52] or through a two-channel architecture [53].
An important aspect of the arrangement in Fig. 4.9 is that most of the digital
circuitry required to control the trim DACs is either static during normal operation
or can run at very low speeds. This means that there is often no concern about the
digital energy overhead for calibration.
In modern technologies, and when near-minimum-size devices are being used,
the comparator offsets may necessitate relatively complex trim-DACs to span the
required range with suitable resolution. One idea to mitigate this issue is to include
redundant comparators, and to enable only the circuits that fall into the (reduced)
trim-DAC range after fabrication [54]. This scheme can yield good power effi-
ciency as it attacks the mismatch problem along two degrees of freedom. Larger
offsets can now be accommodated using smaller trim-DACs, which in turn imposes
a smaller overhead in terms of area and parasitic capacitance introduced to signal-
path nodes.
The idea of using redundant elements can be pushed further to eliminate the
trim-DACs and the trim-range issue altogether. In [41], redundancy and compar-
ator reassignment are employed to remove all offset constraints. A similar
concept is proposed in [40] (see also [55]), but instead of a static comparator
reassignment, a fault tolerant digital encoder is used to average the statistics
along the faulty thermometer code. While this approach is conceptually very
elegant, it requires a relatively large number of logic operations per sample in the
high-speed data path. In light of the data from Fig. 4.8 (for low ENOB values),
such a solution may be efficient only for very aggressively scaled technologies, i.
e., 45 nm and below.
Thus far, the concept of using redundant elements has not yet found widespread
use. It remains to be seen if these techniques will become a necessity in technologies
for which the variability of minimum-size devices cannot be efficiently managed
using trim-DACs. Extrapolating beyond flash ADCs, it may one day be reasonable
and advantageous to provide redundancy at higher levels of abstraction, for instance
through extra channels in a time-interleaved ADC array [56].
4.5.4 Digital Linearization of Amplifiers
As pointed out in Sect. 4.4, power-efficient and “minimalistic” design approaches

are typically unsuitable for high-resolution applications, unless appropriate digital
error correction schemes are used to enhance conversion linearity. A generic block
diagram of such a scheme is shown in Fig. 4.10. In [32], it was demonstrated that a
simple open-loop differential pair used in a pipeline ADC can be digitally
linearized using only a few thousand logic gates. In ADCs, the concept of digital
amplifier linearization has so far been applied to the pipelined architecture [32,
57–62]. However, it is conceivable to implement similar schemes in other
architectures, e.g., delta-sigma modulators [63].
One key issue in most digital linearization schemes is that the correction
parameters must track changes in operating conditions relatively quickly, prefera-
bly with time constants no larger than 1–10 ms. Unfortunately, most of the basic
statistics-based algorithms for coefficient adaptation require much longer time
constants at high target resolutions [59, 60]. Additional research is needed to extend
the recently proposed “split-ADC” [64, 65] and feed-forward noise cancelation
techniques [66] for use in nonlinear calibration schemes.
4.5.5 Digital Correction of Dynamic Errors
Most of the digital correction methods developed in recent years have targeted the
compensation of static circuit errors. However, it is conceivable that in the future
the correction of dynamic errors will become attractive as well.
Analog Digital
Nonlinearity Inverse
Dout Dout, corr

Vin
Para-
meters
Modulation
Parameter
Measurement
Fig. 4.10 Digital linearization of signal path amplifiers

One opportunity is the digital correction of errors due to finite slew rate, incomplete
settling or incomplete reset of amplifiers [34, 58, 67, 68]. Another opportunity lies in
correcting high-frequency nonlinearities introduced at the sampling front-end of high-
speed ADCs [69–71]. Correcting high-frequency distortion digitally is particularly
attractive in applications that already have a complex digital back-end that can easily
provide additional resources (e.g., inside an FPGA (field-programmable gate array)).
Such applications include, for instance, wireless base stations [69] and test and
measurement equipment. Most digital high-frequency linearization schemes proposed
to date are based on relatively simple, low-order non-linearity models. However, if
digital capabilities in nano-scale technologies continue to improve, dynamic compen-
sation schemes based on relatively complex Volterra series may become feasible [72].
4.5.6 System-Synergistic Error Correction Approaches
In the discussion so far, ADCs were viewed as “black boxes” that deliver a set
performance without any system-level interaction. Given the complexity of today’s
applications, it is important to realize that there exist opportunities to improve ADC
performance by leveraging specific system and signal properties.
For instance, large average power savings are possible in radio receivers when
ADC resolution and speed are dynamically adjusted to satisfy the minimum
instantaneous performance needs. The design described in [73] demonstrates the
efficiency of such an approach.
In the context of digital correction, it is possible to leverage known properties of
application-specific signals to “equalize” the ADC together with the communica-
tion channel [74–79]. For instance, the converter described in [77] uses the system’s
OFDM (orthogonal frequency-division multiplexing) pilot tones to extract compo-
nent mismatch information (see Fig. 4.11). In such an approach, existing system
ADC7
baseband processor
pilot
+ ADC6 estim.
–
–
FFT +
Vin
pilots pilot
errors
+ ADC2
– calibration
logic
+ ADC1
– offset adjustments
Fig. 4.11 Block diagram of a digitally assisted, time-interleaved ADC using a system-synergistic
calibration approach
hardware, for instance the FFT (fast Fourier transform) block, can be re-used to
facilitate ADC calibration.
4.6 Trends in Delta-Sigma A/D Converters
As considered in the previous sections, designers of analog circuits have to face

various challenges associated with each new technology generation. New concepts
at the system level as well as for sub-circuit blocks must be developed due to
reduced supply voltage and reduced intrinsic gain per transistor while benefitting
from increased transit frequency fT. The following section illustrates recent
approaches in delta-sigma ADCs in order to deal with these limitations. As will
be seen, many of the concepts considered in Sects. 4.4 and 4.5 have already found
their application in this kind of ADC, however, mostly as “proof of concepts” in
prototypes implemented by research-orientated institutions, not in industrial mass-
market products.
Delta-sigma ADCs have been with us for some decades. Today, they achieve
medium resolution in signal bandwidths of several tens of megahertz, e.g., 12 bits
in a signal bandwidth of 20 MHz [80], according to the well-known equation for
the peak signal-to-quantization-noise ratio (SQNR) of an ideal modulator given
by [81]
2Nþ1
3p B OSR
SQNRp;ideal ¼ ð2 1Þ ð2N þ 1Þ
2
: (4.8)
2 p
In (4.8), N is the order of the loop filter and B the number of bits of the internal
quantizer according to Fig. 4.12. The oversampling ratio OSR is defined as
a b
loop
B bit
loop B bit
filter filter quantizer
quantizer fS
x(t) x(t)
L0(s) fS L0[z]
y[n] v[n] y[n]
v(t)
u(t) u[n]
L1(s) L1[z]
e[n] e[n]
kq + kq +
DAC[z]
DAC(s)
Fig. 4.12 System-theoretical model of a single-stage delta-sigma modulator, (a) Continuous-time

modulator, (b) Discrete-time modulator
fs
OSR ¼ (4.9)
2fb
where fs is the sampling frequency and fb the signal bandwidth.

As illustrated in Fig. 4.12, the heart of any delta-sigma ADC, i.e., the delta-
sigma modulator, may be implemented in either the discrete-time (DT) or the
continuous-time (CT) domain. The loop filter is mostly implemented by means of
active RC or gmC integrators in the case of CT implementations, or by means of
switched-capacitor (SC) or switched–op-amp (SO) integrators in case of DT
implementations. Thus, the specifications for the unity-gain bandwidth product of
operational amplifiers for CT implementations are relaxed in comparison to SC
implementations due to relaxed settling requirements, an important characteristic
particularly in high-frequency applications. Consequently, CT implementations are
considered to be more power efficient.
At first glance, this characteristic seems to be confirmed by Fig. 4.13a, where
the FOM according to (4.1) is plotted versus the signal bandwidth. These data
were collected from the survey provided in [9], extended by delta-sigma ADCs
published in IEEE Journal of Solid-State Circuits from 2000 to 2010. Most
outstanding is a factor of 10 between the FOM of the most efficient CT and
DT implementation for a signal bandwidth of 20 MHz [80, 82]. However, in
order to perform a fair comparison, i.e., following Sect. 4.2.1 by considering the
FOM of implementations with similar bandwidth and resolution, Fig. 4.13b
should be considered. Obviously, two pairs of designs achieve almost the same
resolution in a bandwidth of 20 MHz, i.e., 50 dB and 63 dB [82–85]. Interest-
ingly, however, and contrary to the above, the FOM of the DT implementation
outperforms the FOM of the corresponding CT implementation by factors of
2 and 3, respectively.
How can this discrepancy be explained? First, the calculation of the FOM does
usually not take into account that CT modulators offer inherent anti-aliasing
filtering without additional power consumption. By contrast, this filtering must
be explicitly implemented outside of a DT modulator, thus resulting in additional
power consumption. Second, it is often not clearly stated in publications whether
further sub-circuit blocks of a delta-sigma modulator are implemented on-chip or
off-chip, e.g., clock generation, reference generation, or digital cancelation filters
(DCFs) as needed for cascaded delta-sigma modulators (see Sect. 4.6.1). Even if
so, the power consumption of each sub-circuit block is not always given sepa-
rately. Indeed, the CT cascaded design achieving 63 dB included the implemen-
tation and thus the power consumption of the on-chip clock and reference
generator as well as of the DCF [83]. By contrast, the calculation of the FOM of
its DT counterpart was based on the power consumption of the modulator only
[82]. Similar observations were made for the CT and DT designs achieving 50 dB
[84, 85]. Thus, care must be taken even in comparing designs with similar
characteristics. In order to clearly demonstrate or demystify the promoted
power efficiency of CT implementations, it would be interesting to see a compar-
ison between the implementation of a DT and a CT delta-sigma modulator
a
FOM (pJ/conversion−step) 102
[84]
101
[85]
[83]
[82]
0
10
discrete−time
continuous−time
10−1
10 20 30 40 50 60 70
bandwidth (MHz)
b
85
80
75
70
65 [83]
SNDR (dB)
[82]
60
55
[85]
50
[84]
45
discrete−time
40
continuous−time
35
5 10 15 20 25 30
bandwidth (MHz)
Fig. 4.13 Comparison of FOM and SNDR between DT and CT delta-sigma modulators
exhibiting a bandwidth greater than 2 MHz. FOM (a) and SNDR (b) versus bandwidth
designed according to identical specifications. Furthermore, the same sub-circuit

blocks should be implemented while being optimized according to the minimum
requirements of a DT and a CT implementation.
For these reasons, it has yet to be seen in the future whether CT modulators will
become the vehicle for low-power implementations or whether DT
implementations keep pace with them. At present, CT architectures are the favored
architectures for the implementation of broadband or high-speed delta-sigma ADCs
as illustrated in Fig. 4.13. From this trend, it may be concluded that they will also
become the architecture of the future. In the following, considerations are thus
illustrated based on CT delta-sigma modulators.
From the linearized model shown in Fig. 4.12a, i.e., replacing the strongly
nonlinear quantizer by a linear model consisting of a quantizer gain kq and an
additive white noise source e [22], the signal transfer function (STF) and noise
transfer function (NTF) result in
y½n ¼ yx ½n þ ye ½n

¼ STFx½n þ NTF½ze½n
kq L0 ðsÞ 1
¼ x½n þ e½n (4.10)
1 þ kq L1 ½z 1 þ kq L1 ½z
where L1 [z] is the discrete-time equivalent of the product of DAC(s) and L1 (s) due
to the sampling process. Note that no argument for the STF is given since it is equal
to a mixture between a continuous-time and a discrete-time transfer function. This
aspect results in the anti-aliasing property, as will be discussed shortly.
Formerly, high resolution was achieved by applying a high OSR in a first-order
modulator, i.e., larger than 100, while using a single-bit quantizer for the implementa-
tion of the internal quantizer. The latter was also of advantage since no additional
dynamic element matching technique must be implemented in order to linearize
the transfer characteristic of a multi-bit DAC. With steadily increasing demand for
higher signal bandwidths, the OSR has become quite small. Today, an OSR between
10 and 20 represents a typical value in applications with signal bandwidths up to
20 MHz (e.g., [80, 86]). To compensate for the loss of resolution due to the reduction of
the OSR, either the order of the loop filter or the resolution of the internal quantizer has
been increased. However, both approaches have their limitations, because of either
architectural or technology constraints, of which those for the loop filter will be
considered first. At the same time, trends for overcoming these limitations will be
outlined.
4.6.1 Loop Filter
The loop filter of a delta-sigma modulator is typically implemented either as a

cascade of integrators in feedback (CIFB) or in feed-forward (CIFF) [22]. Both
Fig. 4.14 Typical loop filter a

implementations illustrated x(t) v(t)
by the example of a second- + k1fss–1 + k2fss–1
order filter, (a) Cascade of – –
integrators in feedback u(t)
(CIFB). (b) Cascade of
integrators in feed-forward
(CIFF) b
+
x(t) v(t)
+ k2fss–1 k1fss–1 +
–
u(t)
a 10000
b 10000
8000 8000
6000 6000
4000 4000
2000 2000
0
0
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
sampling value / Vref sampling value / Vref
Fig. 4.15 Signal swing at the output of the integrators of the loop filters shown in Fig. 4.14.
(a) First integrator, (b) Second integrator (CIFB: solid line; CIFF: solid line, circles; CIFF with
feed-in path from x(t) to v(t): solid line, crosses)
types are illustrated in Fig. 4.14 by means of a continuous-time second-order loop

filter. An implementation of the same transfer function L1(s), and thus the same
NTF according to (4.10), is possible by either architecture.
However, it is most likely that the CIFF architecture will become the dominant
architecture in the near future. In comparison to a CIFB architecture, the output swing
of each integrator is highly reduced, except for the one in front of the quantizer. If
an additional feed-in path from x(t) to v(t) is inserted (Fig. 4.14b, dashed line),
the loop-filter has to process almost only the quantization noise. Thus, the last
integrator also exhibits a highly reduced signal swing at its output. These
characteristics are illustrated in Fig. 4.15. For simulation purposes, a sine wave
input signal was applied to the modulators (fsig ¼ fb, û ¼ 1.9 dBFS (decibels
relative to full scale), OSR ¼ 165, k1 ¼ 6/5, k2 ¼ 45/32, 4-bit quantizer).
According to the simulation results, a reduction of the output swing by a factor
approximately equal to five can be noticed. Thus, it can be concluded that this
20
10
0
|STF| (dB)
−10
−20
−30
CIFB
−40 CIFF
CIFF with feed−in
−50
10−2 10−1 100
f / fs
Fig. 4.16 Signal transfer functions of the different architectures based on the loop filters shown in
Fig. 4.14
architecture offers not only a better robustness against distortion but also a high
attractiveness for implementations in new technology nodes in which the supply
voltage has been reduced. On the other hand, the suppression of the implicit anti-
aliasing filter around multiples of fs, as offered by the STF of any CT architecture, is
less effective in a CIFF architecture than in a CIFB architecture. This characteristic
is illustrated in Fig. 4.16.
Although the loop-filter architectures shown in Fig. 4.14 implement the same
filter function L1(s), L0(s) results in
2
fs
L0;CIFB ðsÞ ¼ k1 k2 ; (4.11)
s
2
fs fs
L0;CIFF ðsÞ ¼ k2 þ k1 k2 ; (4.12)
s s
2
fs fs
L0;CIFFþfeedin ðsÞ ¼ 1 þ k2 þ k1 k2 : (4.13)
s s
From (4.12) and (4.13), it can be seen that the signal path around the second
integrator, or both integrators, acts as a short for frequencies larger than fs.
Thus, the suppression of high frequencies by L0(s) and thus the implicit anti-
aliasing feature of a CIFF architecture is less effective than that of a CIFB
architecture.
Taking the scaling of the loop filter into account, a more detailed analysis of the
achievable SQNR of an ideal single-stage delta-sigma modulator results in [87]
!
Y
SQNRp;scaling ¼ 10 log ðSNRp;ideal Þ þ 20 log kq ki ; (4.14)
i
i.e., the resolution also depends on the set of scaling coefficients and the quantizer
gain. Typically, the product of the scaling coefficients is smaller than one, particu-
larly in high-order single-stage modulators where less aggressive scaling must be
applied for stability reasons [88]. The achievable SQNRp,scaling is thus reduced
compared to SQNRp,ideal and single-stage modulators with loop filter order higher
than four are not attractive.
It can be foreseen that, driven by the ever increasing demand for higher bandwidth
and resolution in combination with reconfigurability demanded, for example, by new
telecommunication standards, delta-sigma modulators must be pushed beyond their
present performance limits in the near future. Towards these goals, cascaded delta-
sigma modulators offer a promising alternative to single-stage architectures. In these
architectures, as illustrated by the example of a CT 2-2 modulator in Fig. 4.17, the
quantization errors of all stages except for the last are canceled via digital cancelation
filters (DCFs) at the output y[n]. For the quantization error of the last stage, the cascade
provides noise-shaping with the order of the overall modulator, i.e., four in the case of
the 2-2 modulator. Taking scaling into account, the SQNRp,scaling of an ideal cascaded
modulator consisting of M stages with order Ni in the ith stage results in
SQNRp;scaling ¼ 10 log ðSNRp;ideal Þþ

" ! #
Y
M Y Ni Y
M1
(4.15)
20log10 kij kqM kiNi ðiþ1Þ 1 ;
i¼1 j¼1 i¼1
fs
x(t) y1[n]
+ k11I(s) + k12I(s) DCF1[z]
– –
DAC1(s)
y[n]
k1221 k1021 +
–
+ fs
y2[n]
+ k21I(s) + k22I(s) DCF2[z]
– –
DAC2(s)
Fig. 4.17 Continuous-time 2-2 cascaded modulator

i.e., the performance loss is given by the product of the single-stage scaling
coefficients, the quantizer gain of the last stage and the scaling coefficients of
each signal path connecting a quantizer input to the input of the next stage.
Obviously, the major advantage of cascaded modulators consists in the cascad-
ing of low-order modulators. Thus, the performance loss due to scaling is less
severe since each modulator in the cascade can be scaled more aggressively than a
single-stage modulator whose order is equal to the overall order of the cascade.
Furthermore, the values of the scaling coefficients between stages may become
larger than one if multi-bit quantizers are used. Indeed, it has been reported that an
ideal CT 3-1 cascaded modulator with four-bit quantizers outperforms a fourth-
order single stage modulator in terms of resolution by three bits, and a CT 2-2 cascaded
modulator by even more than four bits [89]. However, whether these values are
achievable in an implementation, where several sources of non-idealities have to be
considered as well, is still to be proven.
By switching later stages in or out of the cascade, cascaded architectures also
offer a promising approach to reconfigurability in terms of bandwidth and resolu-
tion, as postulated in Sect. 4.5.6. The reasons why they have not yet replaced single-
stage architectures despite all these advantages are manifold.
First, they rely on perfect matching between the analog and the digital parts of
the modulator, i.e., the loop filters and the digital cancelation filters. Otherwise,
low-order shaped quantization noise of earlier stages in the cascade leaks to the
output, drastically reducing the resolution. However, since digitally assisted circuits
may become available nearly “for free” in the future, these matching problems may
easily be overcome. Different on-line and off-line approaches to their compensation
have already been presented [90, 91].
Furthermore, a new architecture for a cascaded delta-sigma modulator has been
presented recently [92]. As illustrated in Fig. 4.18, this kind of cascaded modulator
performs a cancelation of the quantization error of the first stage without the need
for digital cancelation filters. Thus, the non-ideality of a mismatch between the
scaling coefficients of the analog and digital parts of a classical cascaded modulator
is overcome. Additionally, it was proven by means of a first prototype that opera-
tional amplifiers with a DC gain as low as 35 dB may be used for a successful
implementation, thus making it attractive for the implementation in a deep sub-
x[n] y[n]
L10[z] + +
– –
L11[z]
–
e1[n]
+
Fig. 4.18 System-theoretical L20[z] +

model of a sturdy multi-stage –
noise-shaping (SMASH) L21[z]
delta-sigma modulator
micrometer CMOS technology with reduced intrinsic gain per transistor [93]. For
these reasons, this kind of cascaded delta-sigma modulator is called a sturdy
multistage noise-shaping (SMASH) delta-sigma modulator. However, at first
glance, it seems that a SMASH modulator sacrifices the apparently most important
advantage of a MASH modulator in comparison to a single-stage modulator: the
guarantee of overall stability of the cascade by cascading low-order modulators. As
is obvious from Fig. 4.18, the second stage is embedded inside the loop-filter of the
first stage, its order increasing accordingly.
A second reason is given by the characteristic that cascaded architectures
consume more power and need more area for their implementation since, besides
the DCFs, at least one additional quantizer becomes mandatory.
These characteristics are even more apparent in a time-interleaved delta-sigma
modulator, another alternative for the implementation of a broadband ADC apply-
ing small-to-moderate oversampling. Time-interleaved delta-sigma modulators
consist of n identical delta-sigma modulators working in parallel, as illustrated in
Fig. 4.19 (see also Sect. 4.5.2). Thus, the sampling rate fs of each modulator can be
reduced by a factor of n while the same resolution is provided as offered by a single-
stage modulator clocked at fs. This reduction becomes feasible since the outputs of
all n modulators are combined in order to achieve the same resolution at an output
rate fs. However, only a few time-interleaved delta-sigma modulators have been
presented so far, consisting of two paths only [84, 94–96].
Interestingly, the concepts for cascading and time-interleaving delta-sigma
modulators have been known for more than 15 years [97]. However, they gained
renewed interest only recently, particularly due to demand for broadband ADCs
with the capability of reconfiguration in bandwidth and resolution. In the near
future, it is to be expected that research on these architectures will be increased in
order to overcome major non-idealities and to reduce their power consumption and
area requirement, e.g., by op-amp sharing as presented in [84, 96].
Finally, first prototypes of minimalistic delta-sigma modulators as introduced in
Sect. 4.4 were recently presented, i.e., the loop-filter was implemented based on
simple inverters instead of traditional op-amps [27, 28].
x[n]
n ΔΣ-ADC1 n z−(n−1)
z−1 n ΔΣ-ADC2 n z−(n−2) +
y[n]
z−(n−1) n ΔΣ-ADCn n +
Fig. 4.19 System-theoretical model of a DT time-interleaved delta-sigma modulator

4.6.2 Quantizer
In the last subsection, trends for the implementation of the loop filter of a delta-sigma
modulator were considered. Therein, special emphasis was placed on broadband
applications, i.e., bandwidths in the range of several tens of megahertz, which allow
only low oversampling to be applied. Furthermore, the aspects of highly scaled
CMOS technologies were taken into account, i.e., reduced supply voltage and
reduced gain per transistor. In the following, another key component of a delta-
sigma modulator is considered: the quantizer.
Traditionally, the quantizer is implemented by means of a flash ADC. However,
with upcoming supply voltages below 1 V, multi-bit flash quantizers can hardly be
implemented. Considering the example of a 4-bit quantizer, the LSB (least signifi-
cant bit) becomes as small as 62.5 mV. Since this value represents a typical value
for the offset voltage of a CMOS comparator, either an offset compensation
technique must be applied or a rather large input transistor must be implemented.
Besides, multi-bit flash ADCs are very power consuming.
First attempts to overcome these hindrances covered the replacement of the
multi-bit flash ADC by means of a successive approximation ADC or a tracking
quantizer [98, 99]. However, the first approach requires an additional clock running
at multiples of the original sampling frequency, which makes it less attractive for
broadband applications. The second approach only reduces the number of
comparators while still three comparators were needed for the replacement of a
4-bit flash ADC in order to guarantee good tracking capability.
Recently, two new concepts were presented that directly comply with the benefits
of scaled CMOS technologies for digital applications. They thus may become the
favorite architectures for the implementation of a multi-bit quantizer in the near future.
4.6.2.1 Voltage-Controlled Oscillator Based Quantizer
In the first approach, the flash ADC is replaced by a voltage-controlled oscillator

(VCO), as illustrated in Fig. 4.20 [100]. The voltage Vtune, which is used to tune the
frequency of the VCO, is equal to the output voltage v(t) of the loop filter in
Fig. 4.12. By means of registers, it is observed whether a VCO delay cell undergoes
a transition or not within a given clock period. This is performed by comparing
samples of the current and the previous states of the VCO by XOR gates. The
resulting number of VCO delay cells that undergo a transition is a function of the
delay of each stage as set by Vtune. Thus, a quantized value of Vtune is obtained.
The advantages of this approach with respect to future CMOS technology nodes
are obvious. First, this implementation is a highly digital implementation; it relies on
inverters, registers, XOR gates and one adder. An area- and power-efficient imple-
mentation thus becomes feasible. Second, any offset-related non-idealities are over-
come since a CMOS inverter achieves full voltage swing at its output. For the same
reason, the number of delay elements and thus the resolution can easily be increased.
Vtune
ring
oscillator
clock
register I
register II
XOR gates
digital out
Fig. 4.20 Voltage-controlled oscillator based quantizer
Third, it is also to be expected that this architecture will take advantage of future
technology nodes since the resolution versus sampling speed tradeoff directly
improves with the reduced gate delays in modern CMOS processes. This characteristic
makes it very attractive for high-speed operations with small latency. Furthermore,
this architecture provides first-order noise-shaping for the quantization error, whereby
the order of the loop filter can be reduced by one; thus, one op-amp can be saved. It also
provides inherent data-weighted averaging for its digital output code, for which the
linearity requirements imposed on the DAC in the feedback path become less stringent
while making an explicit technique for dynamic element matching (DEM) obsolete.
A first prototype applying a VCO-based quantizer in a third-order delta-sigma
modulator (0.13 mm CMOS, 1.5 V supply voltage) achieved 67 dB signal-to-noise-
and-distortion ratio (SNDR) in a bandwidth of 20 MHz [100]. In a redesign, the
SNDR was improved to 78 dB by quantizing the phase instead of the frequency of
the VCO [101]. Thereby, harmonics generated by the nonlinear tuning gain of the
VCO caused by its nonlinear voltage-to-frequency transfer characteristic were
sufficiently suppressed. However, an explicit DEM technique became mandatory.
4.6.2.2 Time Encoding Quantizer
In the second approach, the multi-bit flash ADC is replaced by a time-encoding

quantizer [102]. It consists of a 1 bit quantizer embedded in a feedback system
followed by some filter blocks and a decimator, as illustrated in Fig. 4.21. The
fOSD
v(t) u[n] y[n]

+ H[z] sinc[z] M
u(t)
H(s) DAC(s)
Fig. 4.21 Time encoding quantizer
feedback loop forces u(t) to oscillate at frequency fc with an amplitude A if no input

signal is applied, i.e., v(t) ¼ 0, and if the quantizer is not sampled with fOSD. In order
to establish this limit cycle, the filter H(s) in the feedback loop must be designed
according to
4fc s= ð4 fc Þ
H ðsÞ ¼ A e : (4.16)
s
However, if the quantizer is sampled with a frequency fOSD sufficiently larger

than fc, then the feedback system can be considered as a quasi-continuous-time
system. In this case, the limit cycle is still established and its frequency is almost not
influenced if an input signal v(t) is applied.
By means of this approach, most of the power of the quantization error of the
1-bit quantizer is located around fc, which is designed to be much larger than fb.
The remaining power within the signal band resembles that of a multi-bit quantizer.
Thereby, the 1-bit quantizer can be designed to function like an arbitrary multi-bit
quantizer while overcoming any offset related non-idealities.
The output signal of the feedback system is filtered by an equalization filter H[z].
This filter is a digital replica of H(s), where the transfer function from v(t) to u[n] is
set equal to one within the signal band. Subsequently, the quantization noise power
located around fc is removed by a sinc decimation filter whose zeros are placed
accordingly. Furthermore, the multi-bit encoding is performed by this filter block
and the output rate is decimated to that of an equivalent multi-bit counterpart. Thus,
the following further oversampling ratios have been defined for a delta-sigma
modulator with an internal time-encoding quantizer:
fc
COSR ¼ (4.17)
fs
fOSD
ROSR ¼ : (4.18)
fc
First prototypes achieved an SNDR of 10 bits and 12 bits in signal bandwidths of

17 MHz and 6.4 MHz, respectively [103]. In both implementations, the time-
encoding quantizer replaced a 4-bit flash quantizer. However, fOSD was set equal to
1.632 GHz and thus became a factor of four larger compared to the sampling
frequency fs of an equivalent 4-bit implementation. The application of a time-
encoding quantizer in a broadband modulator may thus be limited.
4.7 Conclusion
This chapter was intended to discuss trends in the context of low-power, high-
performance A/D conversion. Using survey data from the past 13 years, it was
observed that power efficiency in ADCs has improved at an astonishing rate
of approximately 2 every 2 years. In part, this progress rate is based on cleverly
exploiting the strengths of today’s technology. Smaller feature sizes improve the
power dissipation in circuits that are not limited by thermal noise. In circuit elements
that are limited by thermal noise, exploiting the high transit frequency fT of modern
transistors can be of help in mitigating a penalty from low supply voltages.
A promising paradigm is the trend toward minimalistic ADC architectures and
digital means of correcting analog circuit errors. Digitally assisted ADCs aim to
leverage the low computing energy of modern processes in order to improve the
resolution and robustness of simplified circuits. Future work in this area promises to
fuel further progress in optimizing the power efficiency of ADCs.
Overall, improvements in ADC power dissipation are likely to come from a
combination of aspects that involve improved system embedding and reducing
analog sub-circuit complexity and raw precision at the expense of “cheap” digital
processing resources. These trends can already be observed in delta-sigma ADCs,
which have been considered in detail.
References
1. Poulton, K., et al.: In: ISSCC Digest of Technical Papers, pp. 318–496. IEEE (2003)
2. Schvan, P., et al.: In: ISSCC Digest of Technical Papers, pp. 544–634. IEEE (2008)
3. Abidi, A.: IEEE J. Solid-State Circuits 42(5), 954 (2007)
4. Murmann, B.: In: Proceedings of Custom Integrated Circuit Conference, pp. 105–112. IEEE
(2008)
5. Chiu, Y., Nikolic, B., Gray, P.: In: Proceedings of Custom Integrated Circuit Conference,
pp. 375–382. IEEE (2005)
6. Kenington, P., Astier, L.: IEEE Trans. Veh. Technol. 49(2), 643 (2000)
7. Merkel, K, Wilson, A.: In: Proceedings of Aerospace Conference, Big Sky, pp. 2415–2427.
IEEE (2003)
8. Walden, R.: IEEE J. Select. Areas Commun. 17(4), 539 (1999)
9. Murmann, B.: ADC performance survey 1997–2010. http://www.stanford.edu/_murmann/
adcsurvey.html
10. van Elzakker, M., et al.: In: ISSCC Digest of Technical Papers, pp. 244–610. IEEE (2008)
11. Verbruggen, B., et al.: In: ISSCC Digest of Technical Papers, pp. 252–611. IEEE (2008)
12. Verbruggen, B., et al.: In: ISSCC Digest of Technical Papers, pp. 296–297. IEEE (2010)
13. Kawahito, S.: In: Proceedings of Custom Integrated Circuit Conference, pp. 505–512. IEEE
(2006)
14. Verbruggen, B.: High-speed calibrated analog-to-digital converters in CMOS. Ph.D. thesis,
Vrije Universiteit Brussel (2009)
15. Liu, C., et al.: In: Proceedings Symposium VLSI Circuits, pp. 236–237 (2009)
16. Hu, J., Dolev, N., Murmann, B.: IEEE J. Solid-State Circuits 44(4), 1057 (2009)
17. Devarajan, S., et al.: In: ISSCC Digest of Technical Papers, pp. 86–87. IEEE (2009)
18. Bult, K.: In: Proceedings of ESSCIRC, pp. 52–64 (2009)
19. Annema, A., et al.: IEEE J. Solid-State Circuits 40(1), 132 (2005)
20. Bult, K. The effect of technology scaling on power dissipation in analog circuits. In: Analog
Circuit Design, pp. 251–290. Springer, Berlin/Heidelberg/New York (2006)
21. Annema, A., et al.: In: ISSCC Digest of Technical Papers, pp. 134–135. IEEE (2004)
22. Schreier, R., Temes, G.C.: Understanding Delta-Sigma Data Converters. IEEE Press,
Piscataway (2005)
23. Vittoz, E.: In: Proceedings of International Symposium on Circuits Systems, pp. 1372–1375.
IEEE (1990)
24. Nakamura, K., et al.: IEEE J. Solid-State Circuits 30(3), 173 (1995)
25. Draxelmayr, D.: In: ISSCC Digest of Technical Papers, pp. 264–527. IEEE (2004)
26. der Plas, G.V., Verbruggen, B.: IEEE J. Solid-State Circuits 43(12), 2631 (2008)
27. Chae, Y., Han, G.: IEEE J. Solid-State Circuits 44(2), 458 (2009)
28. van Veldhoven, R., Rutten, R., Breems, L.: In: ISSCC Digest of Technical Papers, pp.
492–630. IEEE (2008)
29. Ahmed, I., Mulder, J., Johns, D.: In: ISSCC Digest of Technical Papers, pp. 164–165. IEEE
(2009)
30. Anthony, M., et al.: In: Proceedings of Symposium on VLSI Circuits, pp. 222–223 (2008)
31. Fiorenza, J., et al.: IEEE J. Solid-State Circuits 41(12), 2658 (2006)
32. Murmann, B., Boser, B.: IEEE J. Solid-State Circuits 38(12), 2040 (2003)
33. Nazemi, A., et al.: In: Proceedings of Symposium on VLSI Circuits, pp. 18–19 (2008)
34. Iroaga, E., Murmann, B.: IEEE J. Solid-State Circuits 42(4), 748 (2007)
35. Murmann, B.: Limits on ADC power dissipation. In: Analog Circuit Design, pp. 351–368.
Springer, Berlin/Heidelberg/New York (2006)
36. Sangster, F., Teer, K.: IEEE J. Solid-State Circuits 4(3), 131 (1969)
37. Alon, E., et al.: IEEE J. Solid-State Circuits 41(2), 413 (2006)
38. Toifl, T., et al.: IEEE J. Solid-State Circuits 44(11), 2901 (2009)
39. Borkar, S.: IEEE Micro 19(4), 23 (1999)
40. Hesener, M., et al.: In: ISSCC Digest of Technical Papers, pp. 248–600. IEEE (2007)
41. Black, W., Hodges, D.: IEEE J. Solid-State Circuits 15(6), 1022 (1980)
42. Jamal, S., et al.: IEEE J. Solid-State Circuits 37(12), 1618 (2002)
43. Vogel, C., Johansson, H.: In: Proceedings of International Symposium on Circuits System,
pp. 3386–3389. IEEE (2006)
44. Hsu, C., et al.: In: ISSCC Digest of Technical Papers, pp. 464–615. IEEE (2007)
45. Vogel, C., Mendel, S.: IEEE Trans. Circuits Syst. I, Reg. Papers 56(11), 2463 (2009)
46. Haftbaradaran, A., Martin, K.: In: Proceedings of Custom Integrated Circuit Conference,
pp. 341–344. IEEE (2007)
47. Elbornsson, J., Gustafsson, F., Eklund, J.: IEEE Trans. Circuits Syst. I, Reg. Papers 51(1), 151
(2004)
48. Haftbaradaran, A., Martin, K.: IEEE Trans. Circuits Syst. II, Exp. Briefs 55(3), 234 (2008)
49. der Plas, G.V., Decoutere, S., Donnay, S.: In: ISSCC Digest of Technical Papers, pp.
2310–2311. IEEE (2006)
50. Tamba, Y., Yamakido, K.: In: ISSCC Digest of Technical Papers, pp. 324–325. IEEE (1999)
51. Verbruggen, B., et al.: In: Proceedings of Symposium on VLSI Circuits, pp. 14–15 (2008)
52. Huang, C., Wu, J.: IEEE Trans. Circuits Syst. I, Reg. Papers 52(9), 1732 (2005)
53. Nakajima, Y., et al.: In: Proceedings of Symposium on VLSI Circuits, pp. 266–267 (2009)
54. Park, S., Palaskas, Y., Flynn, M.: IEEE J. Solid-State Circuits 42(9), 1865 (2007)
55. Frey, M., Loeliger, H.: IEEE Trans. Circuits Syst. I, Reg. Papers 54(1), 229 (2007)
56. Ginsburg, B., Chandrakasan, A.: IEEE J. Solid-State Circuits 43(12), 2641 (2008)
57. Daito, M., et al.: IEEE J. Solid-State Circuits 41(11), 2417 (2006)
58. Grace, C., Hurst, P., Lewis, S.: IEEE J. Solid-State Circuits 40(5), 1038 (2005)
59. Murmann, B., Boser, B.: IEEE Trans. Instrum. Meas. 56(6), 2504 (2007)
60. Panigada, A., Galton, I.: IEEE J. Solid-State Circuits 44(12), 3314 (2009)
61. Sahoo, B., Razavi, B.: IEEE J. Solid-State Circuits 44(9), 2366 (2009)
62. Verma, A., Razavi, B.: IEEE J. Solid-State Circuits 44(11), 3039 (2009)
63. O’Donoghue, K.,: Digital calibration of a switched-capacitor sigma-deltamodulator. Ph.D.
thesis, University of California (2009)
64. Li, J., Moon, U.: IEEE Trans. Circuits Syst. II, Exp. Briefs 50(9), 531 (2003)
65. McNeill, J., Coln, M., Larivee, B.: IEEE J. Solid-State Circuits 40(12), 2437 (2005)
66. Hsueh, K., et al.: In: ISSCC Digest of Technical Papers, pp. 546–634. IEEE (2008)
67. Kawahito, S., et al.: In: Proceedings of Custom Integrated Circuit Conference, pp. 117–120.
IEEE (2008)
68. Keane, J., Hurst, P., Lewis, S.: IEEE Trans. Circuits Syst. I, Reg. Papers 53(3), 511 (2006)
69. Nikaeen, P., Murmann, B. In: Proceedings of Custom Integrated Circuit Conference, pp.
161–164. IEEE (2008)
70. Nikaeen, P., Murmann, B.: IEEE J. Sel. Topics Signal Proc. 3(3), 499 (2009)
71. Satarzadeh, P., Levy, B., Hurst, P.: IEEE J. Sel. Topics Signal Proc. 3(3), 454 (2009)
72. Chiu, Y., et al.: IEEE Trans. Circuits Syst. I, Reg. Papers 51(1), 38 (2004)
73. Malla, P., et al.: In: ISSCC Digest of Technical Papers, pp. 496–631. IEEE (2008)
74. Agazzi, O., et al.: IEEE J. Solid-State Circuits 43(12), 2939 (2008)
75. Namgoong, W.: IEEE Trans. Wireless Commun. 2(3), 502 (2003)
76. Oh, Y., Murmann, B.: IEEE Trans. Circuits Syst. I, Reg. Papers 53(8), 1693 (2006)
77. Oh, Y., Murmann, B.: In: Proceedings of Custom Integrated Circuit Conference, pp.
193–196. IEEE (2007)
78. Sandeep, P., et al.: In: Proceedings of GLOBECOM, pp. 1–5. IEEE (2008)
79. Tsai, T., Hurst, P., Lewis, S.: IEEE Trans. Circuits Syst. I, Reg. Papers 56(2), 307 (2009)
80. Mitteregger, G., et al.: IEEE J. Solid-State Circuits 41(12), 2641 (2006)
81. Ortmanns, M., Gerfers, F.: Continuous-Time Sigma-Delta A/D Conversion. Springer, Berlin/
Heidelberg/New York (2006)
82. Paramesh, J., et al.: In: Proceedings of Symposium on VLSI Circuits, pp. 166–167 (2006)
83. Breems, L.J., Rutten, R., Wetzker, G.: IEEE J. Solid-State Circuits 39(12), 2152 (2004)
84. Caldwell, T.C., Johns, D.A.: IEEE J. Solid-State Circuits 41(7), 1578 (2006)
85. Tabatabaei, A., et al.: In: ISSCC Digest of Technical Papers, pp. 66–67. IEEE (2003)
86. Balmelli, P., Huang, Q.: IEEE J. Solid-State Circuits 39(12), 2212 (2004)
87. Geerts, Y., Steyaert, M., Sansen, W.: Design of Multi-Bit Delta-Sigma A/D Converters.
Kluwer Academic, Dordrecht (2002)
88. Norsworthy, S.R., Schreier, R., Temes, G.C.: Delta-Sigma Data Converters - Theory, Design,
and Simulation. IEEE Press, Piscataway (1997)
89. Keller, M.,: Systematic approach to the synthesis of continuous-time multi-stage noise-
shaping delta-sigma modulators. Ph.D. thesis, University of Freiburg (2010)
90. Kiss, P., et al.: IEEE Trans. Circuits Syst. II, Analog Digit. Signal Process. 47(7), 629 (2000)
91. Schreier, G., Teme, G.C.: IEEE Trans. Circuits Syst. II, Analog Digit. Signal Process 47(7),
621 (2000)
92. Maghari, N., et al.: IET Electron. Lett. 42(22), 1269 (2006)
93. Maghari, N., Kwon, S., Moon, U.: IEEE J. Solid-State Circuits 44(8), 2212 (2009)
94. Galdi, I., et al.: IEEE J. Solid-State Circuits 43(7), 1648 (2008)
95. Lee, K., et al.: IEEE J. Solid-State Circuits 43(12), 2601 (2008)
96. Lee, K.S., Kwon, S., Maloberti, F.: IEEE J. Solid-State Circuits 42(6), 1206 (2007)
97. Aziz, P.M., Sorensen, H.V., van der Spiegel, J.: IEEE Signal Process. Mag. 13(1), 61 (1996)
98. Doerrer, L., et al.: IEEE J. Solid-State Circuits 40(12), 2416 (2005)
99. Samid, L., Manoli, Y.: In: Proceedings of International Symposium on Circuits Systems,
pp. 2965–2968. IEEE (2006)
100. Straayer, M.Z., Perrott, M.H.: IEEE J. Solid-State Circuits 43(4), 805 (2008)
101. Park, M., Perrott, M.H.: IEEE J. Solid-State Circuits 44(12), 3344 (2009)
102. Hernandez, L., Prefasi, E.: IEEE Trans. Circuits Syst. I, Reg. Papers 55(8), 2026 (2008)
103. Prefasi, E., et al.: IEEE J. Solid-State Circuits 44(12), 3344 (2009)
Chapter 5
Interconnects, Transmitters, and Receivers
Bernd Hoefflinger
Abstract Interconnects on-chip between transistors and between functions like

processors and memories, between chips on carriers or in stacks, and the commu-
nication with the outside world have become a highly complex performance,
reliability, cost, and energy challenge.
Twelve layers of metal interconnects, produced by lithography, require, includ-
ing the contact vias, 24 mask and process cycles on top of the process front-end.
The resulting lines are associated with resistance, capacitance and inductance
parasitics as well as with ageing due to high current densities.
Large savings in wiring lengths are achieved with 3D integration: transistor
stacking, chip stacking and TSV’s, a direction, which has exploded since 2005
because of many other benefits and, at the same time, with sensitive reliability and
cost issues. On top of this or as an alternative, non-contact interconnects are
possible with capacitive or inductive coupling. Inductive in particular has proven
to be attractive because its transmission range is large enough for communication in
chip stacks and yet not too large to cause interference.
Optical transmitters based on integrated III-V compound-semiconductor lasers
and THz power amplifiers compete with ascending low-cost, parallel-wire
transmitters based on BiCMOS technologies. Parallel mm-wave and THz trans-
ceiver arrays enable mm-wave radar for traffic safety and THz computed-
tomography.
In spite of all these technology advances, the power efficiency of data commu-
nication will only improve 100 in a decade. New compression and architectural
techniques are in high demand.
B. Hoefflinger (*)

132 B. Hoefflinger
5.1 On-chip Interconnects
The interconnection by wires of millions and billions of devices, mostly transistors, on

large-scale integrated circuits (LSICs) has evolved into the most serious obstacle to the
performance increases expected of each new chip generation. The best assessment of
the problem is the ITRS (International Technology Roadmap for Semiconductors) [1].
Regarding on-chip wiring requirements, we extract the data shown in Table 5.1.
The aggressive scaling of the first-level metal (M1) and the incredible wire
lengths of the first six layers of a total of 12 layers, rising to 14 in 2020, present a
massive resistive and capacitive load, which reduces speed and increases energy
and cost significantly. Before we look at this load more closely, we point out the
steps that promise to alleviate this burden:
– Increase the transistor/device efficiency, i.e., reduce the number of transistors/
devices per function (Sect. 3.6).
– Stack transistors in 3D on top of each other and use selective growth/deposition
of Si/alloys/metal for local interconnects over distances of several nanometers as
opposed to lateral wires/contacts, which are more than an order-of-magnitude
longer and expensive (Sect. 3.7).
– Use TSVs (through-Si-vias) for 3D interconnects instead of 2D global wires to
the chip perimeter.
The latter has finally received massive attention, but its effect on the kilometers
projected in Table 5.1 has not yet been assessed more quantitatively. We notice that
the global wiring pitch is not scaled at a level close to 1 mm, and there seem to be
other accepted limits per chip:
Chip area <10 cm2

Pins on the chip perimeter Total <3,000
Signals <1,000
Power <2,000
The standard wire material is thin-film copper, whose characteristic parameters

are listed in Table 5.2. The Cu resistivity in Table 5.2 includes the effects of grain
boundaries and surface roughness. It clearly shows the limits of conventional metal
interconnects and the motives for finding a replacement. We also show how large
Table 5.1 Roadmap for on-chip interconnects. First-level metal (M1), semi-global and global
wires (From [1])
Year 2010 2015 2020
Node 32 18 10
M1 half pitch [nm] 45 21 12
M1 pitch [nm] 90 42 24
Length/cm2 [km/cm2] for M1 + 5 metal layers, signals only 2.2 4.7 8.4
Semi-global wire pitch [nm] 180 84 48
Global wire pitch [nm] 856 856 856
5 Interconnects, Transmitters, and Receivers 133
Table 5.2 Roadmap for Cu first-level metal interconnect and SiO2 dielectric
2010 2015 2020 Comments
M1 pitch [nm] 90 42 24 ITRS 2009
Aspect ratio 1.8 1.9 2.0 ITRS 2009
Cu resistivity [mO cm] 4.08 6.61 9.8 ITRS 2009
Resistance/mm [O mm–1] 11 77 340
Capacitance/mm [fF mm–1] 0.19 0.18 0.16 ITRS 2009
Transistor intrinsic delay tS [ps] 1.6 1.2 0.8 Fig. 3.10
M1 length (mm) for combined delay 2tS for W/L ¼ 1 0.6 0.4 0.2
W/L ¼ 10 6.0 4.0 2.0
the speed penalty is for Cu wires driven by transistors of the same generation and
for two channel width-to-length ratios W/L. At the 2020 node, the minimum
transistor with an intrinsic delay of 0.8 ps and with W ¼ L ¼ 10 nm, driving an
M1 wire 200 nm long (a distance of two simple gates), would mean a total delay
from transistor input to the end of the wire of 1.6 ps. Obviously, a transistor with
W ¼ 100 nm would provide ten times the current and it could drive a wire 2 mm
long for a total delay of 1.6 ps. These delays are a part of the information in Fig. 5.1,
which shows the expected delays for different interconnect lengths.
Figure 5.1 presents a summary of data on the delay times associated with various
types of interconnects as a function of their length. We have studied Cu lines driven by
NMOS transistors, and their characteristics are the superposition of two lines in our
log–log diagram: The intrinsic delay of the 10 nm transistor from Sect. 3.4 (Fig. 3.10),
and the intrinsic delay of a Cu wire with the parameters of Table 5.2, which is charged
by this transistor with the maximum current of 190 mA mm–1 from Fig. 3.9 (center).
The intersection of these two lines can be considered as the characteristic line length,
where the total delay is twice the intrinsic delay of the transistor. Proper design of the
CMOS driver would optimize (reduce) the total delay for a given wire length and load
capacitance at its end: Choose a complementary pair as the driver with the input pair a
minimum pair with width Wref and output capacitance Cref. Express the total capaci-
tance of wire and load in multiples of Cref as NCref. Choose the width of the driver pair
as N1/2 W.We choose this design rule here in order to show that standard wiring with
CMOS drivers can be tailored to provide a relative minimum of delay.
Figure 5.1 also shows reference characteristics for nanocarbon interconnects
such as carbon nanotubes (CNTs) or graphene lines [2] in different transport modes.
These fascinating structures have aroused tremendous research interest because of
some outstanding fundamental parameters, particularly those for electron transport:
Mobility 20,000 cm2 V–1 s–1, 20 Si

Peak velocity 8 107 cm s–1
Diffusion constant 200 cm2 s–1
Ballistic transport length 3 mm
In his 2010 review paper [2], Jim Meindl, a leader in interconnect research,
presented the nanocarbon delay times introduced in Fig. 5.1 with the parameters
listed above. All nanocarbon delay lines are intrinsic delay times without any
consideration of how to drive these lines.
134 B. Hoefflinger
Logic-Gate Pitches (5 ¥ M1 Pitch = 100 nm / 2020)

1 10 102 103
106 1 µs
105 100 ns
104 10 ns
103 on
1 ns
usi
iff
Delay (ps)
D
C ire
Delay
102 CuW 100 ps
= 1+
L
r. W/
ST
CM
O ire
10 20 uW 10 ps
e s 20 0 +C
W av =1
L
pin W/
CS ST
r.
O
CM
1 20 1 ps
20
ic
ist
10–1 all
CB ift 10–1 ps
Dr
C
10–2 –1 10–2 ps
10 1 10 102
Interconnect Length (mm)
Fig. 5.1 Delay times as a function of interconnect length
The intrinsic delay time for the drift mode at 8 107 cm s–1 is obviously the
shortest, rising with an increase in L2 to cross the CMOS transistor-driven Cu line at
1 mm. The diffusion mode is an order-of-magnitude slower than the drift mode and
also has an L2 behavior. It is slower than the transistor driver beyond 300 nm. The
ballistic mode starts out with a delay proportional to line length L and assumes an L2
dependence at a few micrometers, when phonon scattering sets in. It is slower than a
proper CMOS driver beyond 1 mm. For the spin wave, we have entered the
characteristic of a high-speed wave at 107 cm s–1. It has the longest delay in this
comparative study and is slower than any Cu line driven by a CMOS transistor. The
conclusion from this comparison in Fig. 5.1 is that delay times at the lower end are
limited by intrinsic transistor speed at about 1 ps, and that for any interconnect
length, the best and the working solution is the Cu line driven by CMOS transistors.
The line capacitances and the line delays in Table 5.2 are those for a SiO2
dielectric with a relative dielectric constant k ¼ 4. Research on low-k dielectrics for
insulators between interconnects is an ongoing effort with some partial solutions.
It is interesting to note that the more radical solution, air insulation, is returning
40 years after its earlier appearance in beam-lead technology at Bell Labs [3]. Its
renaissance is assisted by the large-scale use of air gaps in MEMS and NEMS
(micro- and nano-electro-mechanical systems). Many examples appear in Chap. 14.
As we move towards 2020, MEMS-type air bridges of 100 mm length are becoming
long-distance interconnects, and gaining a factor of 4 in the reduction of wiring
capacitance means a practical quantum step forward [4].
5.2 On-chip and Chip–Chip Communication
In this section, we consider various techniques to send and receive digital signals
across a chip or between chips in an assembly by amplifier circuitry. In a way, we
have already used this active-circuit technique in the previous section, when we
drove the Cu lines by a CMOS inverter or inverter pair, and we saw the benefits or
the necessity already for load capacitances on the order of femtofarads (fF).
However, on ICs, we encounter loads 1,000 times larger, particularly between
processor and memory elements, and another order-of-magnitude higher (10 pF)
if we have to send or receive signals between chips assembled in a multichip
package or on a chip carrier (CC) or printed-circuit board (PCB). For a data rate
of 1 Gb/s and a voltage swing of 1 V between ONE and ZERO on a load of 10 pF,
we need a power of 10 mW, and this power efficiency [mW/(Gb/s)] has become the
figure-of-merit (FOM) for digital chip communication. It is, in other terms, the
communication energy per bit: energy/bit ¼ CV2.
Remarkable progress has been made regarding this energy, predominantly by
reducing the voltage to the typical level of 1 V in 2010 and by scaling the circuit
dimensions, and here particularly the area of the contact islands, the pads on which
electrical contacts are made to the outside of the chip. The state-of-the-art in 2010 is
highlighted in Table 5.3.
Table 5.3 Speed and power efficiency of chip–chip transceivers

Speed Power efficiency Refs.
[Gb/s per pin] [mW/(Gb/s) per pin]
47-pin chip-to-chip, 45 nm node 10.0 1.4 8.1 in [5]
Clock-forwarded receiver, 65 nm 7.4 0.9 8.2 in [5]
Complete transceiver, 65 nm node 12.5 1.0 20.5 in [5]
Reconfigurable transceiver, 45 nm 25.0 3.8 20.7 in [5]
Inductive-coupling
synchronous 1.0 0.4 [6]
asynchronous 11.0 1.4 [6]
4-channel optical transmitter 12.5 [7]
Optical Ge receiver 10.0 1.5 20.1 in [5]
24-channel 850 nm optical transceivers 15 6.0 [8]
136 B. Hoefflinger
We can summarize the achieved results as
Power efficiency in 2010 for chip–chip transceivers: 1 mW/(Gb/s) per pin

¼ 1 pJ/bit per pin
For a typical voltage of 1 V, this means an effective capacitance of 1 pF, typical of

fine-pitch mbumps with close alignment tolerances. The table also shows the 2010
status of contactless inductive transceivers [6]. Here, thin-film inductive spiral loops
with presently 30 mm diameter serve as 1:1 transformer coils for vertical signal
transmission with near-field ranges about equal to their diameter. This is a very
promising technique for signal transmission in 3D chip stacks. It is one avenue towards
further progress in the power efficiency of chip–chip communication. Fig. 5.2 contains
a projection starting from the 2010 data, and the long-term history is plotted in Fig. 5.3.
From 2010 to 2020, another 100-fold improvement is feasible. The contribution to
the power efficiency from reducing the voltage swing from 1 V to an outer limit of
300 mV would be a factor of 10, but this will require extensive use of LVDS (low-
voltage differential signaling). The other tenfold improvement from reducing the effec-
tive capacitance implies further scaling and creative use of the third dimension, stacking
devices and chips in order to avoid the penalties associated with lateral wiring distances.
A favorite in bridging lateral distances and reducing crosstalk has been optical
communication across air or through fibers or SiO2 waveguides. An example for the
state-of-the-art in Ge detectors is listed in Table 5.2 with 10 Gb/s and a power
efficiency of 1.5 mW/(Gb/s). Improvements are expected by operating the photo-
detector in the avalanche mode. On the transmitter side, 4-channel and 24-channel
laser-diode transmitters have been announced with rates up to 15 Gb/s per diode
channel [7,8]. The expensive optical interconnect will not readily replace the other
alternatives, and its added cost will be carried only in specific application areas,
100
20 10
20 20
20.7
20.5
10
Speed (Gb/s)
8.2 8.1
fJ n)
10 ro
eu
N
(1
1
pJ
1
0.1
10–2 10–1 1 10 102
Power (mW)
Fig. 5.2 The speed and power of digital chip–chip communication. The references (8.1 etc.) refer
to papers at the ISSCC 2010 [5]
102 100
10 10
Power Efficiency
Energy/b
(pJ/b)
Gb/s
mW
1 1
10–1 10–1
10–2 10–2 (Neuron)

2000 2010 2020
Year
Fig. 5.3 The roadmap of power efficiency for chip–chip communication
such as large servers and supercomputers, where millions of such channels with
lengths from centimeters to meters will be needed. A special example would be
mask-less, direct-write nanolithography, where systems of more than 100,000
parallel beams would be driven at rates >100 terabit s–1 (Chap. 8).
The 2020 (short-range) communication target of 10 fJ/bit would take us to the
typical energy level of synapses of a neuron (Chap. 18), where we find power levels
of 1 pW in signals about 10 ms apart. If we revisit Table 5.1 to find a 2015 wire length
of 5 km/cm–2 for six wiring levels with a total height of about 500 nm, we obtain a
volume wiring density of 100,000 km cm3, a wire length that would stretchthree
times around the earth. We conclude that, in communications energy and wiring
density, we are really approaching the characteristics of the human brain.
5.3 Wireless Transceivers
A highly important field for the connection of chips to the outside world, but also for
the contact-less communication between chips, is the wireless transmission and
reception of information. We give an overview, again with an emphasis on figures-
of-merit and on how the trends in the world of digital mobile communication
(Chaps.12 and 13) impact the required chip performance. We start with an overview
of typical power requirements on wireless transmitters and receivers listed in Table 5.4.
We have also listed the radio frequency (RF) power levels in dBm units, commonly
quoted by RF professionals. This means power on a log scale relative to 1 mW:

PowerðWÞ
PowerðdBmÞ ¼ 10 log10
1mW
138 B. Hoefflinger
Table 5.4 RF power levels of wireless digital systems

33 dBm 2W Maximum output UMTS/3 G mobile phone (power class 1)
30 dBm 1W Maximum output GSM 800/1900 mobile phone
20 dBm 100 mW Typical transmission power of a wireless router
15 dBm 32 mW Typical WiFi transmission power of a laptop
4 dBm 2.5 mW Bluetooth transmission power, class 2: 10 m
10 dBm 100 mW Typical maximum receiver signal of a WLAN
70 dBm 100 pW Typical wireless receiver in an 802.11 network
120 dBm 1 fW Receiver signal from GPS satellite
Table 5.5 Wireless transmitters and receivers. On-chip performance 2010 [5]
Frequency Data rate Microwave PAEa Energy per bit Ref.b
[GHz] power [dBm] [%] [nJ/bit]
UWB impulse 3–5 1 Mb/s 80. . .–92 0.92 transmit 11.9
radio 5.3 receive
Receiver 2.4 0.5 Mb/s 82 0.83 11.6
SiGeBiCMOS 60 13 4 11.3
16-ch. transmitter
Receiver 65 nm 60 21 11.4
1.1 V Transceiver 56 11 Gb/s 1 mW Tx 3 0.06/14 mm 23.1
SiGeBiCMOS 160 <5 1 23.2
Transceiver
1 V 65 nm Ampl. 60 17 11.7 23.6
1.2 V 90 nm Amp. 60 20 14.2 23.7
60.5 18 4.9 23.8
1.2 V Amp. 85 6.6 3 23.3
3.3VSiGeBiCMOS 650 54 Rx 23.9
a
Power added efficiency
b
References refer to papers at the ISSCC 2010 [5]
We choose some recent 60 GHz transceivers and the 2–5-GHz range to highlight
the state-of-the-art performance in Table 5.5.
The reported results span a wide range of carrier frequencies with the emphasis
on the 60 GHz band, which is considered for short-range radar and for high-data-
rate (10 Gb/s) indoor wireless communication. The performance levels have been
achieved with typically 1 V and 90–40 nm CMOS technologies with power-added
efficiencies (PAEs), i.e., ratios of RF to DC power, of typically 10%. This is very
promising for media-rich wireless, where these data rates are necessary.
Overall, we have to realize that bandwidth in broadband wireless is becoming
limited (Fig. 5.4) and more precious so that bit-rate compression and coding require
sustained R&D.
The data on the 650 GHz experiments also show that solutions for terahertz
imaging are available with present 130 nm high-speed SiGe BiCMOS technologies,
making up an impressive high-performance technology arsenal.
Nevertheless, the mobile-broadband volume has increased at such dramatic rates,
Fig. 5.5, in particular for wireless video, that quantum-jump improvements in video
data rates are needed to support the future mobile companions envisioned in Chap. 13,
Fig. 5.4 Limited spectrum for mobile broadband [9] (# 2010 IEEE)
Fig. 5.5 Projected mobile-broadband data volume per month [9] (# 2010 IEEE)
140 B. Hoefflinger
and new, highly efficient video cameras with intelligent focal-plane frame processing
(Chap.15) have to be developed.
References
1. www.ITRS.net/reports.html. 2009 Edition

2. Meindl, J., Naeemi, A., Bakir, M., Murali, R.: Nanoelectronics in retrospect, prospect and
principle. In: IEEE ISSCC (International Solid-State Circuits Conference) 2010, Digest of
Technical Papers, San Francisco, 7–11 Feb 2010, pp. 31–35. doi: 10.1109/ISSCC.2010.
5434062
3. Lepselter, M.P., Waggener, H.A., MacDonald, R.W., Davis, R.E.: Beam-lead devices and
integrated circuits. Proc. IEEE 53, 405 (1965). doi:10.1109/PROC.1965.3772
4. Johnson, R.C.: Air gaps move up from chip- to board-level. www.eetasia.com/
ART_8800616596_480100_NT_3f74ff49.HTM. Accessed 19 Aug 2010. See also: Interconnect
Focus Center at Georgia Tech (Prof. Paul Kohl) and Steve Hillenius, executive VP at the
Semiconductor Research Corp. (SRC)
5. IEEE ISSCC (International Solid-State Circuits Conference). 2010, Digest of Technical Papers
6. Ishikuro, H.: Contactless interfaces in 3D integration. In: IEEE ISSCC (International Solid-
State Circuits Conference) 2010, Forum 1: Silicon 3D Integration – Technology and Systems,
San Francisco, 8–12 Feb 2010
7. Johnson, R.C.: Intel: silicon photonics to replace copper. Electronic Engineering Times, Asia,
29 July 2010
8. Schow, C., Doany, F., Kash, J.: Get on the optical bus. IEEE Spectr 47(9), 32 (2010)
9. Lazarus, M.: The great spectrum famine. IEEE Spectr. 47(10), 26 (2010)
Chapter 6
Requirements and Markets for Nanoelectronics
Bernd Hoefflinger
Abstract The semiconductor market grew 2010 by 70Bio.$ against 2009, more
than in the previous 9 years taken together, and the semiconductor industry
launched the biggest investment program in its history with 100Bio.$ over a 2-
year period. This was the overture to a decade with great potential and great
challenges. We look at the market segments and the required electronic functions,
and we highlight four product and service areas:
• Approaching 6 Billion mobile-phone subscribers
• Access to education for any child
• One Carebot (personal robot) per family
• Efficient and safe personal mobility.
At the level of over four billion active mobile phones 2010, it is clear that mobile
electronic companions have become the drivers of nanoelectronic innovations with
growth only limited by the creation and support of new, attractive features and
services. Energy, bandwidth, size and weight requirements of these consumer
products provide the largest pressure for System-on-Chip (SoC) architectures.
Other exemplary new products are selected for their significance, some for their
lengthy path into the market. Health care is such an example: The non-invasive
glucose sensor and the portable ECG recorder” with automatic, neuroprocessor-
driven event detection in the size of a quarter $ would serve hundreds of millions of
people. Nanoelectronics for self-guided health is an area of public policy in view of
the cost of “a posteriori” medical care.
Access to information and education for any child/student will be provided by 1$
tablets where service contracts and the spin-offs from surfing and cloud-computing
will generate the revenue.
B. Hoefflinger (*)

142 B. Hoefflinger
Personal robots, coined by the ageing Japanese nation as the key product after
the PC and ridiculed by others, will arrive as carebots for education, entertainment,
rehabilitation, and home-service, accepted as a large-scale need by 2020 in most
developed countries including China.
Accident prevention systems on rail and road already would make millions of units
per year, if required on all trucks and busses, and they would save many lives. For
electric bikes, scooters and cars, there is no limit to more intelligent control and energy
efficiency. Effective individual mobility, compatible with the environment, is another
matter of global competition, of public well-being and of a related public policy.
6.1 Progression of the Semiconductor Market
The semiconductor market has seen a history of growth and cycles like no other
industry. It reached the level of $1 billion (109) in 1970, and it grew by 15% per annum
for the next 30 years, much faster than any other industry, to the level of $200 billion in
2000. This corresponded to about 100 billion chips per annum at an average selling
price (ASP) of about $2. More than 80% of this market has been integrated circuits,
and the average number of transistors on these circuits (chips) increased from about
100 in 1970 to about 100 million in 2000. This growth helped the information and
communication industry (ICT) to grow from about $350 billion in 1970 to $3 trillion in
2000 at a rate of 7.5%/a (percent per annum), about half the rate of chips and yet about
twice that of the world economy as a whole. This development is illustrated in Fig. 6.1
together with the first decade of the new millennium and an outlook towards 2020.
We see that between 1980 and 2000
– Chips quadrupled their share of the ICT industry to 7%
– The ICT industry doubled its share of the world economy to 10%.
The first decade of the new millennium has seen diminished growth on all fronts
due to the burst of the internet bubble in 2000 and the financial crisis of 2008/2009, but
also due to the slow start of third-generation mobile-phone products and services.
For the years 2010 and 2011, the semiconductor industry expects to grow by $100
billion, which is more than the growth in the past 9 years, to about $320 billion. As
Fig. 6.1 shows, a conservative estimate will carry the semiconductor industry to $450
billion in 2020, and several sources predict an optimistic growth by 7%/a to $600 billion.
At this level of chip revenue, the revenue of the ICT industry would reach $7 trillion in
2020, so that it would double its share of the world economy or, in other words, it would
continue to support and to assure the growth of the world economy by at least 2.5%/a.
What are the market drivers among chips, and which are the key requirements on
these chips? The fundamentals, in simple terms, have persisted over four decades,
and they have settled recently at these parameters:
1. The cost per transistor drops by 50%/a.
2. The number of transistors per chip increases by 50%/a, which allows new
functions and new applications.
6 Requirements and Markets for Nanoelectronics 143
Fig. 6.1 The world economy (gross product of goods and services), the ICT revenue, and the
semiconductor industry revenue (CHIPS) from 1980 to 2020 [1–5]
3. The number of chips increases by 15%/a.

4. Total available market (TAM) increases by 7%/a.
From a general functionality point of view, we can summarize the expectations/
requirements for the decade 2010–2020 as:
– Memories: Increase the number of bits per chip-scale package 100–300-fold.
– Logic: Increase the operations per second per watt 300-fold.
These broad targets can be seen as general drivers for the advancement of the
chip industry. In the following, we take a closer look at the market segments, the
special functionalities on the chips, and on their relationships, as well as their
significance for progress.
144 B. Hoefflinger
We notice in Fig. 6.2 that computers and communications make up about 65% of
the chip market, with a trend towards communications. These two segments
continue to drive the chip market. Consumer chips have changed by just 1% to
19%, with, however, the introduction of many new products within their segment,
as we will see shortly (Fig. 6.3). Automotive chips have increased their share by
Fig. 6.2 Segments of the semiconductor market in 1995 and 2009 [5, 6] (#WSTS and EE Times)
Fig. 6.3 Revenue details for 2009 within the major market segments [5, 6] (#WSTS and EE
Times)
Table 6.1 Chip revenue ($billion) and 14-year growth by market segment
1995 2009 Change (%)
Computing 70 99 +41
Communication 21 45 +114
Consumer 25 43 +72
Med./Industrial 14 20 +40
Automotive 7 18 +157
Military 3 2 33
Total 140 227 +62
Table 6.2 Market segments (vertical) and special functions (horizontal) of nanochips
Processor Memory Analog/digital Transmitter/ Sensor/actuator
(Chaps. 10, (Chaps. 11, 18) (Chap. 4) receiver (Chaps. 14,
13, 16) (Chap. 5) 15, 17)
Computing xxx xxx xx
Communication xx xx xxx xxx
Consumer x xxx xxx xxx xx
Med./industrial xx xxx xxx xx xxx
Automotive xx x xxx x xxx
Military xxx xxx xxx xxx xxx
60% and their revenue by 157% in that 14-year period. Table 6.1 shows that this
period saw an overall growth in revenue of 62%. Besides automotive, consumer and
communications sectors came in with growth beyond the chip average of +62%.
Within the major market segments, several product areas stand out, which have
become the drivers, as we can see in Fig. 6.3.
Chips for notebooks/netbooks and for mobile phones have provided most of the
revenue growth and will continue to do so. In the consumer segment, all areas have
seen and continue to see remarkable growth. The actual progress will depend on the
continuing, rapid improvement of the chip functionalities, which are listed in
Table 6.2 in a matrix against the market segments. We have listed the chapters of
this book in which the special functions are treated in detail, and we have
highlighted their importance for the different markets. Here we continue by
pointing out the potential and challenges for nanochips in the various markets.
6.2 Chips for Computers
The potential and challenges for nanochips in the computer segment are
summarized in Table 6.3. Computers continue to mark the leading edge in
processing power provided by the most complex chips, and Chap. 10 provides
insight into the performance and progression of mainframe processors. Their
ultimate assembly into supercomputers marks the high end of computing power,
146 B. Hoefflinger
Table 6.3 Challenges for nanochips in the computer segment

Processors Memory Analog/digital Transmit/receive Sense/actuate
Computers Throughput Capacity Bandwidth Video
Speed Bandwidth Noise 3D Display
Energy architecture Energy
Fig. 6.4 Outlook in 2009 for the TOP 500 worldwide computing power measured in FLOPS
(floating-point operations per second) [7] (#TOP 500)
continuously monitored by TOP500 [7], the agency keeping track of the peak in
worldwide computing power, as illustrated in Fig. 6.4.
We see that the world’s leading supercomputer achieved 20 teraFLOPS in 2009
and that the industry believes in a trend line of 1,000/decade, which predicts a
performance of the number 1 supercomputer in 2020 of 20 peta(1015)FLOPS,
equivalent to the total worldwide computing power in 2010. This performance
will either be demonstrated in a brute-force approach of merging the processor
power then available into such a powerful – and power-hungry – system, or by the
DARPA-funded Program on Ubiquitous High-Performance Computing [8]. The
required electrical power for the worldwide processor farms in 2020 will indeed be
the number 1 problem of future high-performance computing, and it will be touched
upon in Chap. 20.
Figure 6.3 tells us that the high-end computer chips, those going into servers and
supercomputers, make up about one third of the total revenue from computer chips.
Chips for desktops make up another, declining third, and chips for portable
notebooks and netbooks represent the remaining, rising third. This sector is the
promising, controversial, and embattled field of growth, because it has an army
approaching from the mobile-phone camp from the communication segment with
fourth-generation (4 G) smart phones and personal assistants/companions. This

convergence is the most dynamic phenomenon of product development at the
start of the decade 2010–2020. These mobile products have the dual requirement
of ultra-low-power and high-performance data and video processing. We leave the
video part to the sections on communication and consumer chips and just highlight
the competition within the data-processing community:
– CISC versus RISC
– Cloud computing
– Processor-memory bandwidth
The CISC (complex-instruction-set computing) league, led by the Wintel alliance
of Microsoft Windows and Intel, naturally stuck with high-power cores and broad
communication between processing and memory elements, has faced a steadily grow-
ing, highly diversified league of RISC (reduced-instruction-set computing) companies,
mostly licensing the architecture and the operating system from ARM (Advanced
RISC Machines), today the world’s most successful licensor of chip technology (see
Chap. 21). In terms of shipments in 2010, there are 300 million PCs/a against 700
million mobile phones/a, and the newcomers, tablet PCs. These are the new hardware/
software/wireless-mobile medium for e-books and e-newspapers/magazines.
The progression of these wireless mobile media can be followed by watching
ventures dedicated to provide computing and information access to children and
students. This movement was pioneered by Professor Negroponte’s, UN-endorsed
project OLPC – One-Laptop-per-Child. It provided exemplary pressure on bringing
the cost and power consumption of a PC down and solving the battery problem in
remote locations (initially with a crankshaft dynamo). While the price for such an
educational PC has been brought down to <$100, the wireless-mobility issue can
now be served by third- and fourth-generation mobile-phone systems. Together
with the user interface provided by the new generation of e-tablets/pads, the post-
Gutenberg galaxy is becoming a reality, where OLPC takes on the new meaning of
Open Libraries for Practically all Children.
We can see that educational e-books are spreading faster in newly developed
countries than in the Old World, and that this global potential offers sufficient
growth for computer chips, if their cost, performance, and power consumption
make the big steps indicated above.
The trend towards mobile thin clients helps the promotion of cloud computing,
the shift of complex-software-dependent and computationally intense tasks to
remote servers in the internet. This system- and services-driven strategy with
sizable investments in public and private infrastructure has evolved into
government–industry
10: Programs for cloud computing,
shifting the PC emphasis away from classical desktop/laptop PCs.

Advances in computing are more and more limited by cost and energy,
mostly expended for the high-speed and highly parallel communication between
148 B. Hoefflinger
processors and memories. This communication expense does not scale effectively
with new technology generations, mostly because of the circuit topology and the
interference between channels. Therefore, the pressure is rising on the classical
monopoly of the von Neumann architecture of processor, data- and instruction-
memory. This architecture will continue to be the best for mathematical tasks.
However, for
– Databases
– Interconnection networks
– Search engines,
– Video and graphics processing
alternative architectures have advanced, such as associative processing, helped
by scaling of transistors and memory cells. This is evident in a review of associative
memories, also called content-addressable memories (CAMs) [9]. The more radical
innovation strategy is a critical-size R&D effort on silicon brains:
10: Programs Fast Analog Computing with Emergent Transient States, FACETS (Europe),
and Systems of Neuromorphic Adaptive Plastic Scalable Electronics, SyNAPSE (DARPA)
([10, 11] and Chap. 18).
These programs mark a renaissance of cellular neural networks (CNNs), origi-

nally conceived in the 1960s and more broadly investigated in the 1980s and in
some European centers continuously (Chap. 16). Now, it seems that there is a
critical interest because of
– Roadblocks for von Neumann and
– Nanoscale devices and 3D integration necessary and now available for complex
brain-like structures.
In any event, the results of these more radical programs will not only influence the
computer segment of chips, but also any of the other segments. Of these,
communications, with billions of subscribers, will certainly be the largest.
6.3 Communication
The potential and challenges for nanochips in the communication segment are
summarized in Table 6.4. The communication segment has been driven by the
digital mobile phone, first introduced in 1985 in Europe with the GSM standard and
in Japan with the PCS standard. Global standardization evolved with code-division
multiple access (CDMA) and then about 2000 with UMTS, the Universal Mobile
Telecommunication Standard, also called 3G (Third generation). This has brought
not only still images but also video to the mobile phone as well as e-mail and
internet so that all mobile phones manufactured today have these features of smart
phones. The demand on
Table 6.4 Challenges for nanochips in the communication segment

Processors Memory Analog/ Transmit/ Sense/actuate
digital receive
Communication Energy Capacity Resolution Bandwidth Video
Signal Nonvolatile Speed Energy 3D display
processing
Architecture Bandwidth Energy
– Transmission bit rate

– Video compression and coding
– Camera and display resolution
– Tactile interface
has put and continues to put tremendous pressure on the rate of progress regarding
– Processing power efficiency (GOPS mW–1)
– Transceiver bandwidth (GHz) and power efficiency (RF power/DC power)
– Overall power management (sleep vs active)
The processing-power efficiency, owing to the inclusion of video capability, had
to be improved 300-fold between 2000 and 2008. This achievement led to the
projection of another factor of 300 within the present decade, not only to bring high-
definition TV (HDTV) resolution (2 k 1 k) wireless anywhere, anytime to the
user but real movie resolution of 4 k 2 k pixels at rates of 25 frames per second. It
is the live sports-event experience with personal zoom and possibly 3D that could
eventually be demonstrated with this extreme technology specification.
The wireless, fiber, and satellite communication infrastructure will have to grow
50-fold to support this growth in data traffic, and also to meet the demand from the
tablet computers mentioned in the previous section. Obviously, e-tablets will also
be introduced by the leaders in the mobile-phone market as their up-scale product in
this continuing convergence of PC and smart phone. The third party in this
convergence is the TV, accounted for within the consumer segment, but really
also a communication and computer product, considering its capabilities listed in
the next section.
6.4 Consumer Segment
The potential and challenges for nanochips in the consumer segment are
summarized in Table 6.5. The TV has advanced to a
– Server
– Broadband transceiver
– 3D movie theater
150 B. Hoefflinger
Table 6.5 Challenges for nanochips in the consumer segment

Consumer Energy Nonvolatile Cost Wireless Video
Cost Cost Audio
Architecture Capacity 3D display
These attributes make the TV a product as challenging as the computing and

communication products discussed above. The requirements on the chips are as
high as for PCs and smart phones, because, in any given technology generation, its
video resolution will be 16–32 times higher, and it has to run virtually continuous-
time, which tells us that we have an energy crisis unless we achieve the same levels
of power efficiency and unless we can handle display drivers with the complexity of
today’s cinema projectors at the energy and cost level of a consumer commodity.
The silicon compact disk, a flash memory in a credit card, is about to replace
CDs and DVDs, and solid-state drives (SSDs), having achieved 32GB capacity in
2009 and now heading towards terabytes (Ch. 11), will replace the disk drives in
MP3 players.
HD graphics and video, heading towards 3D, represent a sizable and highly
competitive market, justifying aggressive development, which will, however, fail the
energy test unless quantum jumps in the efficiency of image processing are achieved. It
is here that new architectures such as neuromorphic networks (Chaps. 16 and 18) have
to be explored on a 10 level-of-effort.
These processing architectures will also help digital cameras to new heights of
stand-alone video and image focal-plane processing and coding. In fact, with the
global push towards 3D integration of device and chip layers, the retinomorphic, 3D
integration of photoreceptor planes and processing planes (Fig. 3.33) will be
introduced on a larger scale. This will be discussed in Chap. 15.
6.5 The Professional Segments of the Chip Market
The three segments medical/industrial, automotive, and military have a number of

requirements in common that have forced the chip industry over decades to meet
high standards. These mean significant levels of process R&D and rigorous quality-
management systems (QMSs). Generally, there are higher operating requirements
for chips in professional applications:
– Extended temperature range
– Electromagnetic protection and compatibility (EMC)
– Harsh environments
– Controlled degradation
– Guaranteed reliability levels (FIT).
A guarantee of 100 FIT (one failure in 107 h), a typical requirement for
an application-specific chip (ASIC) with millions of transistors tells us that the
assurance of the quality and reliability of chips is a major task.
A system of standards and certifications has evolved, which we summarize here
to highlight the discipline in the chip industry that has been responsible for the
unparalleled growth of this industry:
– MIL-STD-883: This has been the global procedure for the qualification of chips
for 40 years.
– QML (Qualified Manufacturing Lines): Total plant qualification established and
performed by RADC (Rome Air Development Center) since 1981 associated
with the development of the chip factories manufacturing wafers and chips under
contract with outside customers. The beginning of the foundry age.
– ISO 90001 (1988) and ISO 9000:2000: International standards for total quality
management of manufacturing enterprises.
– CECC-EN 1001–14: Certification of integrated-circuit manufacturing, specific
for each technology generation.
– QS 9000: Qualification of chip suppliers by automotive customers (adopted by
many professional industrial customers).
These qualifications have to be performed at least for each technology genera-
tion and mostly on a per-product basis, and the semiconductor company in turn has
to certify all suppliers and service providers. Over the past 20 years, such quality
chains have been established generally for most professional businesses worldwide.
It is still a singular phenomenon that the chip industry has had this culture for
40 years, and, in the case of the rapid deployment of new technology nodes, has to
do this every 18 months.
Naturally, those chip makers in the professional segments need a consistent,
advanced R&D commitment. However, there is a large area of mutual interest
between them and those in the volume segments.
We have seen that the three major segments of the chip market, making up more
than 80% of the total market, have to make significant investments in video
processing. On one hand, they can benefit from R&D conducted by the three
remaining, professional segments of the market, where the system R&D level is
naturally higher. On the other hand, the volume introduction of new chips in the big
segments helps the professionals to bring the cost down for these innovations.
6.6 Medical/Industrial Segment
The potential and challenges for nanochips in the medical/industrial segment are
summarized in Table 6.6. Industrial applications of microchips are to a large degree
in electronic control. These often safety-critical tasks means that IC solutions had
and have to go through elaborate approval procedures, for example, performed by
UL (Underwriters Laboratories) or to obtain European certification (the CE mark).
152 B. Hoefflinger
Table 6.6 Challenges for nanochips in the medical/industrial segment

Medical Performance Non-volatile Speed Bandwidth Multi-sense
Industrial Architecture Precision Body-Area Net Lab-on-Chip
Power-drive
Intelligence
Fail-safe
Fault-tolerant
Historically, this meant that the certification laboratories went slowly and orderly
from
– Mechanical to electromechanical (relays),
– To discrete-transistor boards,
– To small-scale ICs (operational amplifiers, then diode–transistor logic (DTL)
and transistor–transistor logic (TTL)),
– To gate arrays,
– To microcontrollers,
– To application-specific chips,
some of these levels taking more than 10 years. The message is that control
hardware and software are slow to be certified and have long lifetimes once
accepted. It means that new computing paradigms will take >30 years to be
accepted in industrial control. Fuzzy logic as a parallel multi-input, rule-based
control architecture, introduced in the early 1980s, has only made it into
selective applications, and cellular neural networks (CNNs), after 20 years, are
just beginning to make it into industrial control, although they have, for exam-
ple, high levels of built-in fault tolerance. Essential properties of chips in
industrial control have to provide
– High reliability under wide variation in operating conditions
– Fault-tolerant architecture
– Fail-safe performance (under degradation, go into a safe, defined state)
– FMEA (failure-mode and effect analysis)
These are just a few criteria intended to show that an aggressive scaling of
technologies to levels where transistors perform marginally and with a large
variance is not the strategy for progress in professional applications., Rather safe
and effective new architectures should be explored that reduce the number of
components and the power density in chips and packages, offering the largest
benefits in cost, energy, and system lifetime/reliability.
Among the many exemplary industrial electronic products, robots stand out as
the most complex systems and with the promise of being the next big thing after the
PC, as predicted by Japanese experts in 2001 [12].
The personal robot, designated by the aging Japanese nation as the key product
after the PC and ridiculed by others, has arrived in various shapes, from vacuum
cleaners to humanoid marching bands and tree climbers. As of 2010, 50% of the
world robotics market is already accounted for by home/service robots, and this
market is predicted to grow threefold within the next 15 years [13], as detailed in
Table 6.7.
Table 6.7 Worldwide robotics market growth 2010–2025 in $billion, after [13]
2010 2025 Growth rate
Home 12 34 2.8
Manufacturing 8 12 1.5
Medical 2 9 4.5
Public sector 2 8 4.0
Bio-industrial 1 3 3.0
Total 25 66 2.6
Robot electronics today is composed of hundreds of outdated commodity

sensors, controllers, and actuators, in industry owing to lengthy certification
routines and otherwise often as a proof concept and devoid of an economy-of-
scale. In order to really make robots adaptable, intelligent, and safe in real-life
situations, a quantum jump towards adaptive, intelligent, fault-tolerant hardware
with multisensor and multiactuator interfaces is needed, and neural architectures fit
these tasks. They will also reduce significantly size, weight, power consumption,
and cost. The new focus on silicon brains (Chaps. 16 and 18) will be key to this
development.
Personal carebots for education, entertainment, rehabilitation, and home-
services are accepted as a large-scale need by 2020 in most developed countries,
including China, not only for the elderly in aging societies but also as companions
and tutors for the young. Several exemplary projects illustrate the 10 level of
these developments:
10 Programs for carebots:

– A Robot in Every Home by 2020 (Korea)
– Japan Robotics Association
– KITECH (Korea Institute of Industrial Technology)
– Taiwan [14]
– CARE (Coordination actions for robotics in Europe)
Different directions are being pursued. Many systems are special-function tech-
nical solutions while multifunction, companion-type carebots tackle the human-
interaction and feel issue with humanoid, also called android, systems [15, 16], as
shown in Fig. 6.5.
Robot learning and the robot–human interaction are continuing fundamental
challenges [17, 18], but also interdisciplinary challenges, where nanochip
strategists and designers have to get more involved given the formidable perfor-
mance levels of functionalities that can be realized today and in the near future.
This call for interdisciplinary development work should also be heard more
clearly regarding the vast opportunities in medical electronics. Pioneering
154 B. Hoefflinger
Fig. 6.5 The android robot Geminoid F by Hiroshi Ishiguro of Osaka University and ATR in 2010
[15] (# 2010 IEEE)
achievements such as the pacemaker of 1964 [19], based on a bipolar circuit just 4
years after the invention of the integrated circuit [20] highlights the importance of
radical innovations for the formation of entirely new industries. In this case, it is the
industry for electronic implants and prosthetics. Among the many systems being
developed at present, we chose retinal implants for the restoration of vision as the
most complex system. Its implantation in the retina of the eye and its interaction
with the visual cortex was proposed in 1996. Experts in ophthalmology, physiology,
neurology, microsurgery, biomedicine, physical chemistry, and microelectronics
worked together, and after about 10 years of R&D [21], tests could be performed
with patients worldwide. An account and perspective is given in Chap. 17.
We can only pinpoint a few areas where medical electronics has significant
impact on public health and on growth within the chip market. Minimally invasive
diagnosis, therapy, and surgery benefit most from the decreasing size and the
increased performance of scaled-down chips. A key example for the state-of-the-
art is the camera pill from Given Imaging [22], which can be swallowed, and, while
traveling through the intestines, transmits images to a receiver on the body’s surface
(Fig. 6.6).
Personal, portable, non-invasive diagnostic and monitoring devices offer great
benefits to patients and a potential mass market for medical electronics. The
percentage reduction in size and weight of these devices would appear directly in
the growth rate of the available market. As an example, the Holter tape today is a
magnetic-tape recorder the size of a Walkman that monitors a patient’s cardiogram
over 24 h. It then has to be reviewed by the physician for critical events. This could
be transformed into a device barely larger than the needed skin contact, containing a
neural-network event-recorder chip, trained for the individual patient, and an event
flash memory. This would deliver directly the critical events and the time when they
Fig. 6.6 The Given Imaging

autonomous camera capsule,
which transmits wireless
images from inside the
intestines after being
swallowed, has been used in
hospitals since 2002 [22].
Source: Given Imaging
occurred, and it could directly alert the patient or his supervision of any problems.
Such a device would have a significant preventive effect and save a posteriori
suffering of patients and expenses of (public) insurers. Therefore, the development
of such a device would be in the public interest, and resources should be provided to
accelerate their introduction into practice.
The same is true for the non-invasive, continuous monitoring of blood sugar for
diabetics. About 20 million people worldwide suffer lifelong from type-I (genetic)
diabetes, and 500 million acquire it due to lifestyle and age. The consequences of
inadequate monitoring – blindness, amputation, and heart disease – are burdensome
and expensive. It is also presently a big business for the pharmaceutical industry
selling test strips, often 5 per patient per day, for intrusive blood tests, which, of
course, are tough on patients and not good enough to detect the critical events of
hypoglycemia. An ear clip with a multispectral, near-infrared spectrometer, contin-
uously trained for the patient’s characteristics, would offer continuous care-free
monitoring. Research on this technology started in the 1980s, with an estimated
total level-of-effort of < $100 million worldwide over 30 years, without and
probably against the pharmaceutical industry. In 2010 there is no such system on
the market. There are many other examples where a public research policy guided
by long-term health requirements is needed, especially because lengthy public
certification procedures limit the investment of venture capital and often are not
survived by start-ups.
The remarkable advances of chips and microsystems in miniaturization and
performance, driven by the big-three market segments, have created a new market
with significant benefits and leverage: Minimally-invasive surgical robots.
The availability of microchips for
– Millmeter-size video cameras and 3D vision
– Microsensors and microactuators
– High-performance video processing
has allowed quantum jumps in medical surgery and related fields [23].
3D medical imaging plays a key role in surgical robots, and it is driven to its
highest levels in resolution and speed in computed tomography (CT), a sector with
156 B. Hoefflinger
remarkable unit growth. It has extreme requirements for integrating large numbers
of Si tiles on the torus of the equipment to sense and process the radiation arriving
on their surfaces. It is now entering a new generation with the large-scale availabil-
ity of the 3D integration of sensing and processing chips connected by thousands to
millions of TSVs (Sect. 3.7).
6.7 Automotive Segment
The potential and challenges for nanochips in the automotive segment are
summarized in Table 6.8. The automotive segment with a revenue of about $20
billion in 2010 is the fastest growing segment in the chip industry. In spite of the
enormous pressure on cost put on the chip industry by the automotive companies,
the chip content/car has doubled each decade to $400 in 2010, and it will double
again by 2020, driven by the fundamental interest in efficient and safe personal
mobility, which has, in the meantime, become a mass issue in the newly developed
countries as well, with over 1 billion potential customers. This mobility issue and its
support by electronics and computing became the topic of the largest-ever joint
research effort in Europe:
10: PROMETHEUS: Program for European Traffic with Highest Efficiency and Unprecedented
Safety.
Fifteen car companies and over 70 research institutes started a joint, long-term
research project in 1986, the centennial of the automobile, on all aspects of informa-
tion and communication technology (ICT) to innovate road traffic [24]. Soon,
automotive suppliers and ICT companies joined so that, at the peak of the program,
about 50 companies and over 100 institutes participated in PROMETHEUS until
about 1996, when it was continued as many, more specific programs. It was probably
the largest seeding program for new automotive and traffic technology. Programs
with similar goals were set up in the US and Japan, for example,
10: IVHS: Intelligent Vehicle–Highway Society (USA),
an industry–government cooperative still in existence today. A 1993 special issue

of IEEE Micro [25] reflects on some topics of these programs.
– Navigation
Table 6.8 Challenges for nanochips in the automotive segment

Automotive Fail-safe Non-volatile Speed Bandwidth Multisense
Architecture Cost Precision Security Power-drive
Intelligence Cost Cost Display
Cost Cost
Fig. 6.7 Traffic-sign detection and classification [26] (#Springer )
– ESP (Electronic Safety Program) (Chap. 14)

– ACC (Autonomous Cruise and Anti-collision Control)
are some of the prominent results in the market. A wealth of prototypes have
seen tens of thousands of operating hours and hundreds of thousands of
kilometers and are under consideration or waiting, indicative of the lengthy
procedures necessary for certification or clarification of privacy or liability
issues. Presently, accidents with considerable fatalities and damage, caused by
trucks and buses due to driver fatigue or negligence, are rising steeply, while
fatigue-warning, lane-keeping, and collision-avoidance systems are among the
most tested functions [26] (see an example in Fig. 6.7).
However, governments are slow to make these systems mandatory, keeping
market demand low and cost high. Instead, their introduction should be mandatory
and temporarily subsidized in view of public-safety interests. Naturally, in the
introductory phase of these new systems, developed in the advanced countries,
high-level manufacturing jobs would be produced there, replacing the migration of
other manufacturing jobs to low-cost regions.
A recent area of high growth is the transition to hybrid and electric vehicles
because of the never-ending optimization of their electronic control. The intelligent
personalization of these cars to the driver and to his/her daily commute route will be
one application of the coming intelligent neuro-controllers (Chap. 16).
6.8 Military Segment
The potential and challenges for nanochips in the military segment are summarized
in Table 6.9. Military electronics, as we have seen, has been the leader in challeng-
ing specifications and the driver for fail-safe, high-quality, and reliable electronics.
This continues to benefit the other professional segments.
158 B. Hoefflinger
Table 6.9 Challenges for nanochips in the military segment

Military Performance Capacity Precision Bandwidth Multisense
Architecture Bandwidth Speed Security Power-drive
Intelligence Non-volatile Body-area networks
Secure
References
1. www.EITO.com
2. www.Gartner.com
3. www.ICInsights.com
4. www.ISuppli.com
5. www.WSTS.org
6. Rhines, W.: Comment: is IC industry poised for consolidation? EE Times Asia, 7 April 2010.
www.eetasia.com
7. www.TOP500.com
8. http://www.darpa.mil/Our_Work/I2O/Programs/Ubiquitous_High_Performance_Computing_
%28UHPC%29.aspx. Accessed March 2011
9. Pagiamtzis, K., Sheikholeslami, A.: Content-addressable memory (CAM) circuits and
architectures: a tutorial and survey. IEEE J. Solid-State Circuits 41, 712 (2006)
10. facets.kip.uni-heidelberg.de/. Accessed March 2011
11. www.darpa.mil/Our_Work/DSO/Programs/Systems_of_Neuromorphic_Adaptive_
Plastic_Scalable_Electronics_(SYNAPSE).aspx. Accessed March 2011
12. Japan Robot Association: Summary Report on Technology Strategy for Creating a “Robot
Society” in the 21st Century (2001). www.jara.jp/e/dl/report0105.pdf. Accessed March 2011
13. Japan Robotics Association: Worldwide Robotics Market Growth, as quoted by many journals
and companies such as The Robotics Report and Microsoft (2010). www.jara.jp
14. Taiwan aims to capture share of the robotics market. NY Times, 30 Oct 2008
15. See: Guizzo, E., Geminoid, F.: more video and photos of the female android. http://spectrum.
ieee.org/automaton/robotics/humanoids/042010-geminoid-f-more-video-and-photos-of-the-
female-android. Accessed March 2011
16. Division for Applied Robot Technology, KITECH, Ansan, Korea: http://news.nationalgeographic.
com/news/2006/05/android-korea-1.html (2006)
17. Matarić, M.J.: Where to next? In: The Robotics Primer. MIT, Cambridge, MA (2007), Chap. 22
18. Marques, L., de Almeida, A., Tokhi, M.O., Virk, G.S. (eds.): In: Advances in Mobile Robots,
Proceedings of 11th International Conference on Climbing and Walking Robots and the
Support Technologies for Mobile Machines, Coimbra, Portugal. World Scientific, Singapore
(2008)
19. Lillehei, C.W., Gott, V.L., Hodges Jr., P.C., Long, D.M., Bakken, E.E.: Transistor pacemaker
for treatment of complete atrioventricular dissociation. JAMA 172, 2006 (1960)
20. Noyce, R.N.: Semiconductor device-and-lead structure. US Patent No. 2981877, filed 30 July
1959. Issued 25 Apr 1961
21. Graf, H.-G., Dollberg, A., Schulze Sp€ untrup, J.-D., Warkentin, K.: “HDR sub-retinal implant
for the vision-impaired”, in High-Dynamic-Range (HDR) Vision. In: Hoefflinger, B. (ed.)
Springer Series in Advanced Microelectronics, vol. 26, pp. 141–146. Springer, Berlin/
Heidelberg (2007)
22. www.givenimaging.com
23. MTB Europe: surgical robot market share, market strategies, and market forecasts, 2008–2014.
www.mtbeurope.info/bi/reports/
24. Hoefflinger, B.: Computer-assisted motoring. In: Calder, N. (ed.) Scientific Europe, pp. 28–33.
Natuur &Techniek, Maastricht (1990)
25. Hoefflinger, B.: Special issue on automotive electronics, traffic management, intelligent
control, roadside information, vision enhancement, safety and reliability. IEEE Micro. 13(1),
11–66 (1993)
26. Knoll, P.M.: “HDR vision for driver assistance”, in High-Dynamic-Range (HDR) Vision. In:
Hoefflinger, B. (ed.) Springer Series in Advanced Microelectronics, vol. 26, pp. 123–136.
Springer, Berlin/Heidelberg (2007)
Chapter 7
ITRS: The International Technology Roadmap
for Semiconductors
Bernd Hoefflinger
Abstract In a move singular for the world’s industry, the semiconductor industry
established a quantitative strategy for its progress with the establishment of the
ITRS. In its 17th year, it has been extended in 2009 to the year 2024. We present
some important and critical milestones with a focus on 2020. Transistor gate
lengths of 5.6 nm with a 3 sigma tolerance of 1 nm clearly show the aggressive
nature of this strategy, and we reflect on this goal on the basis of our 10 nm
reference nanotransistor discussed in Sect .3.3. The roadmap treats in detail the
total process hierarchy from the transistor level up through 14 levels of metallic
interconnect layers, which must handle the signal transport between transistors and
with the outside world. This hierarchy starts with a first-level metal interconnect
characterized by a half-pitch (roughly the line width) of 14 nm, which is required to
be applicable through intermediate layers with wiring lengths orders of magnitude
longer than at the first local level. At the uppermost global level, the metal pattern
has to be compatible with high-density through-silicon vias (TSV), in order to
handle the 3D stacking of chips at the wafer level to achieve the functionality of the
final chip-size product. At the individual wafer level, the full manufacturing process
is characterized by up to 40 masks, thousands of processing steps and a cumulative
defect density of hopefully <1/cm2.
7.1 History and 2010 Status of the Roadmap
The organization of the US semiconductor industry, the SIA (Semiconductor

Industry Association), launched its first roadmap in 1992. From the experience
with Moore’s law of 1965 [1] and 1975 [2] (Fig. 2.10) and of the “scaling” law of
1974 [3] (Fig. 2.11), a forecasting scheme was derived with a 15-year horizon. That
B. Hoefflinger (*)

162 B. Hoefflinger
effort was joined by the semiconductor industry in Europe, Japan, Korea, and
Taiwan in 1998, and the first International Technology Roadmap for Semi-
conductors (ITRS) was issued in 1999. A fully revised roadmap has been issued
since in every odd-numbered year. For our discussion here, we rely on the 2009
roadmap reaching out to the year 2024 [4]. Its aim is to provide the best present
estimate (of the future) with a 15-year horizon.
The roadmap started out in the 1990s with shrink factors of 0.7/2 years for
minimum dimensions and other dimensional parameters such as film thicknesses,
lateral tolerances within layers and between layers, etc., which had to be scaled
concurrently. This target orientation had a tangible influence on progress, evident in
the 1999 speed-up on scaling to 0.7/18 months (Fig. 2.17). In the meantime, the
roadmap process and structure has evolved into a highly complex, multidimen-
sional effort with major enhancements:
2001: system drivers
2005: emerging research devices (ERD)
2007: emerging research materials (ERM).
As a result, numerous international working groups have been established.
Table 7.1 is a listing, together with references to the chapters in the present book
in which related issues are addressed by specialists.
Broad global assumptions and stiff models of exponential progress are charac-
teristic of the roadmap, supported so far by 40 years of history. We start with the
S-curve for the introduction of new products as shown in Fig. 7.1.
In Fig. 7.1, first conference paper means first product announcement (typically at
the ISSCC). It is also characteristic that the specific equipment for a new generation
becomes available in an early (alpha) version only 3 years before the mass produc-
tion of chips. Furthermore, lead-time from samples to volume is less than
12 months. If the ITRS or the industry talks of a technology node, it means the
start of volume production. For calibration: 2010 is the year of the 32 nm node. The
32 nm here in the ITRS defines the smallest half-pitch of contacted metal lines on
any product. Model assumptions are also made for the overall innovation process,
as we see in Fig. 7.2.
Table 7.1 Structure of the 2009 ITRS and related chapters in the present book
System drivers Chaps. 6, 10–13
Design Chap. 9
Test and test equipment Chap. 21
Process integration, devices and structures Chap. 3
RF and analog/mixed-signal technologies Chaps. 4, 5
Emerging research devices Chaps. 3, 23
Front-end processes Sect. 3.2
Lithography Chap. 8
Interconnects Sect. 3.6, Chap. 5
Factory integration Chap. 21
Assembly and packaging Sect. 3.8, Chaps. 12, 21
7 ITRS: The International Technology Roadmap for Semiconductors 163
Fig. 7.1 Production ramp-up and technology-cycle timing [ITRS 2009]. ERD/ERM emerging
research devices/materials, PIDS process integration and device structures (# SEMATECH)
Fig. 7.2 Typical ramp from research to production. The example illustrates the assumption of
introducing high-electron-mobility materials for the transistor channel by 2019 [4]. FEP: Front
end processing (# SEMATECH)
The assumed R&D lead-time of 10 years in the example emphasizes that any
new chip technology with an economic impact in 2020 should have been presented
in a credible reduction to practice by 2010.
Based on these fundamentals, the core of the roadmap continues to be the overall
roadmap technology characteristics (ORTC).
164 B. Hoefflinger
7.2 ORTC: Overall Roadmap Technology Characteristics
These characteristics are the smallest lengths for gates and linewidths of first-level
metal introduced in a product in a given year (Table 7.2).
The trends expressed in this data are best illustrated by the lines in Figs. 7.3 and 7.4.
The basic observation on these trend lines is that progress is slowing: The 0.7
shrink, which classically happened every 2 years and even saw a speed-up in the
decade 1995–2005 to 18 months, has slowed to 3 years and to almost 4 years for the
physical gate length in logic chips.
The 2009 roadmap draws a magic line at the 16 nm node in 2016. The roadmap
data is called near-term to this year (2016), and it happens to bear significance,
because, after 2016 and the 16 nm node, the solutions to most of the technology
issues related to a straight scaling strategy had to be declared as unknown in
Table 7.2 Key lithography-related characteristics by product and year [ITRS 2009]
Near-term years 2010 2012 2014 2016
Flash Poly Si ½ pitch [nm] 32 25 20 16
MPU/ASIC first metal ½ pitch [nm] 45 32 24 19
MPU physical gate length [nm] 27 22 18 15
Long-term years 2018 2020 2022 2024
Flash poly Si ½ pitch [nm] 12.6 10.0 8.0 6.3
MPU/ASIC first metal ½ pitch [nm] 15.0 11.9 9.5 7.5
MPU physical gate length [nm] 12.8 10.7 8.9 7.4
Fig. 7.3 ITRS trend lines for metal-1 half-pitch of DRAM and flash memory [4] (# SEMATECH)
Fig. 7.4 ITRS trend lines for gates and metal-1 half-pitch in microprocessors and high-
performance ASICs [4] (# SEMATECH)
2009, whereas with the lead-time model of Fig. 7.2, they should have been known
since 2004–2006. As a case in point, lithography and transistor performance are
highlighted here: As outlined in Chap. 8 on nanolithography, there is in 2010 no
known lithography technology for 16 nm and beyond other than low-throughput
multiple-electron-beam (MEB) lithography. As described in Sect. 3.2, MOS
transistors with an equivalent oxide thickness (EOT) of 1 nm and L < 15 nm are
a step back in current/mm, have fundamentally large variances of threshold and
current and mostly no intrinsic gain.
In Sect. 7.4, we discuss to what extent emerging research devices can remove
this roadblock by 2020. In any event, the historic scaling strategy is coming to an
end in a window of 16–10 nm. Is this the end of the chip industry as the world leader
in market growth? To answer this question, it is worthwhile forgetting the ORTC
scaling trend lines as a self-fulfilling prophecy or driver for a moment and to look at
the solely important product and service drivers.
7.3 System Drivers
The system-drivers working group was established in 2001 in order to focus on the
requirements of the various chip markets, in a way a late recognition of a market-
driven strategy (market pull) after the Y2K crash and after 40 years of an unparal-
leled technology push provided by a Moore’s law implemented by scaling or, in
other words, forceful 2-year successions of new technology generations.
166 B. Hoefflinger
The International Technology Working Groups (ITWGs) are a mix of market

(horizontal) and specialty (vertical) groups in the sense of the matrix in Chap. 6 of
this book, recalled here as Table 7.3. Our expert authors deal with the requirements
and challenges in the specialty areas in the chapters indicated in the table.
The extensive listings of requirements and challenges of the 16 working groups
of the ITRS show common denominators for the required rates of progress in terms
of product performance:
– Double the computing/communication speed every 2 years
– Double the memory density every 2 years.
This cannot be achieved by scaling anymore, and, in fact, it has not been achieved
by scaling since the mid-1990s, but rather by “engineering cleverness”, as demanded
by Moore in the 1980s. The multiprocessor cores in computing and the multilevel
storage/transistor and stacked memory chips are striking examples of this cleverness.
Clearly, the energy per function, CV 2, demands continuous and forceful reduc-
tion of the capacitance C and the operating voltage V, and the decisive contributions
have to come from
– The reduction of wire lengths per function
– The reduction of transistor sites per function
– The reduction of the bandwidth between processors and memories
– The reduction of voltage swing (e.g., low-voltage differential signaling, LVDS).
The progress on these issues is critical, and it will be addressed again in Sect. 7.4.
As to system requirements, we have to look some more at the 2/2 years rate, as
aggressive as it may be. It promises a 32 progress per decade. Just looking at the
personal mobile assistant of 2020 (Chap. 13), we see that advances 300 between
2010 and 2020 have been demanded and have been considered to be feasible. This
strategizing clearly requires quantum jumps in performance beyond the 2/2 years
improvement. The multi-bit/cell flash RAM is the most prominent example of this
phenomenon (Fig. 7.5).
Operations per second per milliwatt is the key metric for graphics and image
processing, and it is here that rich media and 3D vision require a 300-fold improve-
ment between 2010 and 2020. HIPERLOGIC in Sect. 3.6 and digital neuro-
processing in Chap. 16 are indicative of such quantum jumps in performance
Table 7.3 Markets (vertical) and special functions (horizontal) of nanochips

Processor Memory Analog/ Transmitter/ Sensor/actuator
Chaps. 10, 13, Chaps. 11, digital receiver Chaps. 14, 15, 17
16 18 Chap. 4 Chap. 5
Computing xxx xxx xx
Communicating xx xx xxx xxx
Consumer x xxx xxx xxx xx
Med./Industrial xx xxx xxx xx xxx
Automotive xx x xxx x xxx
Military xxx xxx xxx xxx xxx
Fig. 7.5 Gigabits/chip for DRAM and flash RAM [4]. SLC single-level cell, MLC multi-level cell
(# SEMATECH)
beyond just an extension of the mainstream by 2/2 years. In view of the 16 nm

barrier, the ITRS puts high hopes on emerging research devices.
7.4 ERD: Emerging Research Devices
Numerous keywords for future devices and new information paradigms are men-
tioned by this working group. With reference to the innovation curve in Fig. 7.2 as
applied to product introduction by 2020, we focus on enhanced MOS transistors with
gates 10 nm long. The roadmap shows a possible progression, reproduced in Fig. 7.6.
The most important, 10, R&D tasks are clearly identified in this figure:
1. Transition to metal gate
2. Transition to high-k (high-dielectric-constant) gate dielectric
3. Transition to fully depleted silicon-on-insulator (FD SOI)
4. Transition to multiple-gate (MuG) transistors
While items 1 and 2 have been introduced into production, continuous improve-
ment is necessary. Item 3, thin Si films on insulator, with thicknesses <5 nm, fully
depleted, are a matter of complex trade-offs and closely related to the challenging
introduction of multiple-gate transistors, either as
– Dual-gate transistors or as
– Surround-gate transistors.
168 B. Hoefflinger
Fig. 7.6 A possible progression to a 10 nm transistor in 2020 [4] (# SEMATECH)
These transistor types upset the physical-design topographies so much engraved

in the industry that their introduction by any company will really mean a quantum
step in the industry. While dual-gate and tri-gate transistors have been introduced in
2011, and several interesting and powerful functionalities have been described in
Sect. 3.5, the surround-gate transistor is a long-term topic closely associated with
the nanotube-transistors (Sect. 3.8).
Figure 7.6 also points out new materials for the transistor channels. Strained Si
and SiGe have already been introduced to achieve higher electron mobilities and
hole mobilities coming closer to those of ideal electrons for speed and symmetrical
operation. III–V materials for MOS transistor channels are considered with opti-
mism to benefit from a fivefold or higher mobility. However, we should keep in mind
that what counts in switching circuitry is maximum current, and this is determined
by maximum velocity, (3.15), and this would be at best two times higher. While this
might still be worth the effort, the processing and compatibility challenges and the
loss of a technology with complementary transistors have to be weighed in.
The emphasis on speed is only justified for the gighertz and terahertz community
of superprocessors and broadband communication. For all others, energy/function
is the issue. Here we deal with the world below 400 mV. We studied this world
extensively in Sects. 3.2 and 3.3, and the most important messages we learned are:
1. The maximum operating voltage for a 1 nm (EOT) gate transistor is 500 mV.
2. At <500 mV, MOS transistors operate in the barrier-control (diffusion, bipolar,
or leakage) mode, where the potential barrier is controlled by both the gate and
the drain voltages (3.17, 20).
3. The transconductance of a MOS transistor at room temperature is a decade of
current per120 mV at best, with a fundamental 4s spread to <300 mV for
L ¼ 10 nm.
4. The intrinsic voltage gain for tI ¼ 1 nm (EOT) and L ¼ 10 nm is unity, i.e., no

gain without channel engineering (3.22).
5. The maximum current is again determined by the maximum high-field velocity
and not by low-field mobility (3.16, 21).
Again, the mobility advantage of III–V materials and Ge is not relevant. Fur-
thermore, the temperature dependence of currents in the barrier-control mode is
rather strong. This means that on–off current ratios shrink rapidly at higher
temperatures, a problem for Si and likely the out for Ge as a channel material.
6. The fundamental variance of nanotransistors due to Poisson distributions of
charges (acceptors, donors, channel charges) requires that any signal regenera-
tor/amplifier has to be made up of complementary transistors, and, in fact,
differential signaling (LVDS), and cross-coupled complementary transistor
pairs are required (Sect. 3.3) for safe logic operations.
These stringent criteria have to be observed in the evolution of scaled-down
MOS transistors but also for any possible replacements in the real world of chips for
the masses, which is at 20 C and mostly significantly higher on-chip.
Beyond the scenario of Fig. 7.6 and a 2020 horizon, emerging research pursues
many directions in the absence of a winner. The ERD committee has assessed more
than 15 different technologies in a multidimensional performance evaluation. We
select the evaluation of two types of transistors, which we discussed in Sect. 3.8:
– The carbon-nanotube (CNT) MOSFET
– The single-electron transistor (SET)
Their evaluation, as shown in Fig. 7.7, shows eight well-chosen performance
criteria:
– Performance
– Energy efficiency
– Gain
– Scalability
– Compatibility with CMOS architecture
– Compatibility with CMOS technology
– Operational temperature
– Reliability
Apart from the quantitative assessment of the criteria, the important observation
on these diagrams is the evolution of the ratings over time: From green in 2005
through blue in 2007 to red in 2009. Although a span of four years is short on a
research time scale, the apparent trend is that almost all values of 2005 looked
optimistic compared with the assessment in 2009. This appearance of more realism
is frequent in the 2009 issue of the roadmap. It is an indicator that more fundamental,
radical innovations are seen as having an impact farther in the future than assumed in
recent years. This does not mean that the progression of electronic chips has slowed,
but that other disruptive innovations have taken center stage, which are less nano and
quantum-driven. Interconnects are a superb example of this phenomenon.
170 B. Hoefflinger
a
2005 (1D Structures) Scalability Performance 2009
3
2007 (1D Structures)
2
CMOS Architectural
Energy Efficiency
Compatibility
1
CMOS Technological Gain

Compatibility
Operational Temperature Operational Reliability

b
2005 (SETs) Scalability
3
Performance 2009
2007 (SETs)
2
CMOS Architectural Energy Efficiency
Compatibility
1
CMOS Technological Gain

Compatibility
Operational Temperature Operational Reliability
Fig. 7.7 Evaluation of (a) carbon nanotube MOSFETs and (b) single-electron transistors [4]
(# SEMATECH)
7.5 Interconnects
For the integrated circuit functions on a chip, thousands to billions of transistors

have to be interconnected by thin-film metal layers, each separated from those
above and below by inter-level dielectric insulators. The schematic cross section in
Fig. 7.8 of the surface of a chip illustrates the resulting structure.
At the bottom, we see a pair of complementary transistors, each in its well for
isolation. Blue tungsten plugs reach up to the first metal interconnect layer. The
manufacturing process up to this level is called the front-end of line (FEOL). What
Fig. 7.8 Schematic cross section through the surface of a CMOS chip illustrating the wiring
hierarchy. Blue and orange: metal, gray: insulators [4] (# SEMATECH)
follows in the example is three layers of intermediate-level wires, two layers of

semi-global, and another five layers of global interconnects. This hierarchy visibly
demonstrates the most serious roadblock for nanoelectronics:
The multi-level interconnect with its problems of
– Resistance, capacitance, inductance
– Loading the transistors
– Cross talk
– Area and volume
– Overlay tolerance
– Cost and yield
– Reliability
The scaling of this maze of inverted skyscrapers with skywalks at up to 14 levels
is a highly complex challenge, and the roadmap demands a progression as shown in
Fig. 7.9, for which manufacturable solutions beyond 2015 are not known or have
severe drawbacks in resistance, crosstalk and reliability. In any event, the pressure to
– Reduce operating voltages
– Use low-voltage differential signaling (LVDS)
172 B. Hoefflinger
Fig. 7.9 Projected density of 1011 x

contacts, vias, TSV’s, and
bumps
1010 x
Contacts
109 x
On-Chip Vias
108
Density (cm–2 )
107
106
TSV´s
105
mbumps
104
1990 2000 2010 2020
Year
mounts steadily.
At the same time, lateral wiring lengths are becoming excessive, because of the
wider pitch, which has to be accepted at the higher, intermediate, and global levels,
and also because of the slower pace of scaling transistors and, in fact, the end of the
roadmap for lateral transistor dimensions.
This scenario has finally caused a widespread commitment since about 2005 to
reduce lateral wiring lengths and chip sizes by stacking chips on top of each other,
Fig. 7.10 Chip cross sections showing TSV first, at the end of FEOL, and last [4] (#ITRS)
as we discussed in Sect. 3.7. Many alternatives for stacking exist, depending on chip
functionality and the manufacturers involved, leading to new networks and action
chains. We highlight these possible scenarios with the illustration of three
alternatives for producing vias, which have to go through the (thinned) chips to
carry signals and power from one chip to the next (Fig. 7.10).
Assuming a SOI chip, TSV-first means that the via is formed literally before the
formation of the transistor layer. In the center, the alternative TSV-middle is shown,
where the TSV is formed at the end of the FEOL, together with the first metal.
At the back-end-of-line (BEOL) on the right, a voluminous TSV-last is shown
going through the chip, indicative of the aspect ratios that have to be realized by
appropriate cutting and/or etching processes.
Bearing in mind the serious lithography limits on a volume-production level
below 15 nm and the conductance limits of metal wires at this scale, the further
scaling-down of interconnects is clearly the number one area for alternative
wires/interconnects such as carbon nanotubes and graphene lines because of
their fundamental advantages but also because they can be produced without
lithography [5–7].
References
1. Moore, G.: Cramming more components onto integrated circuits. Electron. Mag 38(8), 19 Apr
1965
2. Moore, G.: Progress in digital integrated electronics. IEEE IEDM (International Electron
Devices Meeting) , Technical Digest, pp. 11–13 (1975)
3. Dennard, R.H., Gaensslen, F.H., Yu, H.N., Rideout, V.L., Bassous, E., LeBlanc, A.R.: Design of
ion-implanted MOSFET’s with very small physical dimensions. IEEE J Solid-St Circ 9, 256
(1974)
174 B. Hoefflinger
4. www.itrs.net/reports.html. Semiconductor Industry Association. The International Technology

Roadmap for Semiconductors, 2009 Edition. International SEMATECH: Austin, TX 2009
5. Appenzeller, J., Joselevich, E., Hoenlein, F.: Carbon nanotubes for data processing. In: Waser,
R. (ed.) Nanoelectronics and Information Technology, 2nd edn. Wiley-VCH, Weinheim (2005)
6. Wolf, E.L: Chap. 5, Some newer building blocks for nanoelectronics. In: Quantum
Nanoelectronics, Wiley-VCH, Weinhiem (2009)
7. Hoenlein, W., Kreupl, F., Duesberg, G.S., Grahem, A.P., Liebau, M., Seidel, R., Unger, E.:
Carbon nanotube applications in microelectronics. In: Siffert, P., Krimmel, E.F. (eds.) Silicon:
Evolution and Future of a Technology, pp. 477–488. Springer, Berlin/Heidelberg (2004)
Chapter 8
Nanolithography
Burn J. Lin
Abstract Nanolithography, the patterning of hundreds to thousands of trillions of

nano structures on a silicon wafer, determines the direction of future nano-
electronics. The technical feasibility and the cost-of-ownership for technologies
beyond the mark of 22 nm lines and spaces (half pitch) is evaluated for Extreme-
Ultra-Violet (EUV = 13 nm) lithography and Multiple-Electron-Beam (MEB)
direct-write-on-wafer technology. The 32 nm lithography-of-choice, Double-
Patterning, 193 nm Liquid-Immersion Optical Technology (DPT) can possibly be
extended to 22 nm with restrictive design patterns and rules, and it serves as a
reference for the future candidates.
EUV lithography operates with reflective mirror optics and 4 multi-layer
reflective masks. Due to problems with a sufficiently powerful and economical
13 nm source, and with the quality and lifetime of the masks, the introduction of
EUV into production has been shifted to 2012.
MEB direct-write lithography has demonstrated 16 nm half-pitch capability, and
it is attractive, because it does not require product-specific masks, and electron
optics as well as electron metrology inherently offer the highest resolution. How-
ever, an MEB system must operate with more than 10,000 beams in parallel to
achieve a throughput of ten 300 mm wafers per hour. A data volume of Peta-Bytes
per wafer is required at Gigabit rates per second per beam. MEB direct-write would
shift the development bottleneck away from mask technology and infrastructure to
data processing, which may be easier.
The R&D expenses for EUV have exceeded those for MEB by two orders of
magnitude so far. As difficult as a cost model per wafer layer may be, it shows
basically that, as compared with DPT immersion optical lithography, EUV would
be more expensive by up to a factor of 3, and MEB could cost less than present
B.J. Lin (*)

TSMC, 153 Kuang Fu Road, Sec. 1, Lane 89, 1st Fl., Hsinchu R.O.C. 300, Taiwan
e-mail: burnlin@tsmc.com

176 B.J. Lin
optical nanolithography. From an operational point-of-view, power is a real con-

cern. For a throughput of 100 wafers per hour, the optical system dissipates 200 kW,
the MEB system is estimated at 170–370 kW, while the EUV system could
dissipate between 2 and 16 MW.
This outlook shows that the issues of minimum feature size, throughput and cost
of nanolithography require an optimisation at the system level from product
definition through restrictive design and topology rules to wafer-layer rules in
order to control the cost of manufacturing and to achieve the best product.
Optical patterning through masks onto photosensitive resists proceeded from

minimum features of 20 mm in 1970 to 22 nm in 2010. The end of optical
lithography was initially predicted for 1985 at feature sizes of 1 mm, and hundreds
of millions of dollars were spent in the 1980s to prepare X-ray proximity printing
as the technology of the future, which has not happened. The ultimate limit of
optical lithography was moved in the 1990s to 100 nm, initiating next-generation-
lithography (NGL) projects to develop lithography technology for sub-100-nm
features and towards 10 nm. This chapter is focused on two of these NGL tech-
nologies: EUV (extreme-ultraviolet lithography), and MEB (multiple-electron-
beam lithography). These are compared with the ultimate optical lithography,
also capable of delineating features of similar size: DPT (double-patterning,
liquid-immersion lithography).
8.1 The Progression of Optical Lithography
Optical lithography has supported Moore’s Law and the ITRS (Chap. 7) for the first
50 years of microelectronics. The progression in line-width resolution followed the
relationship
l
Resolution ¼ k1 :
NA
Shorter-wavelength optical and resist systems, and the continuing increase of the
numerical aperture (NA) as well as reduction of the resolution-scaling factor k1
provided a remarkable rate of progress, as shown by the data on the period
1999–2012 (Figs. 8.1 and 8.2).
The medium for optical patterning was air until 2006, when k1 ¼ 0.36 at the
193 nm line of ArF laser sources was reached. The single biggest step forward was
made when the lens–wafer space was immersed in water, and the NA effectively
increased by 1.44 according to the refractive index of water at 193-nm wave-
length, so that, with the inclusion of restricted design rules and more resolution-
enhancement techniques (RETs), the 32-nm node could be reached in 2010.
8 Nanolithography 177
Fig. 8.1 Lens and reticle parameters, OPC, and RETs employed for the technology nodes from
1999 to 2012 [1]. OPC: Optical proximity correction, OAI: Off-axis illuminatiom, HLC:
Handcrafted layout compensation, RB OPC: Rule-based OPC, MB OPC: Model-based OPC,
HB OPC: Hybrid-based OPC, AF: Assist features, DFM: Design for manufacturability, S2E
DFM: Shape-to-electric DFM, MB AF: Model-based assist features, RDR: Restricted design
rules. # 2009 IEEE
l
Resolution = k1
NA
Fig. 8.2 Resolution for half pitch [(linewidth + space)/2] as wavelength reduction progresses [1].
# 2009 IEEE
With the introduction of DPT, the mask cost is doubled, the throughput is
halved, and the processing cost more than doubled, making the total wafer-
production cost unbearably high and causing serious economic effects.
178 B.J. Lin
A 2010 state-of-the-art optical-lithography system with ArF water-immersion

DPT can be characterized by the following parameters:
Minimum half pitch 22 nm at k1 ¼ 0.3, l ¼ 193 nm, NA ¼ 1.35

Single-machine overlay <2 nm
Reticle magnification 4
Throughput ~175 wph (300 mm diameter wafer)
Price ~€40 M
Related equipment ~US$15 M
Footprint ~18 m2 including access areas
Wall power 220 kW
This most advanced equipment is needed for the critical layers in a 32-nm
process:
– Active
– Gate
– Contact
– Metal 1–n
– Via 1–n.
These layers form the critical path in the total process, which can involve more
than 10 metal layers resulting in a total mask count >40. Liquid-immersion DPT
can be pushed to 22 nm with restricted design rules, limited throughput, and
a significant increase in cost. It is desirable to use less expensive alternative
technologies at and beyond the 22 nm node. The following sections are focused
on the two major contenders for production-level nanolithography at 32 nm and
beyond: EUV and MEB. Nanoimprinting holds promise for high replicated resolu-
tion. However, fabrication difficulties of the 1 templates, limited template dura-
bility, expensive template inspection, and the inherent high defect count of the
nanoimprint technology prevent it from being seriously considered here [7].
8.2 Extreme-Ultraviolet (EUV) Lithography
The International SEMATECH Lithography Expert Group decided in 2000 that

EUV lithography should be the NGL of choice. The EUV wavelength of 13.5 nm
means a > 10 advantage in resolution over optical, as far as wavelength is
concerned, and sufficient progress overall, even though the k1 and the NA with
the necessary reflective-mirror optics would be back to those of the early days with
optical lithography. The system principle is shown in Fig. 8.3.
The 13.5 nm radiation is obtained from the plasma generated by irradiating a Sn
droplet with a powerful CO2 laser. This is realized with an efficiency of 0.2–0.5% from
CO2 laser irradiation. The numbers in the figure illustrate the power at each component
along the path of the beam, required to deliver 1 mJ/cm2 of EUV light at the wafer at
the throughput of 100 wph (wafers per hour). About 0.1% of the EUV power arrives on
In-band
EUV light On collector
26.8P W 9.36P Watt
2nd normal
On 1st NI incidence
mirror 6.37P mirror
Watt
On mask
550P mW
Grazing On wafer
incidence 2
1P mJ / cm ,
mirrors 30P mW for
M6
M1 100 wph
370P mW
M2 M3
M5
M4
Fig. 8.3 EUV illuminator and imaging lens. It is an all-reflective system [1]. # 2009 IEEE
the wafer. Evidently, the overall power efficiency from the laser to the wafer is
extremely small. This poses a serious problem: If we consider a resist sensitivity of
30 mJ/cm2 at 13.5 nm, hundreds of kilowatts of primary laser power are needed.
Other extreme requirements [2] are:
– Roughness of the mirror surfaces at atomic levels of 0.03–0.05 nm, corres-
ponding pictorially to a 1-mm roughness over the diameter of Japan.
– Reflectivity of mirrors established by 50 bilayers of MoSi/Si with 0.01 nm
uniformity.
– Reticle cleanliness in the absence of pellicles and reticle life at high power
densities.
– Reticle flatness has to be better than 46.5 nm, which is 10 better than the
specification for 193-nm reticles.
– New technology, instrumentation, and infrastructure for reticle inspection and
repair.
Nevertheless, EUV lithography has received R&D funding easily in the order
of US$1 billion by 2010. Two alpha tools have been installed, producing test
structures like those shown in Fig. 8.4.
In any event, EUV lithography puts such strong demands on all kinds of
resources that, if successful, it can be used by only a few extremely large chip
180 B.J. Lin
CD = 52 nm
Focus = 0nm
CD = 53nm CD = 54nm
N32 SRAM contact holes

Focus = –40nm Focus = + 40nm
Fig. 8.4 Micrographs of SRAM test structures: 32-nm contact holes with a pitch of 52–54 nm [1].
# 2009 IEEE
manufacturers or consortia of similar size and technology status. Attaining lower

imaging cost than DPT cannot be taken for granted. A 2012 system can be esti-
mated to have the following characteristics:
Minimum half pitch 21 nm at k1 ¼ 0.5, l ¼ 13.5 nm, NA ¼ 0.32

Single-machine overlay <3.5 nm
Reticle magnification 4
Throughput 125 wafers (300-mm dia. wafer)
Price €65 M
Related equipment US$20 M (excluding mask-making equipment)
Footprint 50 m2 including access areas
Wall power 750–2,000 kW
If we consider that a Gigafactory requires 50 such systems for total throughput,

we can appreciate that this technology is accessible to a few players only. Further
data on these issues will be discussed in Sect. 8.4.
8.3 Multiple-Electron-Beam (MEB) Lithography
Electron-beam lithography has been at the forefront of resolution since the late
1960s, when focused electron beams became the tools of choice to write high-
resolution masks for optical lithography or to write submicrometer features directly
on semiconductor substrates. As a serial dot-by-dot exposure system, it has the
reputation of being slow, although orders-of-magnitude improvements in through-

put were achieved with
– Vector scan (about 1975)
– Variable shaped beam (1982)
– Cell projection through typically 25 Si stencil masks (1995)
A micrograph of a 1997 stencil mask with 200-nm slots in 3-mm Si for the gates
in a 256 Mb DRAM is shown in Fig. 8.5.
Electron beams have many unique advantages:
– Inexpensive electron optics
– Large depth-of-field
– Sub-nanometer resolution
– Highest-speed and high-accuracy steering
– Highest-speed On-Off switching (blanking)
– Secondary-electron detection for position, overlay, and stitching accuracy
– Gigabit/s digital addressing
The systems mentioned so far were single-beam systems, using sophisticated
electron optics. A scheme to produce a large array of parallel nanometer-size
electron beams is used in the MAPPER system [3], as shown in Fig. 8.6, where
13,000 beams are created by blocking the collimated beam from a single electron
source with an aperture plate. The beam-blanker array acts upon the optical signal
from the fiber bundle to blank off the unwanted beams. The On-beams are
Fig. 8.5 SEM micrograph of a Si stencil mask with 200-nm slots in 3-mm Si for the gates in a
256 Mb DRAM (Courtesy INFINEON and IMS CHIPS)
182 B.J. Lin
Electron Source
Fiber bundle
Collimator lens
Aperture Array
Beam Blanker Array
Beam Deflector Array

Projection lens Array
Fig. 8.6 The MAPPER MEB system [1]. # 2009 IEEE
subsequently deflected and imaged. All electron optics components are made with
the MEMS (micro-electro-mechanical system) technology, making it economically
feasible to achieve such a large parallelism [6].
The MAPPER write scheme is shown in Fig. 8.7. For a 45-nm half-pitch system,
the 13,000 beams are distributed 150 mm apart in a 26 10 mm2 area. The
separation ensures negligible beam-to-beam interaction. Only the beam blur
recorded in the resist due to forward scattering calls for proximity correction. The
beams are staggered row-by-row. The staggering distance is 2 mm. The deflection
range of each beam is slightly larger than 2 mm to allow a common stitching area
between beams, so that the 150-mm spaces in each row are filled up with
patterns after 75 rows of beam pass through. Each beam is blanked off for unexposed
areas [4].
Massive parallelism is not allowed to be costly. Otherwise the cost target to
sustain Moore’s law still cannot be met. One should forsake the mentality of using
the vacuum-tube equivalent of conventional e-beam optics and turn to adopting the
integrated-circuit-equivalent highly-parallel electron optics, whose costs are
governed by Moore’s law. With cost of electron optics drastically reduced, the
dominant costs are now the digital electronics, such as CPU (central processing
Fig. 8.7 MAPPER writing scheme (EO: Electron optics) [1]. # 2009 IEEE
unit), GPU (graphics processing unit), FPGA (field programmable gate array), and
DRAM, required to run the massively parallel, independent channels. For the first
time, the cost of patterning tools is no longer open-ended but rather a self-reducing
loop with each technology-node advance.
Another multiple-beam system uses a conventional column but with a program-
mable reflective electronic mask called a digital pattern generator (DPG) as shown
in Fig. 8.8.
The DPG is illuminated by the illumination optics through a beam bender. The
DPG is a conventional 65-nm CMOS chip with its last metal layer turned towards
the illuminating beam. The electrons are decelerated to almost zero potential with
the Off-electrons absorbed by the CMOS chip while the On-electrons are reflected
back to 50 keV and de-magnified by the main imaging column to expose the wafer.
Even though there are between 1 M and 4 M pixels on the DPG, the area covered is
too small compared to that of an optical scanner. The acceleration and deceleration
of a rectilinear wafer stage to sustain high wafer throughput are impossible to reach.
Instead, a rotating stage to expose six wafers together can support high throughput
with an attainable speed. Again, all components are inexpensive. The data path is
still an expensive part of the system. The cost of the system can be self-reducing.
The fascination with these systems, beyond the fundamental advantage that they
provide maskless nano-patterning, lies in the fact that MEB lithography directly
benefits in a closed loop from the
– Advances in MEMS and NEMS (micro- and nano-electro-mechanical systems)
technology for the production of the aperture and blanking plates as well as
arrays of field-emission tips, and
– Advances in high-speed interfaces to provide terabits/s to steer the parallel
beams.
184 B.J. Lin
Digital
Pattern
Generator
Electron Gun
Projection
Optics
EXB Illumination
Filter Optics
Beam
Bender
Multiple Wafer
Rotary Stage
Fig. 8.8 Second generation reflective electron-beam lithography (REBL) system. EXB: charged
particle separator E B
MEB direct-write lithography demonstrated 16-nm resolution (half-pitch) in

2009.
For high-volume manufacturing (HVM), the basic MEB chamber with 13,000
columns and 10 wph has to be clustered into 10 chambers to reach 100 wph, as
illustrated in Fig. 8.9.
An MEB system with 10 chambers has about the same footprint as the 193-nm
optical immersion scanner. A 2012 assessment would be as follows:
Minimum linewidth 22 nm
Minimum Overlay 7 nm
No reticle Maskless
Number of e-beams 130,000
Throughput 100 (300 mm) wph
Data rate 520 Tbit/s
Estimated Price US$50 M
Footprint 18 m2
HVM clustered production tool:

1m
>13,000 beams per chamber
(10 WPH)
10 WPH x 5 x 2 = 100WPH
Footprint ~ArF scanner
Fig. 8.9 MEB maskless-lithography system for high-volume manufacturing [1]. # 2009 IEEE
The following calculation of data rates illustrates the challenges and possible
system strategies for these maskless systems [5]:
– Total pixels in a field (0 and 1 bitmap)
¼ ð33 mm 26 mm=2:25 nm 2:25 nmÞ 1:1ð10% over scanÞ

¼ 190 Tbits
¼ 21:2 TBytes
– A 300-mm wafer has ~90 fields
– 10 wph ) <0:9 s/field
– 13;000 beams ) data rate >3:75 Gbps=beam!
– In addition to data rate, also challenge in data storage and cost
– Use more parallelism to reduce data rate
MEB lithography by 2010 has seen R&D resources at only a few percent
compared with those for EUV, and no industry consensus has been established to
change that situation. However, this may change in view of the critical comparison
in the following section.
8.4 Comparison of Three Nanolithographies
The status in 2010 and the challenges for the three most viable nanolithographies
for volume production of nanochips with features <32 nm are discussed with
respect to maturity, cost, and factory compatibility. Table 8.1 provides a compact
overview.
186 B.J. Lin
Table 8.1 Maturity, cost, infrastructure, and required investment

Wafer cost Mask cost Infra- Maturity Investment
structure
MEB ML2 Potentially low 0 Ready Massive Badly needed
parallism
needs
most work
Immersion DPT More than 2 SPT Ever increasing Ready Mature No question
EUV From slightly Can be higher Expensive, Not yet Plenty but
below to much than DPT to be never
higher than developed enough
DPT
1000
Ten 10-wph MEB tools

Data equipment on subfloor 100
Footprint of
8400
EUV Scanner
and light
source
1010 Scanner
footprint
with light
4600
4600
900 1000 source

on
subfloor
1000
2020 2020 3795

4020 4020 5795
MEB cluster Scanner EUV Scanner
Fig. 8.10 Footprint comparison of MEB cluster, optical scanner, and EUV scanner [1]. # 2009
IEEE
Single-patterning technology optical immersion lithography (SPT) is taken as

a reference. Given the 2010 status, neither EUV nor MEB is mature for production.
Lithography has become the show-stopper for scaling beyond 22 nm so that serious
investment and strategy decisions are asked for. Cleanroom footprint in Fig. 8.10,
cost per layer in Table 8.3 and electrical power in Table 8.2 are used to provide
guidance for the distribution of resources.
We see that the footprint for the throughput of 100 wph with optical and e-beam
technology is about the same, while that for the EUV scanner is more than double.
Cost estimates, difficult as they may be, are important in order to identify
fundamental components of cost and their improvement (Table 8.3). The numbers
for optical DPT versus SPT are fairly well known. The exposure cost is a strong
function of throughput, so it doubles for DPT. However, because of the extra
etching step required by DPT, the combined exposure-and-processing cost more
Table 8.2 Electrical power (kW) for 32-nm lithography in a 150-MW Gigafactory for 130,000
12 in. wafers/month (HVM: High-volume manufacturing) [1]
kW Immer EUV HVM MEB HVM
scanner
Supplier Supplier 30 mJ/cm2 30 mJ/cm2 resist + Ten Share
estimate estimate instead of conservative collector 10 wph datapath
10 mJ/cm2 and source efficiencies columns
Source 89 580 1,740 16,313 120 120
Exposure unit 130 169 190 190
Datapath 250 53
Total per tool 219 749 1,930 16,503 370 173
Total for 59 tools 12,921 44,191 113,870 973,648 21,830 10,222
Fraction of 8.61% 29.46% 75.91% 649.10% 14.55% 6.81%
scanner
power in fab
130 k wafers per month 1200 fab, 150,000 kW
Table 8.3 Cost per lithography layer relative to single-patterning (SP) water-immersion ArF
lithography [1]
All costs normalized to SP Imm SP Imm SP EUV goal EUV MEB goal MEB
exposure cost/layer possibility possibility
Normalized exposure tool 3.63E + 06 3.63E + 06 4.54E + 06 4.54E + 06 4.54E + 06 2.72E + 06
cost
Normalized track cost 5.58E + 05 5.58E + 05 4.88E + 05 2.09E + 05 4.88E + 05 4.88E + 05
Raw thruput (WPH) 180 100 100 20 100 100
Normalized exposure 1.00 1.80 2.17 10.37 2.17 1.37
cost/layer
Normalized consumable 0.91 2.27 1.43 1.43 1.05 1.05
cost/layer
Normalized expo+consum 1.91 4.07 3.61 11.80 3.22 2.42
cost/layer
than doubles. For EUV lithography, if the development goal of the EUV tool
supplier is met, the combined exposure and processing cost is slightly less than
that of DPT with ArF water-immersion lithography. If the goal is not met, either due
to realistic resist sensitivity or source power, a fivefold smaller throughput is likely.
It drives the total cost per layer to 3 that of optical DPT. With multiple electron
beams meeting a similar throughput-to-tool-cost ratio as the EUV goal, a cost/layer
of 80% against DPT is obtained. Further sharing of the datapath within the MEB
cluster could bring the tool cost down by 40% so that overall cost/layer might drop
to 60% of DPT.
The total electrical power required in a Gigafactory just for lithography will no
longer be negligible with next-generation-lithography candidates. The optical
immersion scanner as a reference requires 219 kW for a 100-wph DPT tool.
A total of 59 scanners to support the 130,000 wafers/month factory, would consume
8.6% of the factory power target of 150 MW.
188 B.J. Lin
We saw in Sect. 8.2 that the energy efficiency of the EUV system is extremely
small. Even the supplier estimate results in a power of 750 kW/tool or 44 MW for
59 tools, corresponding to 30% of the total fab power. If the resist-sensitivity target
of 10 mJ/cm2 is missed by a factor of 3 and the actual system efficiency from source
to wafer by almost another factor of 10 from the most pessimistic estimate, then the
EUV lithography power in the fab rises to unpractical levels of between 141 MW
and 974 MW. The cost to power these tools adds to the already high cost estimate
shown in Table 8.2. It is also disastrous in terms of carbon footprint.
The overall-power estimate for MEB maskless systems arrives at a total power
of 370 kW for the 10-column cluster with 130,000 beams, of which 250 kW is
attributed to the datapath. This would result in 22 MW for the fab, corresponding to
16% of the total fab power, also high compared to that of DPT optical tools. It is
plausible to predict that the power for the datapaths could be reduced to 50 kW with
shared data, resulting in a total of 173 kW for each cluster and a total lithography
power in the plant of 6.8% or less.
References
1. Lin, B.J.: Limits of optical lithography and status of EUV. In: IEEE International Electron
Devices Meeting (IEDM ), 2009
2. Lin, B.J.: Sober view on extreme ultraviolet lithography. J. Microlith. Microfab. Microsyst.
5, 033005 (2006)
3. www.mapperlithography.com
4. Lin, B.J.: Optical lithography: here is Why. SPIE, Bellingham (2010)
5. Lin, B.J.: NGL comparable to 193-nm lithography in cost, footprint, and power consumption.
Microelectron. Eng. 86, 442 (2009)
6. Klein, C., Platzgummer, E., Loeschner, H., Gross, G.: PML2: the mask-less multi-beam solution
for the 32 nm node and beyond. SEMATECH 2008 Litho Forum. www.sematech.org/meetings/
archives/litho/8352/index.htm (2008). Accessed 14 May 2008
7. Lin, B.J.: Litho/mask strategies for 32-nm half-pitch and beyond: using established and adven-
turous tools/technologies to improve cost and imaging performance. Proc. SPIE 7379, 1 (2009)
Chapter 9
Power-Efficient Design Challenges
Barry Pangrle
Abstract Design teams find themselves facing decreasing power budgets while
simultaneously the products that they design continue to require the integration of
increasingly complex levels of functionality. The market place (driven by consumer
preferences) and new regulations and guidelines on energy efficiency and environ-
mental impact are the key drivers. This in turn has generated new approaches in all
IC and electronic system design domains from the architecture to the physical
layout of ICs, to design-for-test, as well as for design verification to insure that
the design implementation actually meets the intended requirements and
specifications.
This chapter covers key aspects of these forces from a technological and market
perspective that are driving designers to produce more energy-efficient products.
Observations by significant industry leaders from AMD, ARM, IBM, Intel, nVidia
and TSMC are cited, and the emerging techniques and technologies used to address
these issues now and into the future are explored.
Topic areas include:
– System level: Architectural analysis and transaction-level modeling. How archi-
tectural decisions can dramatically reduce the design power and the importance
of modeling hardware and software together.
– IC (Chip) level: The impact of creating on-chip power domains for selectively
turning power off and/or multi-voltage operation on: (1) chip verification, (2)
multi-corner multi-mode analysis during placement and routing of logic cells
and (3) changes to design-for-test, all in order to accommodate for power-gating
and multi-voltage control logic, retention registers, isolation cells and level
shifters needed to implement these power saving techniques.
B. Pangrle (*)
Mentor Graphics, 46871 Bayside Parkway Fremont, CA 94538, USA
e-mail: barry_pangrle@mentor.com

190 B. Pangrle
– Process level: The disappearing impact of body-bias techniques on leakage

control and why new approaches like High-K Metal Gate (HKMG) technology
help but don’t eliminate power issues.
Power-efficient design is impacting the way chip designers work today, and this
chapter focuses on where the most significant gains can be realized and why power-
efficiency requirements will continue to challenge designers into the future. Despite
new process technologies, the future will continue to rely on innovative design
approaches.
9.1 Introduction
Design teams find themselves facing decreasingly tighter power budgets while,
simultaneously, the products that they are designing continue to require increas-
ingly complex levels of integration and functionality. Consumer preferences, new
regulations and guidelines on energy efficiency and environmental impact are key
drivers in the market place. This in turn has generated new approaches in all IC and
electronic-system-design domains from the architecture to the physical layout of
ICs, to design-for-test, as well as for verification to insure that the design imple-
mentation actually meets the intended requirements and specifications.
This chapter covers key aspects of these forces from the technological and
market perspectives that are driving designers to produce more energy-efficient
products. The observations of key industry leaders from AMD, ARM, IBM, Intel,
nVidia, and TSMC are cited, and the emerging techniques and technologies used to
address these issues now and projected into the future are explored.
Product power efficiency significantly impacts the consumer-electronics industry.
It is a major driver in terms of consumers’ requirements and preferences for mobile
and portable handheld devices. As far back as June 18, 2007, when Apple Inc.
announced that the battery-life runtimes increased for its iPhone, Apple stock
jumped 3.81% that day (As a reference comparison, the NASDAQ composite
index closed down a fraction of a percent that same day). Consumers’ preference
for lower power devices has definitely commanded the attention of the chief officers
that run large corporations producing consumer-electronic products.
The so-called “green” movement to create environmentally friendly products
also has an impact on the industry and extends well beyond small handheld devices
to computers, household appliances, and other electronics. The U.S. Environmental
Protection Agency’s (EPA) Energy Star program created in 1992 is one such
example. The agency sets up requirements for Energy Star certification that allows
companies to place the Energy Star logo onto their products. This program has
gained consumer mind-share, and customers who want to save energy look for this
label on products when comparison shopping. This program has proven so successful
that it has also been adopted in Australia, Canada, Japan, New Zealand, Taiwan,
and the European Union.
9 Power-Efficient Design Challenges 191
9.2 Unique Challenges
For the designers of electronic devices, the task of reducing power has introduced
new challenges. Traditionally for CMOS designs, most of the logic could be
designed under the assumption that the power and ground supply lines on the
chip were constants. The only variations were mainly caused by coupling noise
and voltage drop due to the inherent resistance of the supply and ground lines. In
order to squeeze as much energy efficiency out of their designs as possible,
designers now have to design logic blocks where the supply voltage may be set
for multiple operating points and in some instances totally turned off altogether.
This brings on a whole new set of challenges that did not exist under the old
constant-supply-source design style.
When blocks of a design operate at different voltages, it is necessary to check
that the signals that are used to communicate between these blocks are driven at
appropriate voltage levels for correct operation. For example, if one block (Block A)
has its supply voltage (Vdda) set to 0.8 V and it is sending a signal to another block
(Block B) that has its supply voltage (Vddb) set to 1.25 V, it is quite possible that a
signal originating from A and arriving in B may not be able to send a logical ‘1’
value that is recognizable by B. Since 0.8 V is only 64% of Vddb, it is actually quite
likely that B will see this as an “unknown” value, that is, not being a valid ‘1’ or ‘0’.
This can lead to major problems with the functionality of the design and will be
discussed in more detail later in this chapter. The same applies to signals that cross
between blocks that are powered up and powered down independently. Special
precautions need to be taken so that input signals to blocks that are still powered ON
are not just left floating. For blocks that are completely powered down, the designer
also has to answer the question of what is the intended state of the block when it is
powered back ON. Is it sufficient to just reset the registers in the block or should the
original state be restored? Perhaps only part of the state needs to be restored. These
and other questions need answers in order to produce a working design.
Joe Macri, CTO at Advanced Micro Devices, has said, “Platform power is as
important as core silicon power” [1]. For IC engineers, it is important to understand
that the chip power is only part of the story. At the end of the day, the chip has to go
into a package, and that package has to go onto a board. The chip probably also has
some software running on it. It is the power characteristics of the end product that
the consumer is going to experience. Placing all types of hooks for power manage-
ment into the IC is of little value if the software running on it doesn’t make good use
of those capabilities.
The diagram in Fig. 9.1 shows the interdependent relationship between system-
level design, chip design, package design, and PCB design. The Open SystemC
Initiative (OSCI) has released version 2.0 of its Transaction Level Modeling speci-
fication (referred to as TLM 2.0). Transaction level models are a way to let system
designers create high-level models in order to explore the design space and perform
power, performance, and area (PPA) analysis. The ability to separate the function-
ality from the power and timing allows system designers to focus on the models for
192 B. Pangrle
HLS
TLM Architectural Analysis
Power-Aware Models
System Level Verification
Power-Aware
Design
PCB Chip
Design Design
Package Test
PCB Power-Aware Test
Power Integrity
Design
Place & Route
Multi-Corner Multi-Mode
Fig. 9.1 Power impacts the whole system. TLM transaction level modeling, PCB printed circuit
board, HLS high-level synthesis
the desired behavior and then perform successive refinements to tie the power and
timing characteristics to specific implementations of that behavior. Using high-level
synthesis (HLS) tools is one way to generate register transfer level (RTL)
implementations from the behavioral models written in C or C++. These tools
appear promising in terms of providing a way to quickly generate multiple
implementations in order to search the design space for the best possible solutions.
Power-management techniques such as power-gating and dynamic voltage and
frequency scaling (DVFS) that turn off the power to portions of the design and
change the voltage level on the supply rails impose new requirements on the
verification process. Power-aware verification takes into account the fact that
logic cannot produce valid output unless it is powered ON. As previously men-
tioned, checks are also needed to insure that signals crossing power domains, i.e.,
portions of the chip that are running on different supply rails, have the appropriate
voltage levels to insure that good logical ‘0’ and ‘1’ values are communicated.
Likewise for test structures that create scan chains out of the registers for shifting in
test vectors and then shifting out the results, it is important to understand the power
domain structure on the chip to insure the test integrity.
Power management schemes that use multiple voltage levels increase the num-
ber of scenarios under which the design has to meet timing constraints. The number
of modes of operation for power-efficient chips significantly increases the com-
plexity of closing timing in a manner that the chip meets timing for all possible
modes of operations and across all of the process, voltage, and temperature corners.
Power integrity is an important issue at the board level. Many ICs use off-chip
voltage regulators placed on the printed circuit board (PCB). As the number of
voltage levels increases per chip, the board complexity also increases. Another
important factor from an integrity standpoint is the amount of current that needs to
be delivered. Since dynamic power is proportional to the supply voltage squared,
reducing supply voltages has been an important weapon in the designer’s arsenal
for fighting power. Back in the days of 0.5 mm CMOS when the supply voltage was
Trend
14.0
12.0
Requirement
10.0 Static Dynamic
8.0
(W)
6.0
4.0
2.0
0.0
2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024
Fig. 9.2 ITRS 2009 SoC consumer portable power trend
typically set at 5.0 V, a 25 W chip would require 5 A of current. Today, a 25 W chip

may be largely supplied by 1.0 V or less, thus requiring 25 A of current! For large-
scale microprocessors and graphics processors, these current levels can exceed
100 A. Coupled with the speed of signals and other noise sources, maintaining
PCB power integrity is challenging.
People often wonder if all of this additional effort for managing power is only a
temporary phase that the industry is going through and whether some technological
breakthrough will allow the return to a less complex design style. It seems inevitable
that eventually some technology will come along and supplant CMOS. Unfortu-
nately though, at the time of writing, it doesn’t look like it will happen very soon.
Figure 9.2 shows a chart from the 2009 International Technology Roadmap for
Semiconductors (ITRS) [2] indicating the expected power trend for system-on-chip
(SoC) consumer-portable ICs. The bars on the chart indicate static and dynamic
power. Static power is the amount of power that is used just to have the device turned
on without any useful work being performed. Typically in CMOS designs, this is due
to current “leaking” in the transistors used as switches. In an ideal switch, the device is
either “OFF” or “ON”. When the ideal switch is ON, current can flow freely without
any resistance and when it is OFF, no current flows at all. The real-world transistor
implementation differs from the ideal, and one of the differences is that, when the
transistor switch is OFF, some current can still get through the switch. As threshold
voltages have decreased and as gate insulators have become thinner, the switches
have become “leakier”. Dynamic power refers to the power used to perform useful
computation and work. For CMOS logic, this involves the transistor switches
turning OFF and ON to perform the intended logic computations. CMOS found
early success in low power applications, e.g., digital watches, because the vast
majority of the power was used only for the dynamic switching portion of the
logic computations. Ideally, as the transistor switches turn ON and OFF, current
194 B. Pangrle
flows to charge up a given capacitive load or discharge it, but does not flow on a
direct path from power to ground. This is accomplished by creating complementary
circuits that only allow either the path to power or the path to ground to be activated
for any given logic output, hence the name complementary metal oxide semicon-
ductor (CMOS). The dynamic power for CMOS circuitry can be approximated by
using a capacitive charge/discharge model, yielding the following common equa-
tion: Pdynamic afCV2, where a represents the switching activity factor. A major
point of importance here is the quadratic impact of voltage on dynamic power.
Clearly, the trend in Fig. 9.2 is heading in the wrong direction, although you
might detect brief periods where there are slight dips in the bars attributable to the
projected use of technologies such as fully depleted silicon on insulator (FDSOI) or
multi-gated structures such as Fin-FETs. While I won’t hazard to make predictions
10+ years out in an industry as dynamic as the semiconductor industry, it is pretty
clear that the expected trend going forward is for the power management problem to
increasingly become even more challenging.
The Chief Technology Officer at ARM, Mike Muller, has presented a number of
talks where he has mentioned “dark silicon” [3]. He points out that, by the year
2020, we could be using 11-nm process technology with 16 times the density of
45-nm technology and the capability of increasing clock frequencies by a factor of
2.4. Since dynamic power is proportional to clock frequency, roughly speaking we
would need to reduce the power on a per device basis by approximately a factor of
38 (16 2.4). Unfortunately, Mr. Muller also points out that the expectation is that
the 11-nm devices’ energy usage will only drop to a third of the 45-nm level and is
quoted as saying that this will leave engineers “with a power budget so pinched they
may be able to activate only 9% of those transistors” [3]. Clearly, there is a need for
more innovation in terms of getting the most capability out of the underlying
capacity of the process technologies.
9.3 When Power Does Not Scale
The electronics industry has a history of pushing technologies until the limits
from an energy-efficiency standpoint are reached, starting with vacuum tubes
through bipolar-transistor emitter-coupled logic (ECL), n-channel metal oxide
semiconductor (NMOS) to complementary metal oxide semiconductor (CMOS)
(as a side note, I find it interesting that the industry stayed with the term MOS
even after moving to polysilicon self-aligned gate processes and now, with the
advantages of high-k metal gate [HKMG] processes, MOS is truly MOS again). In
each case, the technologies were initially limited by the levels of integration. As the
technologies advanced though, the ability to create higher-density designs exceeded
the reduction in per-device energy usage. Eventually, power becomes a primary
driver from a design perspective. If you want to go faster, you need to find ways to
become more power efficient. All other things being equal, if one design uses less
power than another, then the clock frequency on the lower power design can be
Power Density Clock Frequency

1000 100000
Rocket Nozzle
Inflection Point Inflection Point
~30 GHZ
Nuclear Reactor
10000
100 P4D
Watts / cm2
C2D P4
P4 C2D Ci7
Pentium® ΙΙΙ
1000
MHz
Ci7 Pentium® ΙΙΙ
Pentium® ΙΙ P4D
Pentium® ΙΙ
10 Hot Plate Pentium® Pro
Pentium® Pro
100
Pentium®
i486 Pentium®
i486
1 i386 i386 10
1m 0.5 m 0.25 m 0.1 3 m 65 nm ’87 ’92 ’97 ’02 ’07 ’10
Fig. 9.3 Inflection point for power density and clock frequency
cranked up until the power matches the power of the higher-power design, thus
reaping the benefits of the higher clock rate. Power budgets limit the amount of
functionality that is integrated into a single chip. The more energy efficient a chip
is, the more functionality that can be integrated into it. This is very relevant in light
of Mike Muller’s comments about activating only 9% of the transistors.
In many ways, the impact of power on CMOS design has been like the impact of
driving along at 200 mph and hitting a brick wall, but that doesn’t mean that this
wasn’t foreseen. The charts in Fig. 9.3 are an extension of two charts shown by
Fred Pollack, an Intel Fellow, at Micro32 in 1999 [4]. His slides went up to the
Pentium III processor and I have added more recent processors to the charts to
illustrate what has happened since then. At the inflection point indicated in Fig. 9.3,
the flattening of the power density curves corresponds to an equal flattening in the
clock frequency curves. In 2002, then Intel CTO Pat Gelsinger predicted that by
2010 x86 processors would be running at 30 GHz clock frequencies [5]. In all
fairness to the former Intel CTO, it is pretty clear to see how he arrived at that
prediction. For roughly 15 years, processor designs were heading straight up the
line on that log chart. Continuing that line to 2010 implies a clock frequency of
~30 GHz.
The brick-wall incident appears to take place around 2004 and not coincidentally
at the 90 nm technology node. What happened? Going back to 0.5 mm again when
the supply voltage was typically set at 5.0 V, the threshold voltage for NMOS
transistors was in the neighborhood of 1.25 V. Leakage current for these devices
was relatively negligible in the overall scheme of energy consumption. As the
technology scaled down further to 0.35 mm, the supply voltage scaled down to
3.5 V, at 0.25 mm to 2.5 V, at 0.18 mm to 1.8 V. This scaling continued down to
roughly 1 V at 100 nm, and the 100 nm node is a pretty reasonable marker for where
the inflection point occurs. A general rule of thumb for performance reasons was to
set the threshold voltage at approximately one fourth of the supply voltage so at a
supply voltage of 1.0 V the threshold voltage would be somewhere in the vicinity of
0.25 V. This drop in threshold voltage from ~1.25 V at 0.5 mm to ~0.25 V at 100 nm
196 B. Pangrle
has a huge impact on the leakage current of the transistors. In short, in order to limit
subthreshold leakage current in the transistors, the threshold voltages have been
held relatively constant and therefore so have the supply voltages at around roughly
1.0 V. Remembering that the dynamic power is proportional to the voltage squared,
it becomes even more apparent why the inability to scale down the supply voltage at
the same rate as the feature size is so important to the amount of energy consumed
at successive technology nodes.
The graph on the right in Fig. 9.3 shows the sudden halt to increasing clock
speeds. The industry had become accustomed to the exponential increase, indicated
by the straight line on the log graph, in clock speeds over time. As was mentioned
previously, dynamic power is proportional to the clock frequency at a given voltage.
One trick that people who “overclock” processors use is to increase the supply
voltage along with adding exotic cooling to remove all of the thermally dissipated
energy. As was also mentioned earlier, dynamic power for CMOS circuits is
proportional to the supply voltage squared. In effect, if it is necessary to increase
the voltage in order to run at higher clock frequencies, then it can be said that the
power increases proportional to the cube of the clock frequency. This approximation
is at least often true for relevant operating regions, and you may see this relationship
mentioned in other literature. It is no small wonder then that, after the release of the
Pentium 4 570 J at 3.8 GHz in November 2004, there has been a long string of years
without the introduction of x86 processors at a higher clock frequency. In fact, the
clock rates for the subsequent generation of processors actually fell.
I/O
I/O
Mem
Clock Clock
MPU1 MPU2
Mem
Logic
Logic
I/O Clock Clock
Logic
ASP1 ASP2
Mem
Logic I/O Mem
Fig. 9.4 Power budgets for two general-purpose and two application-specific processors (With
kind permission from Springer Science+Business, # 2009 [6])
For processor designs, a significant portion of the energy consumed by the

circuitry is actually used for clocking. Figure 9.4 shows power-budget pie charts for
two general-purpose (microprocessor unit, MPU) processors and two application-
specific (ASP) processors. This diagram appears in Slide 1.30 on p.20 in [6]. Across
these four designs, the power budget for clocks is between approximately 20% to
almost 50% with the values being higher for the MPUs than for the ASPs. For
instance, a special-purpose processor designed for networking applications could
easily have a higher percentage of its power budget allocated for I/O in order to deal
with the transfer of large amounts of data. One thing that’s interesting to note about
all four of the processors shown is that the power budget for the clocks rivals the
power budget for the logic.
A number of issues become more challenging as clock frequencies are increased.
First, the constraints on skew and jitter are tighter, and this often leads to longer
insertion delays and longer buffer chains that consume more power. Second, as the
skew decreases, more registers are triggered in a narrower window of time causing
more activity to occur virtually simultaneously. Higher activity levels create a need
for more current to sustain the logic circuitry, and this can lead to noise and IR-drop
issues. Third, higher current levels can exacerbate local hot spots and together with
higher current densities can also adversely impact reliability. From a power stand-
point, it’s very important to pay attention to clocking schemes and to use efficient
clock-tree synthesis tools and methods in order to generate energy-efficient designs.
9.4 Advanced Technology Adoption Drives

Power-Aware Design
The halt in voltage scaling at about 100 nm creates significant challenges for
designers designing with sub-100 nm technologies. Initially, this impacted only
the design teams that were working on the very leading-edge technologies. As of
the time of this writing, 32 nm processors are available off-the-shelf, and designers
are working on 22-nm versions. The 22-nm node is four process generations beyond
90 nm. This implies that a lot of design teams are already designing at 90 nm and
below. Often the perception is that the adoption of new technology nodes has
slowed. This perception though depends upon the observer’s point-of-view. If
you take the viewpoint from silicon foundries, the demand for new technology
nodes is as strong as ever. TSMC reported for Q2 2011 that 64% of their revenue
was due to wafers at 90 nm and below and that 90 nm accounted for 9%, 65 nm for
29% and 40 nm for 26% [7]. From this data, revenue from the 40 nm node is on the
verge of eclipsing 65 nm and 65 nm is already falling from a peak of 32% of
revenue in Q1 2011. Since Q3 2009, the cumulative 90 nm and under percentage
has gone from 53% to 64% of wafer revenue while total revenue has also increased
significantly. Clearly from TSMC’s viewpoint, there’s been no slow down in the
demand for new technology nodes. If the data is looked at from the perspective of
198 B. Pangrle
new design starts, then a different picture is formed. Design starts in newer
technology nodes appear to be decreasing. Certainly, the costs for designing in
newer nodes are increasing, and this puts pressure on the number of projects that are
economically viable at these nodes, so what we are really experiencing are fewer
designs with higher volumes moving first to the new nodes. Another way that
designers can make use of newer technology nodes, is to design for Field-Program-
mable Gate Arrays (FPGAs) that are implemented in these newer technology nodes.
To the foundries though, an FPGA is one design. Intuitively, this would all seem to
make sense, and it looks similar to other industries as they mature and start to
consolidate.
The added complexity of elaborate power-management schemes is part of that
additional cost of designing in more advanced technology nodes. One of the real
changes that have occurred is how prevalent power issues are across the design
spectrum once the 100 nm barrier is crossed. Originally, power was primarily a
concern for designers that were designing chips for portable and battery operated
devices. Before 130 nm, most other design teams were primarily concerned about
meeting timing (performance) and area constraints. Power was pretty far down the
list and was often left as an optimization exercise for subsequent versions of the
design. In other words, the priorities were get the chip out meeting timing and area
and then worry about the power later. This approach hit a number of teams very
hard at the inflection point shown in Fig. 9.3. Chips have to go into packages and
onto boards. As mentioned early, power is a platform issue and not just a chip issue.
If a board has a 25 W power budget and a chip comes in at say 22 W, unless that
chip is the only chip going onto the board, that’s probably a problem. If a chip
misses the maximum power limits for its intended package and the increased cost of
the next available package that can accommodate the additional thermal power
dissipation makes the combined part cost too expensive for its target market, that’s
a problem too. Chip area and yield fade a bit in importance when the cost of the
packaging becomes significantly higher than the cost of the silicon. If the only way
to get into a lower-cost package is to reduce the power consumption of the chip,
suddenly power shoots up the list of design priorities. This became a concern for all
types of designs and not just portable, battery operated applications.
The change in priorities has caused designers to re-think a number of strategies.
Back in 1996 when Power Compiler ® was released, it had a feature for inserting
clock-gating into designs. The prevailing thinking at that time was that adding any
logic onto the clock tree was generally a “bad” idea. As might be guessed from the
pie charts shown in Fig. 9.4, clock-gating is now a generally accepted design
practice and can be a powerful technique for reducing dynamic power. Not using
clock-gating often means that the design team is missing out on significant power
savings. In response to increased leakage power that typically started to appear
around 130 nm, foundries started to offer multiple cell libraries at the same
technology node where each library corresponded to cells that were built using a
threshold set for that library. As an example, one library might use transistors with a
0.35 V threshold and the other 0.25 V. The cells in the higher threshold voltage
library leak less than the cells in the lower threshold library but at the expense of
having slower timing characteristics. Basically, the idea is to use as many high-
threshold cells as possible and only insert the faster, leakier low-threshold cells
where necessary in order to meet timing. There could easily be a factor of 30 in the
leakage between the two libraries implying that, if only 3% of the cells needed to be
low-threshold cells, the overall leakage would still be roughly twice as high as a
design only using high-threshold cells. Nonetheless, twice as high is still much
better than 30 times as high. The technique could allow designers to meet timing
and reduce leakage by an order of magnitude compared to using only low-threshold
cells. High-threshold libraries are often referred to as “low-power” libraries. One
needs to be somewhat careful about reading too much into that though. The high-
threshold libraries are low power from a leakage standpoint but, from a perfor-
mance and energy-efficiency standpoint, lower-threshold libraries will have better
power characteristics for some designs. Remembering the “overclockers” trick of
raising the voltage to get higher clock speeds, if in using a higher-threshold library
it is necessary to raise the supply voltage to meet the same timing that could be
achieved with a lower voltage using a low-threshold library, the lower-threshold
library may win in terms of overall power consumption.
In response to the technological challenges of increased static and dynamic
power in the smaller technology nodes, designers have turned to techniques for
varying the supply voltages. One effective way to reduce static power is to turn off
circuitry when it’s not in use. This is often referred to as “power gating”. At first, it
may sound strange that there would be much circuitry not being used but for
applications like mobile phones, most of the time, most of the designed functional-
ity is not being used. For example, advanced mobile phones have features for
making phone calls, playing music, watching videos plus numerous other personal
computer-like applications. In reality though, the phone will likely spend most of its
time in a mode that occasionally checks for incoming phone calls and perhaps
nothing more. By cutting the power off to all of the circuitry that is not currently in
use, designers have dramatically increased the standby lifetimes for these devices.
At a finer level of granularity, even being able to turn off circuitry for time periods
on the order of milliseconds can have a beneficial impact on the reduction of static
power. For reducing dynamic power, the quadratic impact of voltage is the key. The
ability to tune the supply voltage to match the necessary performance level enables
significant potential dynamic-power savings. Portions of the design that use
dynamic frequency scaling (DFS), i.e., are designed to run at different clock
frequencies, are also good potential candidates for dynamic voltage scaling
(DVS). By pairing up each of the operating frequencies with its own operating
voltage, each frequency mode runs at a more efficient power level. The combina-
tion of the two is often referred to as dynamic voltage and frequency scaling
(DVFS). Traditionally, power and ground had been assumed to be non-varying
constants for most CMOS logic designs. The use of these power-gating and
dynamic voltage-scaling schemes though makes those assumptions no longer
valid and add additional complexities to the design and verification process.
Another technique for reducing static power is to dynamically change the
threshold voltages of the transistors. Figure 9.5 includes a graph showing how the
200 B. Pangrle
Fig. 9.5 Diminishing impact of body bias on the threshold voltage (With kind permission from
Springer Science+Business, # 2009 [6])
threshold voltage varies with changes in the substrate bias along with a simple
NMOS transistor diagram. This graph appears in Slide 2.13 on p.31 in [6]. MOS
transistors are often thought about as three-terminal devices: source, gate, and
drain, but the substrate is a fourth terminal that can also be controlled. Some
designers have made use of this fourth terminal to adjust threshold and therefore
correspondingly the leakage levels of transistors in their designs. It has the advan-
tage that the savings can be had without modifying the power-supply lines for Vdd
or ground lines in the design. Unfortunately, as the graph in Fig. 9.5 shows, this
technique becomes less sensitive to changes in the substrate bias as the transistor
sizes continue to shrink.
9.5 Multi-core Processors
Microprocessor clock frequencies have hit a wall but the performance capabilities
have not. Designers continue to invent new ways to improve performance even in
the face of hard limits on clock frequencies. The biggest factor in performance
improvements since the halt in clock frequencies has been the incorporation of
multiple cores onto the same piece of silicon. The thinking goes that if the
performance can’t be doubled by doubling the clock frequency then perhaps
doubling the number of processor cores can do the trick. Figure 9.6 shows a
progression in AMD Opteron designs since 2003, when the first single-core
Opteron was released. I find it striking that we see the first dual-core processor at
90 nm just as the 100 nm threshold was crossed. Each successive node has led to
more cores on a single piece of silicon, and it appears that this trend will continue
for the foreseeable future with major processor manufacturers having already
announced plans for designs with many more cores.
Fig. 9.6 Four generations of AMD Opteron processors (Photos courtesy of AMD)
Each core presents a unique opportunity from a power-savings standpoint. It is

its own unit that can be individually controlled depending upon the processing
needs at any given moment. From a logical viewpoint, each core can operate in a
separate domain with its own supply voltage and clock frequency, which actually
translates reasonably well into a physical implementation. In the server market,
where some of these processors compete, energy efficiency is an important aspect
of the design. Corporations with large server farms, such as Google or Amazon,
have to provision these facilities to handle peak-demand loads so that operations
always run smoothly. These facilities may only experience peak loads for a
relatively small portion of the time that they are in service and the power bills
from running these facilities are a significant expense. Companies do not want to
have to pay for peak-load power in off-peak hours. In order for server CPUs to be
competitive, these CPUs need to scale back on energy usage, especially in nonpeak
demand modes. Turning cores OFF as they are not needed is one way to accomplish
this, as well as using DVFS schemes to reduce the power of selected cores that do
not need to be running at top speed. It is interesting to note that both Intel and AMD
have introduced capabilities to “overclock” individual processor cores, called
Turbo Boost and Turbo Core respectively, when all of the cores on the chip are
not simultaneously running. Since the part is designed for a specific thermal
202 B. Pangrle
capability, if all of the cores are not running then there is some extra headroom for
the cores that are. This extra margin can be utilized by bumping up the voltage and
frequency of the running cores to improve their performance. These types of
features can make benchmarking and comparisons of different parts quite challeng-
ing but definitely provide additional performance benefits for customers.
Multiple cores increase the level of design complexity. The number of different
operating states or modes increases exponentially with the number of cores, if the
cores are truly independent of each other in terms of power-up and -down
conditions and their voltage and frequency settings. Even within a single core, it
is possible to have sub-blocks that operate in different power and frequency
domains. From an implementation standpoint, the challenge now becomes meeting
all of the constraints across multiple modes simultaneously. Typically, each mode
of operation has to be designed to function correctly across a range of process,
voltage, and temperature “corners” to accommodate for variations in processing,
voltage drop, and temperatures. Figure 9.7 [8] shows an example of a cell phone
SoC that has more than 21 modes and corners that need to simultaneously meet
design constraints.
Figure 9.8 depicts a diagram of a simple design with a core-power domain along
with two other “island” domains that operate independently. The table in Fig. 9.8
illustrates how quickly the possible mode–corner combinations grow as more
islands are added to the design. This has created a need in the marketplace for
implementation tools that can efficiently design and evaluate across multiple modes
and corners in order to close timing and other constraints.
Cell Phone Chip Example
–Operation Modes
• SLEEP (10Khz)
• STANDBY (10Mhz)
• ACTIVE (100Mhz)
–Corner Specification
• PVT corners
– +100C, 1.2 V
– – 40C, 0.8V
• Interconnect corners
–MaxR, OCV 5%
–MaxC, OCV 12%
–MinR, OCV 3.3%
–MinC, OCV 3.3%
–MaxRC, OCV 5%
More than 21 mode/corner scenarios
Fig. 9.7 Multi-corner and multi-mode design example

Fig. 9.8 Example illustrating the growth in modes and corners with design complexity
9.6 Electronic Design Automation Waves
At any given point in time, leading-edge CMOS technologies hold the promise of
certain underlying capabilities that can only begin to be approached by the available
tools and methodologies. When changes from one technology node to the next are
relatively minor, the gap between available capability and serviceable capability
tends to shrink, and the opposite happens when the changes are larger. The gaps that
open when there are large changes, like those created for power management,
present opportunities in the marketplace for new tools and methodologies to appear
and once again close those gaps. For example, in the place-and route-market, gaps
opened and were filled around moves towards SoC and timing-driven design with
the latest now appearing around multi-corner, multi-mode (MCMM) designs.
204 B. Pangrle
The next big wave appears to be shaping up around design implementation. For
many years, designers used schematic capture as a method for creating logic
designs. Starting around the early 1990s, design teams started to transition from
schematic capture to text-based design creation using hardware description
languages (HDLs) such as VHDL and Verilog. This need for a more efficient
design methodology to keep pace with the increasing capability of the underlying
CMOS technology, often referred to as “Moore’s law”, created an opportunity for
the successful entry of new products and methodologies into the marketplace. If, for
arguments sake, we say that this transition to HDL-based design happened at
0.5 mm, then at 32 nm there have been seven generations of process-technology
nodes as defined by the ITRS and a corresponding increase in capabilities of over
two orders of magnitude. Using another rule of thumb, that the industry increases
the level of integration roughly tenfold every 6–7 years, then you also reach the
same conclusion. Pundits have been claiming for a number of years that the industry
is on the verge of moving towards electronic system level (ESL) design and high-
level synthesis (HLS) tools, and it appears that the pressure for higher designer
efficiency to keep pace with the higher levels of available integration are making
that happen.
The move to higher-level design tools and methodologies is also very relevant
from a power-aware design and verification standpoint. As in typical design pro-
cesses, the earliest decisions have the greatest impact on the design. At each
successive stage of design refinement, options are eliminated as the design
crystallizes towards its final implementation. This is also true for power optimization.
The table in Fig. 9.9 shows multiple implementations for an example inverse
fast-Fourier-transform (IFFT) design and how the power, performance, and area
change with each variation of the implementation [9]. This work was performed by
a team at MIT using a high-level synthesis tool to quickly generate multiple designs
for evaluation. There are two important concepts to note here: (1) the range of
variation in power across the designs is almost an order of magnitude and (2) most
Transmitter Design Minimum Frequency to Area Average Power

(IFFT Block) Achieve Required Rate (mm2) (mW)
Comb (48 bfly4s) 1.0 MHz 4.91 3.99 3.99
Piped (48 bfly4s) 1.0 MHz 5.25 4.92

Folded (16 bfly4s) 1.0 MHz 3.97 7.27
Folded (8 bfly4s) 1.5 MHz 3.69 10.90 ~8.6x
Folded (4 bfly4s) 3.0 MHz 2.45 14.40
Folded (2 bfly4s ) 6.0 MHz 1.84 21.10

Folded (1 bfly4) 12.0 MHz 1.52 34.60 34.6
Fig. 9.9 IFFT example design showing power and area trade-off for constant throughput
design teams would not have the time and/or resources to evaluate this many
implementations, if they had to code them all into a RTL HDL description by
hand. The results back many industry expectations that high-level tradeoffs can lead
to an 80% reduction in power. Further downstream optimizations account for a
much smaller percentage.
The reproduction of a presentation slide shown in Fig. 9.10 [10] is from a
keynote talk that Bill Dally, Sr. VP of Research and Chief Scientist at nVidia,
gave at the 46th Design Automation Conference in July of 2009. During his talk,
Dr. Dally stressed the need for tools for exploring power and performance tradeoffs
at an architectural level. With the gap widening between the capabilities of the
underlying silicon and designers’ ability to fully make use of it, the need for better
tools becomes increasingly important. Figure 9.11 is a chart from the 2009 ITRS
Roadmap [2]. It predicts that in order for designers to deal with the graph shown
back in Fig. 9.2 tools that enable that early architectural 80% power savings are
going to be needed, and shows that by 2015 80% of the power optimization will
need to happen at the ESL and HLS portion of the design. Even at the foundry level,
the need for these high-level tools and flows is seen as critical. TSMC included
ESL/HLS tools for the first time in 2010 as part of their reference flow in Reference
Flow 11. Between the need for more productivity per designer to keep pace with the
Power
Power tools for architecture exploration

Rapid evaluation of the power and
performance impact of architecture tradeoffs
End-of-flow tools are not sufficient
Low-power X
Storage arrays, interconnect, etc...
Fig. 9.10 Bill Dally’s slide Custom design for power, not performance
Optimized Vdd, Vt
on the need for high-level
exploration tools [10]
Fig. 9.11 ITRS 2009: growing importance of optimizing for power earlier in the design process
206 B. Pangrle
increased levels of integration and the need for better architectural analysis for
power optimization, a near perfect-storm-like scenario is created for high-level
tools to finally catch on in the marketplace.
9.7 Power-Aware Verification
Power-management schemes vary the Vdd supply-rail levels for portions of the chip,
and this introduces new complexities for verification. At the register transfer level
(RTL), many signals are represented by logical ‘1’ and ‘0’ values. The validity of
those values depends on the underlying physical voltage levels. For example, let us
assume that in order for the circuitry to recognize an input as a ‘1’ or ‘0’, the voltage
level must be greater than or equal to 75% of Vdd or less than or equal to 25% of Vdd
with the remaining 50% in the middle being an “unknown” area. So what happens if
the core logic of the chip is running at 1.25 V and an island is put into a lower power
mode and is running at 0.8 V? An output signal from the 0.8 V block at a logical ‘1’
level will be at best at 0.8 V and probably slightly lower. If that signal is then fed as
an input into the core running at 1.25 V there is a high probability of the signal not
being recognized as a logical ‘1’ value on the input. Even worse, if the signal is fed
into a large buffer or inverter, it may cause both the PMOS and NMOS transistors to
at least partially be “ON” simultaneously, creating a path for current to flow from
Vdd to GND. Remember, earlier we mentioned that, ideally in CMOS designs, only
the path from Vdd to the output or the output to GND is on at a given time. Because
of the physical nature of the transistors implementing the logic, there may be brief
periods of time where transistors on both sides are simultaneously conducting
current across the channel and creating a path from Vdd to GND. This current is
often called “short-circuit” or “crowbar” current and is typically lumped into the
dynamic power equation since it is a function of the logic changing value. It is
definitely not a desired condition. Back to our input signal, since it is at best at
0.8 V, it falls substantially below the threshold of 75% of 1.25 V or approximately
0.94 V, and won’t properly be recognized as a ‘1’.
High-level simulators use logical representations of signals and not the actual
physical voltage levels in order to increase their speed and design capacity. They
work extremely well as long as the underlying assumption that logical ‘1’ and ‘0’
values have the proper physical voltage values associated with them. When two
blocks communicating in a design run at different voltage levels, it is important that
the simulator understands how these differences affect the communication between
blocks. For example, suppose a hypothetical PC notebook designer puts together a
system such that when the notebook computer is closed, it goes into a low power
sleep mode in order to save energy and increase the run lifetime of the battery.
As people will often take closed notebook computers and stick them into backpacks
or briefcases with poor air circulation, the designer builds in a thermal detector to
indicate if the sleeping notebook is getting too hot and should power itself down to
avoid overheating. The thermal detection circuitry raises a ‘1’ signal if it determines
that the notebook is too warm. This warning signal then travels across the chip until
it arrives at a part of the circuitry that is still awake to recognize it and power down
the machine. What happens if the signal happens to pass through a block that has
been placed into a lower-voltage mode to save power on the way to its final
destination? Well, if the simulator only recognizes ‘1’ and ‘0’ and makes no
distinction for signals crossing different voltage boundaries, everything in the
simulation looks fine. Each block receives a ‘1’ and passes it along to the next
block until the ‘1’ shows up at the final destination. What happens in the real
physical world? The voltage on the signal drops below a recognizable ‘1’ value and
is never detected by the final destination, and so the notebook keeps running,
possibly catastrophically.
Commercial simulators now incorporate power-aware capabilities to better
reflect the physical realities. Designers use a format such as the IEEE Std
1801–2009 (also known as the Unified Power Format UPF 2.0) [11] to describe
the power domains or voltage islands on the chip so that the simulator can
determine where the block boundaries are and when signals are crossing them.
The format has a wide range of capabilities and is architected to allow designers to
overlay their power-management scheme on top of their functional descriptions.
This allows the verification of the functionality of a block to occur independently of
the power management, as is traditionally done, and then verify the correctness
with the power management included. This enables better reuse of the design
components since the power management does not have to be hard coded into the
functional behavior of all of the components, and a block can then be reused in
another design with a different power scheme by altering the UPF description.
Figure 9.12 shows a simulation waveform diagram for an example design going
through a simulated power-down and then power-up sequence. It illustrates a
Fig. 9.12 Example waveform diagram for a simple power down/up sequence
208 B. Pangrle
number of conditions that a simulator has to handle in order to properly represent

the design’s behavior. In a typical power-down sequence, a number of actions need
to occur before the power is shut off to the block. First, it is necessary to determine
whether the state of the block needs to be saved. If it does, one popular way to save
the state is to use so-called retention registers that have a separate low-leakage latch
that stays powered ON when the rest of the block is powered OFF. There are
numerous ways to implement retention registers. Some use separate “save” and
“restore” signals, others share one signal for this capability. Some use the clock to
save and restore values, others do not. For an RTL simulator to properly model this
behavior, a description of the intended retention register’s behavior must be sup-
plied. In the example shown in Fig. 9.12, the first step in the process is to quiet the
clocks by setting the clock enable signal low. The next step is to save the value of
the registers and then isolate the output signals. Once these steps are completed, the
voltage to the block can be turned off by setting the Vdd enable signal low. At this
point, the Vdd-on signal goes low to indicate that the block is powered OFF. The
block can now stay in this powered OFF state until it is ready to be powered ON
again. Some items to note here are: the proper sequencing of the signals, the
registered output values being saved and properly isolated (in this case to ‘0’),
and the internal nets of the powered-down block going to corrupted “unknown”
values. If any of the internal nets are somehow escaping and driving powered logic
gates, the “unknown” values should propagate through the simulation and raise a
red flag, indicating that something is wrong. Assertions could also be added to the
simulation to provide additional checks for these conditions. To bring the block
back ON, the process is reversed and Vdd-enable is set high to allow the voltage to
again be supplied to the block. Once the voltage level has reached its proper
operating point, Vdd-on can go high, indicating that the block is once again fully
powered ON. The saved register values can now be restored and the isolation
removed from the outputs of the block. The block is now back up, properly powered
and in a good state so the clock-enable signal can go high, and useful activity for the
block can begin again. The simulation output shows the restored values appearing
in the registers, the clock starting and the isolation being removed from the outputs.
9.8 Testing
There are two main issues to address with regards to power for test: (1) creating
proper test structures for the given power-management scheme (or test for low-
power design) and (2) controlling the power while the chip is under test (or reducing
power during test).
Automatic test-generation tools need to understand the power-architecture
implementation much in the same way as the verification tools do, i.e., detect
where the boundaries of different voltage regions are as well as how the voltages
may vary and even be turned off for each region. Again, a standardized format, such
as the IEEE Std 1801–2009, is useful for providing this kind of design information.
Given this description of the power architecture, the tools can make intelligent
decisions about placing registers into scan chains used to shift in test vectors and
shift out test results. In general, it is beneficial to keep scan chains contained in one
domain. This eliminates concerns about level-shifter insertion on the chains or
chains possibly not functioning correctly because of one domain being powered
down during the test. In order for a scan chain to function correctly, it is necessary
for power to be supplied to the registers in the chain.
The additional cells that are inserted for power management, such as power
switches, retention registers, isolation cells, and level shifters require new tests in
order to demonstrate that the devices are working properly. Retention registers not
only need to function properly as a register when their power domain is powered
ON but also must be able to recover a retained value correctly after going through a
save, power OFF, power ON, and restore sequence. Isolation cells need to properly
transmit input values when not in isolation mode and then, when in isolation mode,
output the proper isolation value. Level shifters need to translate input signals to the
appropriate output-voltage level so that the correct logic value is transmitted across
the domain boundaries.
Managing power under test is another crucial step. For many designs, it is not
uncommon for the switching-activity factor of the register to fall within the range of
15–20% per clock cycle. Designers use these activity levels to estimate the current
necessary for the circuit to function properly and then size the power supply rails
accordingly. Testers are expensive pieces of equipment, and the longer a chip has to
sit on a tester for testing, the more expensive it is to test that chip. Tools that
generate test patterns therefore try to reduce the number of patterns or vectors
necessary to test the chip, and this implies increasing activity levels to try to test as
much of the chip with each vector as possible. Increased activity levels imply
higher current levels and higher power levels. The expectation for shifting in a
random string of ‘1’s and ‘0’s into a scan chain is that about 50% of the registers
will change value on each clock cycle. This is significantly higher than the typical
15–20% level the circuitry may see in normal operation. If the power supply rails
cannot handle the extra current, the voltage may fall due to what is known as
IR-drop. That is, the resistive nature of the supply lines along with the increased
current will cause the voltage to fall along the power rail. If the voltage falls too far,
then it is possible that the circuitry won’t pass on valid ‘1’ values. In situations like
this, it is possible to have a chip that would operate perfectly in its intended
environment but will fail on the tester. The end-effect is that “good” chips fail
test and are thrown away. This is definitely not a desired outcome.
One way to reduce power under test is to not just randomly assign bit values to
portions of the vector that do not impact the test. These bits are often referred to as
“don’t cares”. Instead, if the filler values used for the don’t care bits are set to
produce long chains of the same value then when these bits are shifted into the
registers the switching activity will be reduced. Registers will see long strings of the
same value cycle after cycle and not have to switch as frequently. The tests and
generated vectors can be optimized to reduce switching activity during load,
capture, and unload phases of the testing.
210 B. Pangrle
Complex SoCs using power management operate in many different modes.

Chips intended for smart mobile phones have functionality that is mutually exclu-
sive, for instance playing an mp3 song and handling a phone call. The majority of
the time, most of the functionality is OFF or in a low-power standby state waiting
for an event to wake it up. When testing, it is important to test the chip in actual
modes intended for operation and not to put it into modes that the device should
never see in the final product, especially if those modes pull too much current.
When the current levels are too high, as previously mentioned, a good chip could
fail on the tester. That’s bad. It is even worse though, if the excess power draw is so
high that the elevated temperature actually causes the chip to fail in a destructive
manner. In extreme cases, chips have been known to actually catch fire while on the
tester. Power-management schemes have raised the bar for test-generation tools,
and proper forethought is necessary in terms of partitioning the design and the tests
to insure that all of the necessary functionality is fully covered in a manner that fits
within the design constraints.
9.9 Packaging
Advances in packaging are promising to further raise the level of integration for
electronic devices. Often thrown in the “more than Moore” category, the implica-
tion is that packaging allows additional levels of integration beyond the scaling that
is occurring in silicon. A popular concept for packaging is the stacking of die
commonly referred to as 3D packaging. It offers several potential benefits, amongst
them being a reduction in the size of the delivered functionality. Two or more chips
in one package uses less volume than two or more packages. Multi-chip solutions
also enable better control of yields. Functionality across multiple chips is paired
with known-good tested chips with less area than a corresponding one-chip solution
containing all of the functionality. Semiconductor processes can also be better
tailored for specific functionality, for example: logic, memory, and analog
capabilities. Placing all of the functionality into one package reduces interconnec-
tion parasitics compared to a multi-package solution. Bernard Meyerson, VP/
Fellow at IBM, has said that, “about half of the dissipation in microprocessors
comes from communication with external memory chips. If the chips are stacked
together in 3D, communication energy cost might drop to a tenth.” [12] That is a
significant saving in power. As an example, if a microprocessor is using 100 W and
half (or 50 W) is dissipated in memory communication then a 3D packaging scheme
could potentially reduce the overall power to 55 W. That is an overall reduction of
45% and is huge in terms of the markets that these types of parts are targeted for.
Beyond the potential power savings are the thermal issues that now have to be
modeled in order to insure that hotspots on one chip do not interfere with the proper
operation of adjacent chips in the package. Designers are working with through-
silicon vias (TSVs), edge-connect, and wire-bonding technologies to implement
stacked package designs. Proper thermal modeling is important and parts that don’t
need large heat-sinks or fans are often more desirable from an overall packaging
cost as well as consumer preference standpoint.
9.10 Boards
Power integrity for PCB designs, as well as thermal analysis, are increasingly more
important. The use of multiple supply voltages on the same chip for power
reduction also increases the complexity of the board design. Often the voltage
regulators for the supply voltages are “off-chip” and placed onto the circuit board.
The use of lower supply voltages also complicates matters. Back when 5 V for Vdd
was common, if a chip operated at 25 W, then that implied 5 A of current. With
chips now running at 1 V for Vdd or lower, the same 25 W power budget implies
25 A or more of current. Large processors can easily operate at over 100 W,
implying the need to supply over 100 A of current. As the voltages are lowered
so are the noise tolerances. More current means a larger impact due to IR-drop
effects. Communications processors that have a large portion of their power
budgets targeted for I/O, like the one diagrammed in Fig. 9.4 for example, are
sensitive to changes in the activity levels of the I/O pins. A sudden increase in
activity may require an additional 25 A of current in a matter of clock cycles. Large
changes in current are limited by inductance in the current-supply paths and voltage
levels drop based on L(di/dt) effects.
Figure 9.13 shows a map for an example PCB design. High-frequency serial-
communications links to reduce both power and pin counts on chips are more
commonly used now, and they are also sensitive to the board layout and routing.
Designers need tools to analyze signal integrity for the routes on the board and to
make sure that the power-distribution networks (PDNs) are robust to the effects of
changes in current demands and noise sources.
• Integrated Circuit Power

Trends:
– Lower & multiple voltages/ IC
– Higher currents
– Lower voltage supply tolerances
• Printed Circuit Board (PCB)

Power Distribution Networks
(PDNs) are more complex
– Multiple PDNs on a single PCB
– Requires “jigsaw” of split power /
ground planes
– Over-conservative design increases
cost
Fig. 9.13 Example PCB distribution-network analysis results

212 B. Pangrle
Energy used in the circuitry is mostly dissipated as heat. Chips are designed to
operate within specified temperature ranges based on their target environment.
Requirements for consumer products are typically less stringent than for aircraft,
military, or space applications but often are still quite challenging. For example, on
the main PCB for a personal computer (often referred to as the mainboard or
motherboard), there is a priority ranking for the chips going onto the board.
Typically at the top of the list is the processor chip. This is in order to provide
enough cooling for a high-power-consumption part. Often, other chips on the board
then have to deal with the impact of the heat generated by the processor as well as
other components that go into the case for the machine. With leakage power
significantly increasing at 130 nm and below for current process technologies,
keeping chips cooler is even more important. Leakage power increases exponen-
tially with temperature, so as a chip uses more power and produces more heat, if
that heat isn’t carried away, it will cause the chip to use even more power and run
even hotter. Left to itself, this leads to an unsatisfactory outcome. An important part
of designing PCBs then is performing thermal analysis to determine the expected
temperature ranges across the board so that each chip is guaranteed to operate
within its specified thermal environment.
9.11 Summary
This chapter has presented an overview of power’s impact on the design process
starting from the initial system definition in terms of its architecture all the way
through the implementation process to produce a complete design.
At the system level, architectural analysis and transaction-level modeling were
discussed, and it was shown how architectural decisions can dramatically reduce
the design power and the importance of modeling hardware and software together.
At the chip level, creating on-chip power domains for selectively turning power
off and/or multi-voltage operation has an impact on the following areas: (1) chip
verification, (2) multi-corner multi-mode analysis during placement and routing of
logic cells, and (3) changes to design-for-test, all in order to accommodate power-
gating and multi-voltage control logic, retention registers, isolation cells, and level
shifters needed to implement these power-saving techniques.
At the process level: the disappearing impact of body-bias techniques on leakage
control and why new approaches such as HKMG technology help but do not
eliminate power issues that were covered in this chapter.
Power-efficient design is impacting the way chip designers work today, and this
chapter focused on where the most significant gains can be realized and why power-
efficiency requirements will continue to challenge designers into the future. Despite
new process technologies, the future will continue to rely on innovative design
approaches.
In the end, it will come down to economics more than technology. It won’t be the
technology that is the limiting factor but the cost of implementation.
Acknowledgments I thank Erich Marschner, Keith Gover, Chuck Seeley, TJ Boer, Don
Kurelich, Arvind Narayanan, Jamie Metcalfe, Steve Patreras, Gwyn Sauceda, Shabtay Matalon,
and Jon McDonald for their many discussions and help in compiling this information.
References
1. Magee, M.: AMD says fusion will knock Intel for six. Or five. TG Daily, 11 Nov 2009. http://
www.tgdaily.com/hardware-features/44609-amd-says-fusion-will-knock-intel-for-six-or-five
(2009)
2. International technology roadmap for semiconductors (ITRS). www.ITRS.net/reports.html
(2009) Edition
3. Merrit, R.: Power surge could create ‘dark silicon’, EETimes, 22 Oct 2009. http://www.
eetimes.com/news/semi/showArticle.jhtml?articleID¼220900080 (2009)
4. Pollack, F.: Keynote address: new microarchitecture challenges in the coming generations of
CMOS process technologies. MICRO-32. In: Proceedings of the 32nd Annual IEEE/ACM
International Symposium on Microarchitecture, Haifa, 16 Nov 1999
5. Ross, A., Wood, J.: Spotlight processors: the next little thing. Comput. Power User 2(6),
58 (2002)
6. Rabaey, J.: Low Power Design Essentials. Springer, New York (2009)
7. TSMC Q2: Quarterly Report (2011)
8. Rhines, W.C.: Keynote address: common wisdom versus reality in the electronics industry. In:
DesignCon 2009, Santa Clara, 3 Feb 2009
9. Dave, N., Pellauer, M., Gerding, S., Arvind: 802.11a transmitter: a case study in microarch-
itectural exploration. In: Proceedings of the 4th ACM/IEEE International Conference on
Formal Methods and Models for Co-design, MEMOCODE’06 (2006), Napa Valley,
pp. 59–68. DOI: 10.1109/MEMCOD.2006.1695901
10. Dally, W.J.: Keynote address: the end of denial architecture and the rise of throughput
computing. In: 46th IEEE/ACM Design Automation Conference, Anaheim, 29 July 2009,
p. xv
11. IEEE standard for design and verification of low power integrated circuits, (a.k.a. unified
power format 2.0), IEEE Standard 1801™-2009, 27 March 2009
12. Oishi, M.: 3D technology drives semiconductor evolution. Nikkei Electronics Asia, 4 June
2009. http://techon.nikkeibp.co.jp/article/HONSHI/20090527/170863/
Chapter 10
Superprocessors and Supercomputers
Peter Hans Roth, Christian Jacobi, and Kai Weber
Abstract In this article, we describe current state-of-the art processor designs, the
design challenges faced by technology, and design scaling slow-down, problems
with the new design paradigms and potential solutions as well as longer-term trends
and requirements for future processors and systems.
With technology and design scaling slowing down, the processor industry rapidly
moved from high-frequency designs to multi-core chips in order to keep delivering
the traditionally expected performance improvements. However, this rapid paradigm
change created a whole new set of problems for the efficient usage of these multi-core
designs in large-scale systems. Systems need to satisfy an increasing demand in
throughput computing while at the same time still growing single-thread performance
significantly. The increase in processor cores poses severe challenges to operating
system and application development in order to exploit the available parallelism.
It also requires new programming models (e.g. OpenCL*). Furthermore, commercial
server systems are more and more enriched with special-purpose processors because
these specialty engines are able to deliver more performance within the same power
envelope than general-purpose microprocessors for certain applications.
We are convinced that future processors and systems need to be designed with
tight collaboration between the hardware and software community to ensure the
best possible exploitation of physical resources. In the post-exponential growth era,
hardware designs need to heavily invest in programmability features in addition to
the traditional performance improvements.
P.H. Roth (*) • C. Jacobi

IBM Systems & Technology Group, Technology Development, Schoenaicher Strasse 220, 71032
Boeblingen, Germany
e-mail: peharo@de.ibm.com; cjacobi@de.ibm.com
K. Weber
System z Core Verification Lead, IBM Systems & Technology Group, Technology Development,
Schoenaicher Strasse 220, 71032 Boeblingen, Germany
e-mail: Kai.Weber@de.ibm.com

216 P.H. Roth et al.
10.1 Requirements in Computing, Business Intelligence,

and Rich Media
Over recent decades, the world has seen computers penetrate almost every aspect of
everyday life, and this trend is likely to continue. The rise of the internet, and more
recently smart phones, resulted in an exponential increase of data. At the same time,
companies and other organizations are striving to use the available data more
intelligently in order to reduce operational risk and increase revenue and profit.
The field of “business intelligence” or “business analytics” is becoming more and
more important and is driving a change in the requirements on computer systems.
Traditionally, data has been analyzed using database queries to generate reports.
Business intelligence is using complex mathematical models to transform these
descriptive reporting techniques into predictive and prescriptive models that allow
forecasting of customer behavior or show optimization potential for business
processes (Fig. 10.1).
In addition, the growth of bandwidth for internet connections has allowed users
to create and share content types such as music and videos more easily. However,
in order to help find interesting content, these music and video files need to be
analyzed and tagged for search ability and converted to various formats for acces-
sibility. This process is far more compute intensive than traditional text-based
analysis and drives additional compute and bandwidth requirements into computer
systems.
Fig. 10.1 Business analytics landscape [1]

10 Superprocessors and Supercomputers 217
10.1.1 State-of-the-Art of Multi-core Processors
Microprocessor design has come a long way in the last 20 years. The first POWER*
processor POWER1*, for example, was released in 1990. It contained so much
logic that it didn’t fit onto a single chip, and had to be packaged as a multi-chip
module (MCM). There were separate chips for the Instruction Cache, the Fixed-
Point Unit, and the Floating-Point Unit, and the Data Cache was split across four
chips. With today’s technology, one could fit thousands of those processors on a
single die.
The first high-end CMOS-based mainframe processor was released in 1997 in
the IBM* S/390* G4 system. The entire processor fitted onto a single chip with
roughly two million transistors. The processor was a simple in-order scalar pipeline,
which was clocked at 300 MHz. 12 processors could be assembled on a MCM to
build the symmetric multi-processing (SMP) system.
Numerous major steps happened between the G4 system in 1997, and today’s
high-end IBM zEnterprise* 196 processor. In 1998, branch prediction and binary
floating point was added (traditionally, mainframes use hexadecimal floating
point). In 2000, the IBM eServer* zSeries* 900 system introduced the 64-bit z/
Architecture. The IBM eServer zSeries 990 system in 2003 was the first mainframe
to break through the 1 GHz barrier. The processor implemented a super-scalar
pipeline, with two Load/Store units, two Fixed-Point Units, and one Floating-Point
Unit; execution was in-order, but memory accesses could be done out-of-order. It
was also the first mainframe to have two full processors on a single chip (“dual
core”). The IBM eServer zSeries 990 also introduced the “book concept”, in which
MCMs, each containing multiple CPU-chips, are packaged on a circuit board which
can be plugged into a backplane. Up to four such “books” can plugged in like books
in a bookshelf. This created a SMP with up to 32 processors; the CMOS technology
advancements allowed this to grow to 54 processors in the IBM System z9* in
2005. Processor frequency grew continuously from 300 MHz in G4 to 1.7 G Hz in
the IBM System z9.
From the G4 system to the IBM System z9, the basic processor pipeline was
maintained, adding new features such as superscalar and limited out-of-order
generation by generation. The IBM System z10* processor released in 2007 was
a complete redesign. The old pipeline could no longer provide the required perfor-
mance improvements, so a decision for an ultra-high-frequency pipeline was made;
the system was released with the main pipeline running at 4.4 GHz, with four cores
per chip (“quad core”), and with four books up to 80 CPUs in the system (of which
64 can be purchased by the customer, the remaining cores handle I/O or are used as
spares). In order to achieve this high frequency, the pipeline was simplified to a
two-scalar in-order pipeline.
The IBM zEnterprise 196 is the out-of-order successor of the z10 processor
pipeline. The step to 32-nm CMOS technology allows this to run at 5.2 GHz, the
fastest clocked commercial microprocessor ever. Figures 10.3 and 10.4 show the
functional block diagram and the pipeline diagram, respectively. Figure 10.2 shows
5.2
4.4 GHz
GHz
5000
4000
3000
MHz
1.7
GHz
1.2
2000 GHz
770
420 550 MHz
300
MHz MHz
1000 MHz
0
1997 1998 1999 2000 2003 2005 2008 2010
G4 G5 G6 Z900 Z990 z9 EC z10 EC z196
G4 –1st full-custom CMOS S / 390 z990 – Superscalar CISC pipeline

G5 – IEEE-standard BFP; branch target prediction z9 EC – System level scaling
G6 – Cu BEOL z10 EC – high-frequency in-order pipeline,
z900 – Full 64-bit z / Architecture z196 – high-frequency out-of-order pipeline
Fig. 10.2 Frequency development for system z. BFP binary floating point, BEOL back-end of
line, CISC complex-instruction-set computer
Fig. 10.3 Block diagram for the z196 processor

3 instr Pipeline
D1
decode
D2 LSUs
IDU
N0
AG D$ D$ FM WB Fin
G1 group
write
N1
completion
agen format
back
G2 xfr
AG D$ D$ FM WB Fin
N2
GR
M0 cam RF
FXUs N3
M1 map EX CC WB Fin
R0
ISU
dep mtrx age IssQ EX CC WB Fin
M2 wake R1
wrt mtrx read
chkpnting
reg
S0 S1 S2
read R2
3 BFU
XF F0 F1 F2 F3 F4 F5 F6 F7 F8 WB Fin
R3
FPR
RF DFU R4
F1 F2 ooo WB Fin
RISC
execution out of order execution
units
Fig. 10.4 z196 Processor pipeline diagram
the frequency development for all System z generations starting with the first
CMOS machine G4. The Instruction-fetch/branch (IFB) unit contains an asynchro-
nously running branch predictor, which feeds instruction fetch requests into the
instruction cache (ICM). The instruction cache feeds through instruction buffers
into the synchronous part of the pipeline, starting with the Instruction Decode Unit
(IDU). Through three cycles, the instructions are decoded and grouped into three-
wide dispatch groups. This is also the place where cracking for complicated CISC
ops occurs; for example, a Load Multiple (LM) instruction is cracked into up to 16
micro-ops, each loading one register. A fourth IDU cycle is used to transfer the
decoded and cracked micro-ops into the Instruction Sequencing Unit (ISU). There,
the register renaming takes place before the micro-ops are written into the Depen-
dency and Issue Matrices. The issue bandwidth is five micro-ops per cycle; the five
RISC-like execution units are two Load/Store units (LSUs), two Fixed-Point units
(FXUs), and one floating-point unit, which consists of two sub-units for binary and
decimal floating point (BFU/DFU).
Instructions are completed in program order. At dispatch time, each micro-op is
written into the ISU, but it is also written into the Group Completion Table (GCT),
which keeps track of the execution status of all micro-ops. Once all micro-ops of an
instruction have finished, the completion logic reads out the completion-table entry
and completes the instruction. After completion, error correcting code (ECC)
hardening of the state happens (see Sect. 10.3).
Since the z/Architecture is a CISC architecture, many complex instructions
exist. There are various design features unique to the CISC processor to support
this architecture:
– Many instructions use one operand from memory and a second operand from a
register, with the result being a register again (so-called RX-instruction), for
example R1 ( R1 + mem(B1 + X1 + D1). To prevent cracking of those
instructions into two micro-ops, the concept of dual-issue was invented. A single
micro-op in the ISU issues twice, first to the LSU to load the memory operand
into the target register of the instruction; then a second time to the FXU to read
that target register and the R1 register, perform the arithmetic operation, and
write the final result back into the target register. This saves issue queue
resources and thus allows for a bigger out-of-order window.
– Certain instructions operate directly on memory, for example Move Character
(MVC) or Compare Logical Character (CLC) compare 1–256 bytes of two
memory operands. These instructions are particularly common in commercial
workloads (e.g., databases) and thus critical for the system performance. In order
to provide best-possible performance, a sequencing engine is built into the LSU
to execute those instructions at maximum cache bandwidth, without needing to
crack them at dispatch time into many micro-ops.
– The concept to handle particularly complex instructions in a firmware layer called
“millicode” is also employed in the z196 processor; this concept exists since G4.
Instructions such as Compare String (CLST), which correspond to the cmpstr C-
library routine, execute in this firmware; the hardware implements special
“millicode assist instructions” to facilitate fast execution of the millicode.
– The z/Architecture provides special instructions for fixed-dictionary compres-
sion and for cryptography. These instructions are executed in millicode, but a
hardware co-processor is invoked by millicode to accelerate the computations.
This enables much faster compression and cryptography than would be possible
by discrete instructions.
The instruction cache is 64 kB, the data cache is 128 kB. Both share a 1.5 MB
second-level cache. Also, the translation lookaside buffers (TLBs) are large at 64 2
way and 256 2 way (I-TLB/D-TLB). The D-TLB also supports native 1 MB-pages
in a separate 32 2 way 1 M TLB. The first-level TLBs are backed up by a shared
second-level TLB, which also includes a hardware translation engine (XU). These big
structures particularly help commercial workloads with their big instruction and data
footprint. But running out-of-order at very high frequency also provides very good
performance for compute-intensive workloads, as they become more and more impor-
tant in today’s business intelligence/business analytics workloads.
10.2 Errors, Recovery, and Reliability
The continuous decrease in transistor size has an increasing influence on the amount
of transient errors that happen in a processor. The causes of these transient faults
include radiation, aging, and voltage spikes. Researchers have shown that the
soft-error rate increases around 8% per technology generation [2]. Recent research
in the field of graphics processors used for the Folding@home* project showed
memory error rates of about four fails per week [3]. These numbers show that the
requirements for error-detection and correction are real today and will significantly
increase over the next technology generations. Mainframes have a long history of
built-in error-detection and recovery functionality, but in recent years even small
and mid-sized servers as well as graphics cards used in high-performance comput-
ing saw the addition of ECC logic to their processors. These traditional approaches
to error-detection and correction in hardware have been accompanied by some
interesting developments in software-based reliability.
The 1997 system G4 already contained special circuitry for reliability, for
example the Decode, Fixed- and Floating-Point Units were duplicated to detect
hardware faults. Every result is checkpointed with ECC-hardening in the so called
Recovery Unit. When an error is detected, checkpointing of incoming results is
blocked, and the processor goes through a recovery sequence. During the recovery
sequence, the R-Unit state is read, ECC-checked and corrected, and all architected
registers [e.g., the general-purpose registers (GRs) and floating-point registers
(FRs), the Program Status Word including the Instruction Address, . . .] in the
processor are refreshed from the R-Unit state. The caches, TLBs, etc. are purged
to remove any possible parity fail in them. After the recovery sequence completes,
the instruction processing resumes at the current instruction address that resulted
from the last checkpointed instruction. Thus the recovery sequence cleans up
hardware faults, completely transparent to software.
The concept of checkpointing and recovery still exists today in the z196 proces-
sor, and with shrinking technology it becomes more and more important. However,
due to power limitations, the units are no longer duplicated. Instead, special
checking circuitry is used to detect any hardware faults. Every databus is protected
with parity; computational units either use parity prediction or residue-checking to
detect errors in the computation. Control logic is checked by either local duplica-
tion, or by consistency-checking on the state machine and its transitions. For
example, certain state machines are coded as one-hot states, and an error is detected
should the state ever be not one-hot. (See also [4])
10.3 Scaling Trends
Figure 10.5 shows how each new technology generation is providing less relative
performance improvement and requires alternative solutions to increase perfor-
mance from one machine generation to the next. Some predictions actually show
that technology may even have a negative impact on performance due to power
density considerations.
Moore’s law is still valid today, and it will probably stay valid for the next
decade. However, Moore’s law only addresses the density of the technology, not
Gain by Traditional Scaling Gain by Innovation
100 %
80 %
Relative % Improvement
60 %
40 %
20 %
0%
180 nm 130 nm 90 nm 65 nm 45 nm 32 nm
Fig. 10.5 Improvements by scaling and innovation
other attributes such as frequency or power consumption. However, threshold

voltage is not shrinking at the historical rate, which overall leads to higher
power and power density at constant frequency. This puts a limit on the frequency
growth. The industry trend over the last 5 years clearly showed that the “frequency
race” is over, at least for mainstream processors. For mainframes and other
high-end servers, the situation is slightly different since more effort can be put
into current delivery and cooling; but even with enormous efforts in power and
cooling, the frequency growth will be limited by power consumption and power
density.
Modern microprocessors such as POWER7* or the z196 CPU are very complex
out-of-order processors that aim to extract as much instruction-level parallelism
from the program as possible. However, this approaches the point of diminishing
returns, since the inherent parallelism of single-threaded programs is limited. This
leads to only small improvements in Instructions-per-cycle (IPC) compared to the
past.
Given the limitations in frequency growth and IPC-growth, it seems obvious that
the future lies in thread-level parallelism. Recent years brought multi-cores and
symmetric multi-processing. The POWER7 processor chip for example contains
eight cores, each running four threads in parallel. We expect this trend to continue:
the number of cores will grow approximately with the CMOS shrinking factor, i.e.,
whereas in the past the CMOS density improvements went into more and more
complex processors supporting higher and higher IPC and levels of SMT, we will
now instead see an exponential growth in cores per chip.
The provided thread-level parallelism comes at a high cost: single-threaded

programs cannot exploit the available compute power, and large investments into
multi-threaded programming have to be made.
10.4 Accelerators
Another trend in the industry is the use of accelerators or specialty engines in order
to address the performance needs of a subset of the workloads encountered on
general-purpose processors. This trend started out in the field of 3D computer
games, which require a large number of computations to correctly display a 3D
game scene on the 2D computer screen. The increasing complexity of these game
scenes drove the invention of additional processing capabilities in the Graphics Pro-
cessing Units (GPUs) of PCs. It was also adapted in the Cell Broadband Engine*
processor jointly developed by Sony*, Toshiba*, and IBM with its Synergistic
Processing Elements. Today, the field of accelerators is still dominated by the use
of graphics processor hardware to speed up computations on numerically intensive
and parallel problems. These accelerators usually consist of several hundred
or thousand rather simple processing elements and achieve the best performance
if the computation task consists of a large number of independent calculations
that can be done in parallel. Over the next couple of years, there will be an increase
in all kinds of special- or general-purpose accelerator hardware, examples available
today are IBM’s Smart Analytics Optimizer*, Fixstar’s GigaAccell 180* based on
IBM’s PowerXCell* 8i processor, IBM System z Crypto and Compression unit
available on each System z processor, nVidia* Tesla* compute clusters, or Intel’s
“Intel Many Integrated Core Architecture”* (formerly codenamed “Larrabee”*).
Furthermore, several companies and research projects are investigating the use of
field-programmable gate arrays (FPGAs) for acceleration.
Accelerators suffer from two big problems today, first they need to be connected
via a high-speed interconnect bus otherwise the potential performance improve-
ment may be offset by the latency or bandwidth limitations of the interconnect.
Secondly the integration of these accelerators requires special programming models
in order to use their special purpose hardware. This means that a significant
investment needs to be made in terms of software development in order to convert
an existing application to use accelerator hardware directly. In order to shield the
software developers from having to know the details of the accelerator hardware,
accelerator vendors could provide libraries that abstract the accelerator function for
ease-of-use in software. Especially with FPGAs, it becomes very hard to imagine
that they will ever provide enough ease-of-use for software developers to become
widely used.
For more generic accelerators such as GPUs but also general purpose processors,
the industry is converging on a common programming model as well as an
application programming interface (API) to use these devices and coordinate the
parallel computations. This programming model is called OpenCL*. OpenCL
solves many of the tasks required for parallel programming and different accelera-
tor types. However, the big disadvantage OpenCL faces is that it is based on the
smallest common set of capabilities of the different accelerator types, i.e., it limits
the user to a programming model that works well on GPUs but leaves a lot of the
capabilities of general-purpose processors unused. Examples of features commonly
used in general purpose processors but not available in GPUs would be recursion,
function pointers, and callback functions. The OpenCL standard does not support
these constructs but vendors can define vendor-specific extensions to overcome
these limitations. However, in that case the generic programming model becomes
limited to using the hardware a vendor provides and therefore the application
cannot be run on different accelerators easily.
10.5 Hardware–Software Collaboration
With the trends described above, significant effort needs to be put into collaboration
between hardware and software groups to unlock the potential that modern micro-
architectures provide. This collaboration happens at various levels.
Today’s processor architects interlock regularly with the compiler groups.
Modern compilers (be it static compilers such as C/C++ or dynamic compilers
like a Java* just-in-time [JIT] compiler) internally model the processor pipeline, in
order to influence instruction scheduling. The goal is to create a compiled program
that fits almost perfectly to the internal CPU pipeline, to allow for best possible
instruction-level parallelism. On the z196 processor for example, branch instruc-
tions end the three-wide dispatch groups. The compiler models this and tries to
schedule branch instructions such that groups do not end prematurely.
New instructions are added to the architecture on almost every new generation of
a processor. Often the suggestions for new instruction come from compiler or
software groups. In order to exploit the new instructions, sometimes it is sufficient
to simply change the compiler; in particular with dynamic compilers like the Java
JIT this leads to rapid exploitation. Other enhancements to the architecture have
a longer lead-time. For example, the Power6 and z10 system added hardware for
Decimal Floating Point (DFP) arithmetic. It may take some years until software
explicitly uses this feature, since current software’s data types do not match the
format. Again, Java is an early adopter, exploiting the DFP hardware for the
BigDecimal class; this gives a vast variety of programs a performance boost,
without rewriting those programs [5].
Multi-core and symmetric multi-threaded processors also pose exploitation
challenges. If a computer runs enough independent programs to use the available
hardware threads, exploitation is immediate (at the cost of higher n-way overhead
in the operating system). Many important applications need to be rewritten to give
the operating system enough tasks to exploit the many hardware threads.
Maybe the biggest exploitation challenges exist for accelerators. For example,
it may be relatively easy for hardware to implement a Regular-Expression
or XML (Extensible Markup Language)-parsing accelerator. However, there are

many different software libraries for those tasks, and many software products
even use ad hoc solutions instead of standard libraries. To exploit any specialized
hardware, a rewrite of those libraries or ad hoc functions would be required.
The example explains why on-chip accelerators are rare on today’s mainstream
processor chips. With slower improvements of single-thread performance, the
investment in such features on the hardware and software side may be more
viable to meet future performance targets. Middleware such as WebSphere
Application Server could play an important role in such a development, since
many applications call the Application Server for standard functions like parsing;
this might enable early exploitation, similar to how Java can exploit DFP faster
than other software.
Exploitation of GPUs for general-purpose computation as well as exploitation of
FPGA accelerators may be boosted by the new OpenCL standard. OpenCL defines
a language to describe algorithms device-independently, and an API for appli-
cations to communicate with the accelerator. The program is compiled to the
available accelerator, which could be a rack of blade servers in a hybrid IBM
zEnterprise 196 mainframe, or a rack of GPUs as in nVidia’s Tesla compute
clusters. One drawback of the OpenCL language is its restriction of commonly
used general-purpose programming techniques such as recursion, function pointers,
and callback functions. GPUs do not support those features; for general purpose
accelerator programming, the restriction of those features could negatively impact
OpenCL exploitation.
While there is no direct support for OpenCL from any of the major FPGA
vendors, it is also conceivable to use the OpenCL API and the operating system
services for managing OpenCL devices as an application access path to the FPGA
accelerator.
10.6 Supercomputers
The world of high-performance computing and supercomputers differs from

the general-purpose computing market in that applications on the supercomputer
systems are highly optimized for the specific target architecture. The high degree
of software optimization sometimes results in system architectures that would
rarely be useful in typical commercial applications. One recent example of such
a system-design point is the QPACE* project jointly developed by Deutsches
Elektronen-Synchrotron (DESY), Research Center Juelich, University of
Regensburg, and IBM Deutschland Research and Development GmbH [6]. This
system is highly optimized for research in the area of quantum chromodynamics
and uses a specialized high-throughput 3D torus interconnect for communication
between the compute nodes. Even though this system may not be of interest for
commercial applications, it is an excellent example of successful hardware/soft-
ware co-design.
The majority of the systems listed in the TOP500 list for supercomputers [7]
are built from industry standard components that are extended with specialized
interconnects. In the 1990s, the majority of supercomputers were constructed based
on vector-processor architectures [8], but, over time, the specialized vector processors
almost completely vanished and were replaced by industry-standard processor
architectures like x86 and Power. This trend, among other reasons, is owed to the
fact that vector processors were not attractive for commercial applications, and the
development of specialized processors for the field of supercomputing became too
expensive. Another reason is that today’s x86 and Power processors all contain
Single-Instruction Multiple-Data Floating-Point Units (SIMD FPUs), which deliver
the unprecedented floating-point performance required for supercomputers.
There is a clear trend in the TOP500 list towards hybrid systems consisting of
standard processors coupled with specialized accelerators. The accelerators are
either tightly coupled with the processor, e.g., the Cell Broadband Engine, or
realized through PCI Express cards, e.g., nVidia Tesla. In comparison to commer-
cial computing, the exploitation of accelerators is quicker due to the high degree of
specialized and optimized software, and the inherent parallelism of many mathe-
matical algorithms.
Regarding performance growth over the next years, it is safe to assume that the
historic performance growth rates between 20% and 40% per 6 months will be
maintained [9]. Owing to government funding, a lot of information on the future
high-performance systems is publicly available, for instance the Lawrence
Livermore National Laboratory plans to install a 20 petaflop/s system named
“Sequoia” in 2012 [10,11]. The system will be based on the latest generation of
IBM’s System Blue Gene* called Blue Gene/Q*, and it will contain about 1.6
million processor cores.
10.7 Outlook
The field of processor development has evolved significantly over the last 40 years.
While in the early days technology advancement defined processor performance,
today power density is becoming the limiting factor and performance per watt the
guiding metric. These changes are driving the requirement for innovation into other
areas of processor and system design. Application performance can be influenced
on the following five levels:
– Technology
– Micro-architecture/system design
– Architecture
– Compilers/operating systems
– Applications
With diminishing technology gains, the other four levels are becoming increas-
ingly important and, in order to obtain optimal results, all of them need to be taken
into account during the development cycle, e.g., it does not make sense to introduce
new instructions to the architecture if they cannot be exploited efficiently by
compilers and applications.
This new paradigm will also require hardware designers to get more and more
involved in the software exploitation of the processors they design. Future perfor-
mance gains can only be realized if a strong hardware–software collaboration is
employed during the system design.
We are certain that successfully optimizing the whole stack of system design
will result in another decade of exponential performance growth.
10.7.1 Trademarks
The following are trademarks of the International Business Machines Corpo-

ration in the United States and/or other countries.
AIX* FICON* Parallel Sysplex* System z10

BladeCenter* GDPS* POWER* WebSphere*
CICS* IMS PR/SM z/OS*
Cognos* IBM* System z* z/VM*
DataPower* IBM (logo)* System z9* z/VSE
DB2* Blue Gene Blue Gene/Q zEnterprise
*Registered trademarks of IBM Corporation
The following are trademarks or registered trademarks of other companies.

Adobe, the Adobe logo, PostScript, and the PostScript logo are either registered
trademarks or trademarks of Adobe Systems Incorporated in the United States, and/
or other countries.
Cell Broadband Engine is a trademark of Sony Computer Entertainment, Inc. in
the United States, other countries, or both and is used under license there from.
Java and all Java-based trademarks are trademarks of Oracle, Inc. in the United
States, other countries, or both.
Microsoft, Windows, Windows NT, and the Windows logo are trademarks of
Microsoft Corporation in the United States, other countries, or both.
InfiniBand is a trademark and service mark of the InfiniBand Trade Association.
Intel, Intel logo, Intel Inside, Intel Inside logo, Intel Centrino, Intel Centrino
logo, Celeron, Intel Xeon, Intel SpeedStep, Itanium, and Pentium are trademarks or
registered trademarks of Intel Corporation or its subsidiaries in the United States
and other countries.
UNIX is a registered trademark of The Open Group in the United States and
other countries.
Linux is a registered trademark of Linus Torvalds in the United States, other
countries, or both.
ITIL is a registered trademark, and a registered community trademark of the
Office of Government Commerce, and is registered in the U.S. Patent and Trade-
mark Office.
IT Infrastructure Library is a registered trademark of the Central Computer and

Telecommunications Agency, which is now part of the Office of Government
Commerce.
*All other products may be trademarks or registered trademarks of their respec-
tive companies.
References
1. Davenport, T.H., Harris, J.G.: Competing on Analytics: the New Science of Winning. Harvard
Business School, Boston (2007)
2. Hazucha, P., Karnik, T., Maiz, J., Walstra, S., Bloechel, B., Tschanz, J., Dermer, G., Hareland,
S., Armstrong, P., Borkar, S.: Neutron soft error rate measurements in a 90-nm CMOS process
and scaling trends in SRAM from 0.25-mm to 90-nm generation. IEEE IEDM (International
Electron Devices Meeting) 2003, Technical Digest, pp. 523–526 (2003). doi: 10.1109/
IEDM.2003.1269336
3. Haque, I.S., Pande, V.S.: GPUs: TeraFLOPS or TeraFLAWED? www.cs.stanford.edu/people/
ihaque/posters/sc-2009.pdf. Accessed 20 Sept 2010
4. Meaney, P., Swaney, S., Sanda, P., Spainhower, L.: IBM z990 soft error detection and
recovery. IEEE Transactions on device and materials reliability, Vol. 5, No. 3, September
2005, pg. 419–427
5. Mitran, M., Sham, I., Stepanian, L.: Decimal floating-point in Java 6: best practices. www-304.
ibm.com/partnerworld/wps/servlet/ContentHandler/whitepaper/power/java6_sdk/
best_practice. Accessed Jan 2009
6. Pleiter D.: QPACE: QCD Parallel Computing on the Cell. www.itwm.fhg.de/hpc/workshop/
mic/Qpace_%28Dirk_Pleiter_-_Desy%29.pdf. Accessed 28 Oct 2008
7. www.top500.org.Accessed 29 Sept2010
8. www.top500.org/overtime/list/35/procarch. Accessed 29 Sept 2010
9. www.top500.org/lists/2010/06/performance_development. Accessed 29 Sept 2010
10. https://asc.llnl.gov/publications/leadingHPC.pdf. Accessed 29 Sept 2010
11. https://asc.llnl.gov/computing_resources/sequoia/index.html. Accessed 29 Sept 2010
Chapter 11
Towards Terabit Memories
Bernd Hoefflinger
Abstract Memories have been the major yardstick for the continuing validity of
Moore’s law.
In single-transistor-per-Bit dynamic random-access memories (DRAM), the num-
ber of bits per chip pretty much gives us the number of transistors. For decades,
DRAM’s have offered the largest storage capacity per chip. However, DRAM does
not scale any longer, both in density and voltage, severely limiting its power efficiency
to 10 fJ/b. A differential DRAM would gain four-times in density and eight-times in
energy. Static CMOS RAM (SRAM) with its six transistors/cell is gaining in reputa-
tion because it scales well in cell size and operating voltage so that its fundamental
advantage of speed, non-destructive read-out and low-power standby could lead to
just 2.5 electrons/bit in standby and to a dynamic power efficiency of 2aJ/b. With a
projected 2020 density of 16 Gb/cm2, the SRAM would be as dense as normal DRAM
and vastly better in power efficiency, which would mean a major change in the
architecture and market scenario for DRAM versus SRAM.
Non-volatile Flash memory have seen two quantum jumps in density well
beyond the roadmap: Multi-Bit storage per transistor and high-density TSV
(through-silicon via) technology. The number of electrons required per Bit on the
storage gate has been reduced since their first realization in 1996 by more than an
order of magnitude to 400 electrons/Bit in 2010 for a complexity of 32Gbit per chip
at the 32 nm node. Chip stacking of eight chips with TSV has produced a 32GByte
solid-state drive (SSD). A stack of 32 chips with 2 b/cell at the 16 nm node will
reach a density of 2.5 Terabit/cm2.
Non-volatile memory with a density of 10 10 nm2/Bit is the target for wide-
spread development. Phase-change memory (PCM) and resistive memory (RRAM)
lead in cell density, and they will reach 20 Gb/cm2 in 2D and higher with 3D chip
stacking. This is still almost an order-of-magnitude less than Flash. However, their
B. Hoefflinger (*)

230 B. Hoefflinger
read-out speed is ~10-times faster, with as yet little data on their energy/b. As a read-
out memory with unparalleled retention and lifetime, the ROM with electron-beam
direct-write-lithography (Chap. 8) should be considered for its projected 2D density
of 250 Gb/cm2, a very small read energy of 0.1 mW/Gb/s. The lithography write-
speed 10 ms/Terabit makes this ROM a serious contentender for the optimum in non-
volatile, tamper-proof storage.
11.1 High-Speed CMOS SRAM (Static Random-Access

Memory)
The invention of CMOS technology in 1963 (Figs. 2.7 and 2.8) produced the first
CMOS SRAMs by 1969. The classical 6 T memory cell (Fig. 3.18) is displayed here
as Fig. 11.2, in order to compare it with the most recent 8 T cell, 40 years later [1],
shown in Fig. 11.3.
The dominant issues for static RAM are
– Cell size
– Access time
– Standby power
– Power efficiency or energy/bit.
A representative recent layout is shown in Fig. 11.4 for a 90-nm design.
While the cell size in [2] is about 180 F2 in a standard logic process, where F is
the first-metal half pitch in the sense of a node in the ITRS roadmap, the cell size in
[1] is 121 F2 at the 45-nm node for this 8 T design in a more memory-specific
process. In a 2D planar design, the area for a 6 T SRAM cell is >110 F2. Any
significant improvement of the cell size, i.e., the SRAM density, beyond the scaling
to ~16 nm, will require the 3D integration of the transistors in the cell, as we will see
shortly.
We consider next the speed of the SRAM, the primary reason for operating with
this complex cell. We see in Fig. 11.2 that the current necessary for inverting the
central nodes, carrying the cell information, into the opposite state in the write
mode, is supplied through the VDD and GND nodes per cell, providing a fast write.
In the read mode, again, the current to charge or discharge the bit lines and the
inputs of the sense amplifier at the end of a column is provided by these per-cell
sources, allowing a fast read access time. Obviously, these times depend on the ON
currents of the four transistors in the quad. In these operations, we face a difficult
trade-off between the charging currents, which mean speed of access, rising
with VDD, and the switching energy CVDD2, which decreases with a lower VDD
(see. Sects. 3.2 and 3.3):
1=taccess IDmax =VDD expðVDD VT Þ=VDD

11 Towards Terabit Memories 231
10 Tb/ cm2
32-chip stack 2.5 Tb/ cm2
16-chip stack
1 Tb/ cm2
8-chip stack
640 Gb / cm2
x 250 Gb/ cm2
R OM
80 Gb/ cm2
3b/ cell 50 Gb/ cm2
M
fD RA
lash x Dif 16 Gb/ cm2
F
M LC 13 Gb/ cm2
S 8 Gb/ cm2
O
CM
3D
M/
PC AM
AM RR
DR
1 Gb/ cm2
AM
SR
100 Mb/ cm2

2000 2010 2020
Year
Fig. 11.1 Evolution of CMOS memory density for SRAM, DRAM and Flash memory. MLC
multi-level per cell, PCM phase-change memory, RRAM resistive RAM
in low-voltage barrier-control mode or
ðVDD VT Þ1=2 :
in higher-voltage drift mode.

We have analyzed these relationships in Sect. 3.2, and in Sect. 3.3 on CMOS,
we found the additional complication that, while we pursue an energy priority in
lowering the supply voltage, the inherent DC standby current
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Istdby ¼ Imax Imin ;
where Imax is the maximum transistor on-current and Imin the minimum off-current,
and DC standby power
232 B. Hoefflinger
VDD
Word Line Word Line
Bit
Line Bit
Line
Fig. 11.2 The CMOS 6 T SRAM cell. Its core is the cross-coupled inverter pair, the quad. When
the word line turns on the two access transistors, the content of the cell can be read as true and
complement by a differential sense amplifier without modifying the content of the cell. In the write
mode, the content of the cell can be confirmed or forced to the inverse
RWL0
VCC VCC
WR PU WR “1”
Local RBL (LBL)
contention completion
“0” “1” “0” Nx “1”
RD Port
PCH
WWL0
bit0
VCC
bit1
WBL#
WBL
•••
Global RBL (GBL)
bit7 domino
keeper
write driver
local
merge
Fig. 11.3 A recent 8 T SRAM cell [1], in a 45-nm, CMOS technology. Two additional transistors
form an asymmetrical path for a fast read at low voltages. RWL0 ¼ read word line ¼, WWL0 ¼
write word line 0, RD ¼ read, RBL ¼ read bit line, WBL ¼ write bit line, LBL ¼ local bit line,
GBL ¼ global bit line, PCH ¼ pre-charge (# 2010 IEEE)
BL /BL
Poly-gate
WL
VSS VSS
P-Well diffusion
N-Well
Contact hole F
VDD VDD
BL /BL
WL
VSS VSS
P-Well
PD halo imp.
F
N-Well
VDD VDD
Fig. 11.4 Layout of a 6 T cell in 90-nm CMOS [2]. Memory cell sizes are usually given in
multiples of half-pitch units F. The size of the cell shown here is about 12 15 ¼ 180 F2. PD
halo imp. ¼ PMOS transistor halo implant. (# 2010 IEEE)
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PDC ¼ VDD Imax Imin
means a constant loss, eventually becoming larger than the dynamic switching
power
Pdyn ¼ VDD fImax ;
where f is a duty factor <1/4. Within the total power

pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Ptot ¼ PDC þ Pdyn ¼ VDD Imax ðf þ Imin =Imax Þ;
the on/off current ratio has to be >16 to keep the dynamic power the dominant
component. We have seen in Sect. 3.2 that the low-voltage barrier-control mode
offers larger on/off ratios than the higher-voltage drift mode and the lowest energies
CV2/bit anyway.
The price, however, is access time, and we take a closer look at the access time,
selecting two publications cited before [1, 2]. We have plotted the log of the
measured read access times against the supply voltage in Fig. 11.5.
In the semilog plot we see distinct slopes for the dependence of access time
on supply voltage. At higher voltages, the memory cells work mostly in strong
inversion, where current and access time have a power dependence on voltage,
approximately V1/2. At voltages <0.9 V, we see an exponential dependence
234 B. Hoefflinger
10
x
50
0
m x
V
/d
ec x
35
x
0m
[11.2]
Read Access Time (ns)
V/
x
de
c
x
[11.1]
0.1
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2
VDD (V)
Fig. 11.5 Measured read access times versus supply voltage VDD. [1] ¼ 45 nm, [2] ¼ 90 nm
characteristic of the sub-threshold or barrier-controlled current in the cell with

slopes of 350 mV/decade and 500 mV/decade, respectively.
The energy/bit, also called the power efficiency/bit, has been reduced to
<100 nW Gb–1 s1, or <0.1 fJ/b, for sub-1-ns access times, as shown in
Fig. 11.6, and, by lowering the operating voltage to 0.57 V for the same 45-nm
chip, can be as low as 0.03 fJ/b, while the access time increases to 3.4 ns.
The process design and the operating point of these SRAMs are the result of
a sensitive balance including dynamic and static DC power as well as speed. The
45-nm design quoted here has a DC power of 0.6 mW/b compared with a dynamic
power of 0.3 mW/b, indicating these compromises.
The robustness of state-of-the art SRAMs can be learned from [3], where data
such as the static noise margin (NM) and the 10–90% corridor of minimum supply
voltages were reported. We covered the fundamentals in Sect. 3.3, when we studied
CMOS inverter transfer characteristics and the cross-coupled inverter pair
(Fig. 3.17), the quad, and its noise margin (3.25). Figure 11.7 shows measured
transfer characteristics on CMOS SRAM cells in a 32-nm technology [3].
It is impressive how these characteristics of 2009 still resemble the ideal model-
curve of the 1960s (Fig. 3.15). The extracted noise margins correspond to average
voltage gains of 1.5 at 1.1 V and 2.5 at 07 V, remarkable results for a 32-nm design
and for the sub-threshold regime involved at 0.7 V. The paper also reports that the
10-16 600 eV
20
10
energy / b (J)
energy/ b
10-17 60 eV
2020
10-18 6 eV
0.1 ns 1 ns 10 ns
access time
Fig. 11.6 Dynamic energy/bit of SRAMs versus the access time
Fig. 11.7 Transfer

characteristics of 32-nm
SRAM cells for three supply
voltages [3]. The static noise
margins NM are indicated as
the largest inscribed (dotted)
squares inside a characteristic
and its mirror image (# 2009
IEEE)
90% range of minimum supply voltages for dies on a single wafer reached from
0.79 V to 0.92 V or 130 mV, indicating the variance present at 32 nm and a sign
that this spread will widen quickly for smaller features in the future.
How much room is there for scaling SRAMs? We base our assessment on the
characteristics of the 10-nm nanotransistor treated in Sect. 3.2 and the CMOS
inverter and quad described in Sect. 3.3. Here is a listing of the characteristics of
the 6 T 10-nm cell:
10-nm CMOS 6 T SRAM 2020
W ¼ L ¼ 10 nm
VDD ¼ 0.3 V
236 B. Hoefflinger
VT ¼ 0.3 V
ION ¼ 25 nA; IOFF ¼ 0.25 nA
IDC ¼ 2.5 nA
PDC ¼ 0.75 nW/b ¼ 750 mW/Gb
Duty factor f ¼ 1/4
Pdyn ¼ 1.8 nW/b ¼ 1.8 W/Gb
Access time ¼ 0.6 ns
Energy/bit ¼ 1.1 1018 J/b ¼ 7 eV/b
Power efficiency ¼ 1.1 nW Gb1 s1
Cell size: 120 F2 ¼ 0.12 109 cm2/b ¼ 12 mm2/Gb
Ptotal ¼ 2.55 W/Gb
Power density ¼ 21 W/cm2
VDDstdby ¼ 0.15 V
Pstdby ¼ 120 mW/Gb
The characteristics listed here mark the 2020 performance level expected for a
memory built with the average 10-nm transistors. The treatment of the fundamental
variance of the CMOS nano-quads in Sect. 3.2.1 told us that VDDmin ¼ 0.37 V and
VDDstdby ¼ 0.21 V would be the limits, and manufacturing tolerances would impose
further allowances on this performance data.
We also recognize that these nano-SRAMs will operate at very high power
densities (21 W/cm2) and therefore in the very-low-voltage, deep-sub-threshold
barrier-control mode. Nevertheless, ON/OFF current ratios are limited, and the
power reduction in standby will be just a factor of 20. We have entered these 2020
data in Fig. 11.6. A sub-1-ns access time will be possible, and the energy/bit will be
100-fold smaller than in 2010, at just 7 eV/b. The active memory cell based on the
CMOS quad maximizes speed and minimizes energy at the price of a larger cell
size. Even so, a density of 8 Gb/cm2 can be predicted for a 2020 2D SRAM.
We have seen in Sect. 3.4 that the SRAM cell is the top candidate for stacking
complementary MOS transistors with a common gate and common drains on top of
each other and that this can more than double the density with the same mask count and
fewer metal contacts so that a CMOS SRAM density of 16 Gb/cm2 is possible in 2020.
A bit more speculative is a future 6 T SRAM cell consisting of the comple-
mentary, vertical bipolar quad and NMOS select transistors proposed in Sect. 3.8.
This quad would have the ideal transconductance of 65 mV/decade of collector
current, offering the highest speed and the best ratios of dynamic power versus DC
power and the relatively lowest standby power together with a very high density of
47 F2, corresponding to 21 Gb/cm2 at the 10-nm node.
The vertical-interconnect topography of the cells in Figs. 3.43 will also be
ideal for:
– The flipped stacking of these SRAMs as local memories on top of arrays of
processor elements (PEs).
– The 3D integration of these PEs with performance levels of 1 fJ per 16 16
multiply (Sect. 3.6), together with 3D SRAMs at 1 fJ/kb will advance informa-
tion processing to truly disruptive innovations (Chaps. 13,16,18).
11.2 DRAMs (Dynamic Random-Access Memories)
DRAMs are the famous MOS memories with just one transistor per bit, invented in
1968 and in production since 1971. The basic elements of this memory are
illustrated in Fig. 11.8.
In this electronic-charge-based memory cell, the key elements and the only
elements in the cell are the storage capacitor CS and the select transistor. A high
number of electrons on the storage capacitor means a logical ONE, a small number
a logical ZERO. If the Word signal on the transistor gate is low, the transistor,
ideally, would be OFF, and the charge on the storage capacitor, i.e., the information,
would stay in the cell indefinitely. In reality, it will leak through the transistor into
the substrate and into the circuitry so that it has to be refreshed by re-writing it into
the cell periodically by selecting the cell (raise the word-line voltage to High),
reading the marginal cell content with the bit-line sense amplifier, and re-writing
a perfect content into the cell. In the same way, with each reading of the cell, its
content is destroyed and has to be re-written immediately. These requirements are
a fundamental overhead on the operation of a DRAM so that write, read, and the
full-cycle time of the DRAM make it slower than an SRAM of the same generation.
What does destructive reading in a DRAM mean? The DRAM cell is a passive cell,
in contrast to the SRAM, an active cell. Active means amplification and instant
regeneration: The (differential) amplifier, the quad, inside the SRAM cell not only
restores the charges in the cell immediately as a read access might destroy the
touchy balance in the cell, it also provides charges to charge and discharge the bit
lines as needed and provide healthy signal levels to the bit-line sense amplifier. By
contrast, charges in the passive DRAM cell are lost by leakage or shared with other
capacitors such as the bit-line capacitor CBL once the select transistor is opened.
In fact, if the bit-line sits at 0 V, the cell capacitor at a ONE ¼ VDD, and the cell is
selected, charges would be shared, resulting in a ONE ¼ 0.5 VDD at the input of
the sense amplifier, if CS ¼ CBL. This voltage-divider effect on signal levels is
the reason for making the storage capacitor large enough that leakage and the
capacitance of long bit-lines cannot spoil signal levels too much.
In the early days, the area of the drain diffusion of the NMOS select-transistor
was enlarged to produce a storage capacitance of ~30 fF. The rapid scaling-down of
cell sizes has turned the realization of miniature capacitors of 30 fF and recently
20 fF into a formidable challenge. Because the footprint of this capacitor more-
or-less determines the cell size, folded and deep-trench capacitors with small
footprints have been developed, and the select-transistor has been manufactured
WL
BL Bit-Line
Fig. 11.8 Memory cell, bit- Sense Amplifier
CS CBL
line, and bit-line sense
amplifier of a MOS DRAM.
WL word line, BL bit line GND
238 B. Hoefflinger
on top of the capacitor to save area, an example of 3D-integration at the device

level. The making of dielectrics with a large relative permittivity (high-k dielectric)
has become the show stopper for DRAMs, and they are trailing other chips on the
roadmap by one generation for this reason and also because of
– Sub-threshold leakage of scaled-down select-transistors and
– Higher supply voltages to assure higher immunity against voltage-noise on
bit-lines.
We now take a look at the 2010 state-of-the-art, analyzing two designs from
opposite sides of DRAM applications. One is aimed at large capacity, by 3D-stacking
of 4 2 Gb chips to produce an 8 Gb product [4]. The other is aimed at a high-speed
embedded memory on a processor chip. Table 11.1 lists the technology, the density, the
organization, the access-time, and the dynamic energy per bit in electron volts per bit,
which illustrates how many electrons are moved from 0 V to 1 V to store 1 bit of
information. We recall that 1 eV ¼ 1.6 1019 V A s (J) or 1 mW Gb1 s1 ¼ 1
fJ ¼ 6.25 103 eV.
The 2-Gb chip from [4] with its cell size of 20 F2 has a density of 2 Gb/cm2, and
it requires a sizable dynamic energy of 120,000 eV/b. Its access time of 12 ns
indicates that a larger string of cells, >512, are connected via one bit-line to
a repeater/sense-amplifier.
The embedded DRAM [5] by contrast has a very short string of only 33 cells
connected to a repeater-amplifier, thereby achieving a short access time of 1.35 ns,
competitive with SRAM, while its density is a factor of eight higher at the same
technology node. This advantage shrinks to 4 times compared with an SRAM
built one technology node further. The energy/bit scales only mildly with voltage
and is >10 that of an SRAM in the same technology.
The scaling-down of DRAMs is a problem not only of producing a high-density
capacitor but also of making a very-low-leakage select-transistor. That is why
a future DRAM, projected in [6], is based on the 22-nm node (Table 11.1), and
the 20 fF capacitor would need an equivalent oxide-thickness (EOT) of 0.1 nm,
which means a relative permittivity k ¼ 30 and a physical thickness tI ¼ 3 nm.
SBT (strontium bismuth tantalite) would be a candidate. Even with the aggressive
Table 11.1 Performance of recent and future DRAMs

Chip F[nm] Area[F2] Bit/repeat taccess[ns] VDD[V] Energy/b [eV/b] Density Ref.
[Gb/cm2]
4 2 Gb 50 20 512? 12 1.5 1.2 105 2 [4]
20 fF (20 mW Gb1 s1)
4 292 kb 45 16 33 1.35 1.0 5.6 104 2.8 [5]
18 fF (9 mW Gb1 s1)
64 Gb proj. 22 4 512 1.0 6.2 104 13 [6]
20 fF (10 mW Gb1 s1)
256 Gb 16 8 128 <1 0.5 8 103 50 This book
DiffDRAM 2 5 fF (1.3 mW Gb1 s1)
proj.
cell size of 4 F2, a target memory with 64 Gb would have a size of 5 cm2 at an
energy of 62,000 eV/b, 500 times higher than an SRAM at the same node.
The storage capacitor plus the bit-line capacitance, the associated voltage
division, the bit-line noise, the transistor leakage and the required supply voltage
seriously limit the advancement of DRAMs. While “a single transistor per bit”
sounds attractive, in particular against the 6 T SRAM, it is timely to re-visit the
architecture of these RAMs. The design of [5] points in a possible direction: It has
one single-ended regenerator/33 bit and single bit-lines, while, at the other extreme,
the SRAM has one differential regenerator/bit and two complementary bit-lines.
Regarding the tremendous benefits of low-voltage differential signaling (LVDS),
consider the differential DRAM in Fig. 11.9.
This differential DRAM stores each bit and its complement on two storage
capacitors, it has two select-transistors and two complementary bit-lines per cell,
and it has fully differential signaling and sense-amplifiers/repeaters. Because the
information content is available as TRUE and COMPLEMENT, and because of the
suppression of common-mode noise, and of differential amplification, the storage
capacitors can be reduced to 5 fF and the voltage swing to 0.5 V, also considering a
WL
BL
Bit
CS CBL Differential
GND GND Bit-Line
CS CBL Sense Amplifier
Fig. 11.9 A differential BL
Bit
DRAM
2020
105
DRAM
Intrinsic Dynamic Energy/bit (eV)
104
DiffDRAM
103
MLC Flash
102
101
x SRAM
Fig. 11.10 Energy/bit and

access time for SRAM, 1
DRAM, DiffDRAM, and 10–1 1 10 102 103 104
NAND Flash Read access time (ns)
240 B. Hoefflinger
short string of just 128 cells per repeater, which amounts to a bit-line capacitance of
0.5 fF each at the 16-nm node, for which we find a line capacitance of 0.18 fF/mm in
Table 5.2. Owing to the lower voltage and the differential-signal nature, we move one
technology node further than the projected single-ended design [6]. The differential
sense amplifier could look somewhat like the accelerator in Fig. 3.29, and it would
provide a fast settling (high slew-rate) to full-swing signal levels. With the quad
inside, it would also meet the transistor-variance concerns at 16 nm (see the preceding
section on SRAM and Sect. 3.3, Fig. 3.17). The projected performance of this design
with a cell size of 8 F2 appears in the last line of Table 11.1: A 16-nm fully differential
DRAM (DiffDRAM) chip is projected to have a density of 50 Gb/cm2, an access time
of 1 ns, and a power efficiency of 1.3 mW Gb1 s–1 ¼ 8,000 eV/b.
In a comparison with a normal single-ended DRAM design, the capacitor density
(fF/mm2) is reduced by 50% (EOT ¼ 0.2 nm), the energy/bit eightfold, and the
density/cm2 is increased fivefold. Memory figures-of-merit are shown in Fig. 11.10.
The energy-efficiency of the future SRAM with just 7 eV/b shows the impressive
benefit of the quad, differential amplification in each cell, at the price of cell sizes of
50 F2 (3D) to 110 F2 (2D), but with the potential of aggressive sizing-down to
10 nm and 3D-CMOS (Sect. 3.5). The DRAM is worst in energy per bit, presently
optimal in the speed density (16 F2) product, but its down-sizing will stop soon
at ~22 nm. Advancements such as DiffDRAM may offer a future optimum for
speed density. The ultimate in bit density is just a single transistor per bit and all
transistors connected in series, that is, each transistor sharing its drain with the
source of its neighbor. We treat this configuration in Sects. 11.3 and 11.4.
11.3 NAND Flash Memories
We can store information on an NMOS transistor in a string of NMOS transistors

by setting its threshold voltage VT to a high or low value. In Fig. 11.11, we show
four such strings with vertical bit-lines per string and horizontal word-lines.
The memory transistors each show a floating gate, on which we assume that we
have placed no charge or a negative charge, individually. If we want to analyze
(read) the transistor (cell) under word-line WL1 in the left string, we bias all
non-selected word-lines to a high voltage, VGH ¼ 5 V, to make sure that all those
transistors are turned on, and we set the gate voltage on WL1 to the read voltage
VGR ¼ 2 V. If the selected transistor had a negative charge on its floating gate,
establishing a threshold voltage VT ¼ 2 V, the transistor would not conduct and
there would be no current flow through the string. If the floating gate had no charge
on it, VT ¼ 0 and the transistor would conduct, and current would flow through the
string, because all the others have been turned on anyway. The topography of this
type of memory is attractive, because each cell has just one minimum-size transistor
and shares source and drain electrodes with its neighbors. No storage capacitor
is needed, because we do not evaluate charge but rather measure a current. The
minimum area would be 1 F2 for the transistor gate, 2 F2/2 for source and drain,
BL0 BL1 BLn–1 BLn
string
select
WL0
WL1
WLm–1
WLm
ground
select
source
Fig. 11.11 NAND array architecture of a Flash memory after [7]
resulting in 2 F2, and with the inclusion of the space F to the next string, the
minimum total area would be 4 F2.
The key is how to place the proper charge on the floating gate, and for
re-programming, how to remove it. The dominant mechanism today for energizing
channel electrons to jump over the potential barrier between Si and SiO2 or SiON,
to reach the floating gate, is the extremely high electric field of several 100,000 V/cm
near the drain, producing hot-electron injection to the floating gate if a sufficiently
high programming voltage is applied to the control gate of the selected transistor.
For ON/OFF programming of the transistor, a difference VT,OFF – VT,ON ¼ 2 V is
produced by injecting ~30,000 electrons into the floating gate. In order to erase the
negative charge on the floating gate, all NMOS memory transistors sit in p-type
wells, typically 4 k cells in each well, and the well is biased positive (~+20 V) while
the gates receive a negative bias. This causes a (non-destructive) breakdown of the
oxide so that all electrons are driven into the well and into the n-type substrate in a
Flash, giving the name to this type of non-volatile memory. A superb and detailed
treatment of non-volatile memory technologies with emphasis on Flash has been
given by Brewer and Gill [7].
242 B. Hoefflinger
Fig. 11.12 The distribution of threshold voltages in the 4 b/cell 64 Gb Flash memory[9] (# 2009
IEEE)
We turn our attention to 1995 and paper [8], cited in Chap. 2 as a quantum jump
for non-volatile MOS memories, those which do not lose their information when the
power is turned off. Experiments had been made before to interpret different
amounts of charge on a floating gate as 2 bits or more, but the Intel researchers in
Folsom, CA, impressed the community in 1995 with a sophisticated, mature 32 Mb
prototype, which could be programmed to four threshold voltages, representing
2 b/cell. Figure 2.19 shows the distribution of threshold voltages with their centers
about 2 V apart. Gaining a factor of two or more in NV (non-volatile) memory
density with multi-level-cell (MLC) Flash at the same technology node triggered an
unparalleled worldwide development effort with a 2,000-fold increase in 2D NV
memory capacity within 14 years. A 64 Gb 4 b/cell NAND Flash memory in 43-nm
CMOS was reported in 2009 [9], and we show the distribution of threshold voltages
in Fig. 11.12.
The spacing of threshold voltages has shrunk to just 250 mV, equivalent to about
1,000 electrons. Only 2 years before, the multi-step programming algorithms,
driving and sensing, had reached a resolution of 3,000 e–/step [7], and in the
meantime, the minimum programming steps must have come down to <300eat
speeds well below 1 ms. The state-of-the-art in 2010 is summarized in Table 11.2.
The table also contains estimates of the electrons/bit and of the intrinsic
energy/bit, showing a spread of 64/1 between the lowest level (111) and the highest
(000). It is worth noting that the intrinsic average energy/bit of 8000eis still
a factor of seven better than that for DRAM and that the bits are non-volatile.
However, the read-out time is 1,000-fold longer at several microseconds, which,
for typical read-out currents of 1 mA, means a poor power efficiency of 10 mW
Gb1 s1.
The remarkable density gains of MLC NAND Flash memories are shown in
Fig. 11.13 with a peak density of almost 30 Gb/cm2 achieved in 2010.
We have added explicitly the prefix 2D to our discussion of NAND Flash so far,
and we stay on the 2D scaling path a bit longer, before we address the 3D quantum
jump. The end of the NAND Flash roadmap such as that for DRAM has been
Table 11.2 2010 MLC Bit/chip 32–64 Gb

NAND Flash RAMs
Technology 35–32 nm
Bit/cell 2–4
Cell area 0.005 mm2 ¼ 5 F2
Bit/cm2 26 Gb/cm2
String 66 b
VDD 1.8 V
VSupply 2.7–3.6 V
Programming time 1.3–11.7 ms
Read time 10–100 ms
Electrons/b (3b/cell) 1000e–8000e
DVT 250 mV
Intrinsic energy/bit 250–16,000 eV
Int. power efficiency 2.5 mW Gb1 s1
Fig. 11.13 The evolution of 2D MLC NAND flash density [Mb/mm2] versus the process node
[10] (#EE Times 2010)
predicted at 22 nm. The scaling of Flash is severely limited by the fact that the
insulator thickness cannot be scaled because of charge losses from the floating gate
due to tunneling. An insulator thickness of 6 nm continues to be demanded for
long retention times, and the effective thickness to the control gate is even larger.
Shortening the gate length massively enhances the influence of the drain on the
transistor channel against the control intended to be dominated by the (control)
gate. We treated this drain-induced barrier lowering in Sect. 3.2 and found that
the threshold voltage becomes severely dependent on the drain voltage (3.18, 19).
To show the effect, we simplify a bit by writing
244 B. Hoefflinger
1
VT ¼ VT0 VDS (11.1)
AV
rffiffiffiffiffiffi
eI L N2
AV ¼ 2 (11.2)
eSi tI N1
AV is the intrinsic voltage gain, obviously a strong function of the L/tI ratio and
the ratio of the permittivities, which is 1/3 for the Si/SiO2 system. The doping
concentration N2 deeper in the channel should be higher than N1 near the surface to
suppress the drain effect. In any case, we see that 1/AV becomes large quickly as L is
reduced while tI is not, giving the drain – source voltage a large influence on the
threshold voltage. Depending on the position of a cell within a string, this can make
thresholds overlap, causing an erroneous read. In addition, the variance of the
charges on the gates will force a reduction of the number of distinguishable levels.
The gate will lose control totally, as soon as field lines in the channel reach through
directly from drain to source, causing a punch-through current to flow irrespective
of the programming of the floating gate and the operation of the control gate.
An optimistic projection towards 2020 is given in Table 11.3.
The concerns about scaling made NAND Flash a prime candidate for 3D
integration by stacking chips with
– Through-silicon vias (TSVs) or
– Inductive coupling.
These technologies are very much in a state of flux. Remarkable results in 2010
were
– A 32 GB test chip with a stack of 8 chips, TSVs and 2 b/cell in 60-nm CMOS
[11], which also indicates various architectural choices.
– A 2 Gb/s inductive-coupling through-chip bus for 128-die NAND Flash memory
stacking [12].
Both publications are representative for the almost boundless advancements in
the realization of ever larger memory densities and capacities on the chip scale.
These interconnect schemes with power efficiencies of ~1 pJ/b ¼ 1 mW Gb1 s1
(see Chap. 5) are basically compatible with the on-chip efficiencies of DRAMs and
Flash. Beyond this prerequisite, the other issues for this progression strategy are
– Testing and known-good die
– Power density
– Reliability
Table 11.3 2020 projection for 2D MLC NAND Flash memory

Technology node Bit/cell Electrons/level Density [Gb/cm2]
22 nm 3 500 60
16 nm 2 250 80
– The manufacturing chain and product liability

– Cost
However, the basic notion is that 3D integration of memory by chip-stacking is
a strategy with identifiable milestones and, therefore, that it is the most important
path to reach the 2020 target of DRAM and NAND-Flash capacity of several terabit
on a footprint of a few square centimeters (see Fig. 11.1).
11.4 Electron-Beam Permanent ROM (Read-Only Memory)
Silicon read-only memories with infinite retention times are of interest whenever
the endurance of NV Flash memories is not sufficient or if the manipulation of the
Flash data, modification or destruction, are an issue. So-called mask-programmable
ROMs have filled this niche. On this type of memory chip, the memory transistors
are set to a high or low threshold voltage by masking or not masking transistor gate
areas. In the old days of metal gates, the thin-gate-oxide area of a transistor was
either present in the gate mask or not. If not, the thick field-oxide would remain,
the word-line gate metal would run across it as across all the others, but the field
threshold voltage, being higher than the read voltage, would keep this transistor
OFF forever. With the return of metal gates in nano-CMOS processes, this type
of cross-bar memory with word-line gates orthogonal to source-drain bit-lines,
becomes attractive again because of its ultimate cell area of 4 F2 and its ultimate
scalability to <10 nm, offering a density of 250 Gb/cm2 (Fig. 11.1).
The read-out of the transistor-string would have to detect a current ratio Hi/Lo,
which can be optimized, contrary to the floating-gate transistor, even for 10 nm
transistors by careful transistor and circuit engineering down to voltage levels
of 0.3 V, where [13] is an exemplary paper. As a consequence, energy/bit can be
a factor of 20 smaller and read-out 100-fold faster than 16 nm Flash. The penalty is
read-only, but that would be acceptable or even desirable for important documents
such as digital libraries or 3D high-quality movies as well as permanent embedded
ROM.
Of course, mask-and manufacturing cost have been and could be prohibitive.
However, with the general progress in technology, it pays to re-visit scenarios, and
here it is maskless, direct electron-beam writing of the gate layer of the ROM. With
a single beam, this has been a limited specialty in low-volume ASICs. But due to
the emergence of massively parallel maskless e-beam MEB lithography (Chap. 8),
this process for manufacturing giga-scale MOS nano-ROMs is feasible. We learn
from Chap. 8 that e-beam systems with >100,000 microbeams at <10 nm resolu-
tion will be introduced by 2012. Such a system will be fed by data rates >100 Tbit/
s. For the regular ROM pattern, the transistor grid is stored on the system, and only
1 bit will have to be transmitted per memory cell, namely an elementary shot or not.
Therefore, the e-beam direct-write programming speed can be 10 ms for 1 Tbit,
with no cost for masks and expenses just for machine time.
246 B. Hoefflinger
The capacity of the permanent ROM can be increased by 3D stacking of chips.

The lifetime of this digital document can be enhanced by complete sealing,
eliminating bond-related aging. Power could be supplied, and data could be read
by inductive coupling, making this device a digital Rosetta stone [14].
11.5 Alternative Non-volatile MOS-Compatible RAMs
So far, we have dealt with electronic charge as the token for memory content and
with MOS-memory types, which have been responsible for the relentless advance-
ment of electronics. The non-volatile RAM with possibly unlimited retention time
has continued to challenge researchers. Many solutions have appeared and will
appear. We address those that have potential to make it onto the market by 2020.
Ferroelectric RAMs (FeRAMs): The dielectric in these 1 T per cell memories is
lead zirconate titanate (PZT) or strontium bismuth tantalite (SBT), which change
their electrical polarization as a function of temperature with a large hysteresis.
Once they have been flipped into a state, they keep it over useful temperature ranges
with long retention times.
Magnetic RAMs (MRAMs): Currents are the tokens in programming and reading
the state of the cells in these memories via the electron-spin in a magnetic-tunneling
junction (MTJ). The spin-torque-transfer (STT) variety is the most advanced.
Resistive RAM (RRAM): With the promise of the ultimate density 4 F2 and
possibly multiple-levels per cell (MLC), this crossbar-type memory continuously
attracts inventors. The general direction is to place a material between orthogonal
word- and bit-lines, whose resistance can be changed by programming-current
pulses. In the simplest form, so-called half-voltage select is used for programming
(writing) and reading. An alternative is to insert a select-diode or select-transistor
together with a variable resistor. An example is the insertion of HfO2 between TiN
electrodes.
Phase-Change Memory (PCM): The material between the word-and bit-bars in
this memory is chalcogenide glass, which changes reversibly from low-resistance
crystalline to high-resistance amorphous when heated. An example is Ge2Sb2Te5.
The actual cell has a heater underneath and also a PNP bipolar transistor on top for
cell-select.
A 2010 overview of the status of these RAMs is given in Table 11.4.
Table 11.4 2010 status of non-volatile RAMs

Type Bit/chip Technology node Cell area Density Ref. (paper no.
at ISSCC 2010 [15])
FeRAM 576 kb 130 nm 41 F2 150 Mb/cm2 14.4
STT MRAM 64 Mb 130 nm 41 F2 150 Mb/cm2 14.2
RRAM 64 Mb 130 nm 10 F2 0.6 Gb/cm2 14.3
PCM 1 Gb 45 nm 19 F2 2.7 Gb/cm2 14.8
The 2010 status of the PCM is write speeds comparable to NAND-Flash,

however, with no data on write-energy, and random-access read times of 85 ns/b,
again without data on energy/bit or the equivalent, power efficiency. We have
entered the bit density in Fig. 11.10 and projected that with the indication that
these crossbar-type resistive structures should be scalable to 10 nm, with some
caution regarding the word-select devices.
Phase-change (PCM) and resistive (RRAM) memories are manufactured as
back-end-of-line (BEOL) additions to CMOS so that they are envisioned as possi-
ble multi-layered additions generating a 3D-extension potential of their own,
certainly with extreme challenges regarding power densities.
A chain of SETs (single-electron transistors), see Fig. 2.18, could form a serial
memory. This chain with each transistor holding in its Coulomb box 0, 1, 2,. . .
electrons under the condition of Coulomb blockade, would be a non-volatile
memory. A serial read-out as in a multi-phase charge-coupled device (CCD),
would yield its content but be destructive. Rewrite techniques as in the CCD days
40 years ago would have to be considered. At the 10 10 nm2 node, the Coulomb
energy would be about 20 meV, comparable with the electron thermal energy of
25 meV at room temperature. Therefore, cooling would still be required, but it
shows the potential for this memory concept. Including the tunneling junctions,
multi-phase clocking, and the re-write circuitry, this would offer a density of
10 20 nm2 for a potentially multi-bit per cell, non-volatile serial memory at a
hypothetical 5 nm node. However, contrary to the optimism until early in the first
decade of the new century, no practical circuits have been realized by 2010 so that
these innovations will become critical only after 2025.
Another revolutionary device is the memristor, which, as its name says, should
make an impact on the realization of memories. The memristor M
dF
M¼ ;
dq
dF
L¼ ;
dI
is a passive two-terminal device, and it can be considered as the dual of the

capacitor C:
dU
1=C ¼ ;
dq
dU
R¼ :
dI
While a capacitor stores charge as a function of voltage, the memristor stores

charge as a function of magnetic flux, and this flux results as the time-integral of
voltage:
248 B. Hoefflinger
ð
FðtÞ ¼ UðtÞ dt
This is a monitor of voltage over a certain time, reflecting its history as a time-
continuous analog value, which can mean multiple levels, i.e., multiple bits
per stored value. It is early in the exploitation of this device, and an obvious
interest arises in its application as a synapse in electronic neural networks
(Chaps. 16 and 18).
References
1. Raychowdhury, A., Geuskens, B., Kulkarni, J., Tschanz, J., Bowman, K., Karnik, T., Lu, S.-
L., De, V., Khellah, M.M.: PVT-andageing adaptive wordline boosting for 8 T SRAM power
reduction. In: IEEE ISSCC (International Solid-State Circuits Conference) , Digest Technical
Papers, pp. 352–353 (2010)
2. Nii, K., Yabuuchi, M., Tsukamoto, Y., Hirano, Y., Iwamatsu, T., Kihara, Y.: A 0.5 V 100MHz
PD-SOI SRAM with enhanced read stability and write margin by asymmetric MOSFET and
forward body bias. In: IEEE ISSCC (International Solid-State Circuits Conference) Digest
Technical Papers, pp. 356–357 (2010)
3. Wang, Y., Bhattacharya, U., Hamzaoglu, F., Kolar, P., Ng, Y., Wei, L., Zhang, Y., Zhang, K.,
Bohr, M.: A 4 GHz 291 Mb voltage-scalable SRAM design in 32 nm high-k metal-gate CMOS
with integrated power management. IEEE ISSCC (International Solid-State Circuits Confer-
ence) Digest Technical Papers, pp.456–457 (2009)
4. Kang, U., et al.: 8 Gb 3D DRAM using through-silicon-via technology. In: IEEE ISSCC
(International Solid-State Circuits Conference), Digest Technical Papers, pp.130–131 (2009)
5. Barth, J., et al.: A 45 nm SOI embedded DRAM macro for POWER7TM 32 MB on-chip L3
cache. In: IEEE ISSCC (International Solid-State Circuits Conference) , Digest Technical
Papers, pp. 342–343 (2010)
6. Schroeder, H., Kingon, A.I.: High-permittivity materials for DRAMs. In: Waser, R. (ed.)
Nanoelectronics and Information Technology, pp. 537–561. Wiley-VCH, Weinheim (2005).
Chap 21
7. Brewer, J.E., Gill, M.: Nonvolatile Memory Technologies with Emphasis on Flash. Wiley,
Hoboken (2008)
8. Bauer, M., et al.: A multilevel-cell 32 Mb flash memory. In: IEEE ISSCC (International
Solid-State Circuits Conference), Digest Technical Papers, pp. 132–133 (1995)
9. Trinh, C., et al.: A 5.6 MB/s 64 GB 4b/cell NAND flash memory in 43 nm CMOS. In: IEEE
ISSCC (International Solid-State Circuits Conference), Digest Technical Papers, pp. 246–247
(2009)
10. Young C.: 3 bit-per-cell NAND Flash ready for prime time.www.eetasia.com/
ART_8800596524_499486.htm.27Jan2010.
11. Maeda, T. et al.: International Symposium on VLSI Circuits 2009, Digest, pp. 122–123 (2009)
12. Saito, M., Miura, N., Kuroda, T.: A 2 Gb/s 1.8pJ/b/chip inductive-coupling through-chip bus
for 128-die NAND-Flash memory stacking. In: IEEE ISSCC (International Solid-State
Circuits Conference) Digest Technical Papers, pp. 440–441 (2010)
13. Chang, M.-F., Yang, S.-M., Liang, C.-W., Chiang, C.-C., Chiu, P.-F., Lin, K.-F., Chu, Y.-H.,
Wu, W.-C., Yamauchi, H.: A 0.29 V embedded NAND-ROM in 90 nm CMOS for ultra-
low-voltage applications. In: IEEE ISSCC (International Solid-State Circuits Conference)
Digest Technical Papers, pp. 266–267 (2010)
14. Yuan, Y., Miura, N., Imai, S., Ochi, H., Kuroda, T.: Digital Rosetta tone: A sealed permanent
memory with inductive-coupling power and data link. In: IEEE Symposium on VLSI Circuits,
Digest Technical Papers, pp. 26–27 (2009)
15. IEEE ISSCC (Internatioanl Solid-State Circuits Conferenc.) Digest Technical Papers,
Feb 2010
Chapter 12
3D Integration for Wireless Multimedia
Georg Kimmich
Abstract The convergence of mobile phone, internet, mapping, gaming and office
automation tools with high quality video and still imaging capture capability is
becoming a strong market trend for portable devices. High-density video encode
and decode, 3D graphics for gaming, increased application-software complexity and
ultra-high-bandwidth 4G modem technologies are driving the CPU performance and
memory bandwidth requirements close to the PC segment. These portable multi-
media devices are battery operated, which requires the deployment of new low-
power-optimized silicon process technologies and ultra-low-power design techniques
at system, architecture and device level. Mobile devices also need to comply with
stringent silicon-area and package-volume constraints. As for all consumer devices,
low production cost and fast time-to-volume production is key for success.
This chapter shows how 3D architectures can bring a possible breakthrough
to meet the conflicting power, performance and area constraints. Multiple 3D die-
stacking partitioning strategies are described and analyzed on their potential to
improve the overall system power, performance and cost for specific application
scenarios. Requirements and maturity of the basic process-technology bricks includ-
ing through-silicon via (TSV) and die-to-die attachment techniques are reviewed.
Finally, we highlight new challenges which will arise with 3D stacking and an
outlook on how they may be addressed: Higher power density will require thermal
design considerations, new EDA tools will need to be developed to cope with the
integration of heterogeneous technologies and to guarantee signal and power
integrity across the die stack. The silicon/wafer test strategies have to be adapted
to handle high-density IO arrays, ultra-thin wafers and provide built-in self-test of
attached memories. New standards and business models have to be developed to
allow cost-efficient assembly and testing of devices from different silicon and
technology providers.
G. Kimmich (*)
ST-Ericsson, 12, rue Jules Horovitz, B.P. 217, 38019 Cedex Grenoble, France
e-mail: georg.kimmich@stericsson.com

252 G. Kimmich
12.1 Toward Convergence of Multimedia Devices
The advancements in silicon process technology over the last few decades enabled a
significant increase of integration of about 2 transistors every 2 years as stated in
[1]. This increased density was enabled by smaller process geometries, which also
resulted in smaller transistor gate lengths and gate oxide thickness, which increased
the transistor and hence CPU (central processing unit) performances at similar rates
as the density.
Higher performance levels combined with integration of multiple functions such
as phone modem, CPU, video codecs (coder–decoders), graphic processors, image
and audio processing hardware along with multiple interfaces for audio, display and
connectivity into one piece of silicon are the basis for the convergence of multime-
dia devices.
The second important contribution for this convergence is the development of
the software from the low level software drivers through the operating system used
on the phone, up to the application software, which enables the user to run endless
applications on the devices, including office automation, e-mail, web browsing,
gaming, and mapping.
The sophisticated application software running on a fast processor and the
hardware implementation for multimedia co-processing together with ultra-fast
modem technology enable very rich combinations that are creating new forms of
communication through social networking, online gaming, and information sharing
on a local and global scale (Fig. 12.1).
ANY DEVICE
Phone
Mapping
Laptop MP3
Display
Web
browsing
DSC
Camera
Gaming
PDA
ANY WHERE ANY TIME
Fig. 12.1 Multimedia convergence

12 3D Integration for Wireless Multimedia 253
12.1.1 Multimedia Performance Parameters
Similar to the personal computer evolution, the multimedia processor application

software complexity leads to ever increasing requirements in CPU frequency,
memory access bandwidth, and memory density. On top of the application soft-
ware, the CPU can also be used to run some of the multimedia use cases such as
mp3 decode or video encode/decode.
The CPU performance is commonly benchmarked in D-MIPS (million
Dhrystone instructions per second). The base performance will be defined by the
processor architecture such as instruction and data bus width as well as the CPU
clock frequency. Clock frequencies passed the 1 GHz mark in 2010 and the
D-MIPS performance only a couple of years behind the laptop CPU performances,
a remarkable performance taking into account the stringent power and area
constraints for handheld devices.
To increase the CPU performance, there are multiple solutions, which are
optimized for the best trade-off between performance, software (SW) flexibility,
power dissipation, and silicon area. A good trade-off was achieved with the
introduction of the SMP (symmetric multi-processing) architecture which allows
multi-tasking application software to be run and dynamic power reductions for
lower MIPS use cases on a network of multiple CPUs.
To obtain the best multimedia performance, specific hardware, acceleration
circuitry or co-processors have to be implemented on chip. Audio decode and
encode is not very demanding in terms of MIPS and memory access bandwidth
but needs to be optimized for power to enable long mp3 or similar digital format
audio file playback times. Digital still camera functions require significant amounts
of post-processing for image quality improvements. The image sensor resolution
has quickly increased from 1 Mpixel for the first available camera phones to more
than 10 Mpixels for the smartphones released in 2010.
New HD (high density) display standards for increased video quality increase the
processing and memory bandwidth requirements. 1080p encode/decode with
30 frames per second are standard today on high-end smartphones, which require
a total of about 3.2 GB/s bandwidth to the external LPDDR (low-power dual data rate
RAM) memory. This evolution is not over yet; the frame rate will be doubled for
the next-generation devices to allow slow-motion replay, the display sizes and hence
resolutions will increase beyond the 1080p standard and, last but not least, new
graphic applications require graphic processors that will require more than 10 GB/s
of memory bandwidth alone to reach their full performance (Fig. 12.2).
12.1.2 Mobile Device Constraints
Portable devices need to meet, in addition to the performance standards, also

constraints for volume and power. Small device package footprints enable smaller
254 G. Kimmich
12 10
Ultra high performance to cope with HD 1080p
ever increasing multimedia 60 fps 40 9
10 performance and software
applications 8
>1GHz
Multi Core 7
8 CPU
HD 1080p 6
Gbits / s
D-GIPS
6 5
18
4
4 HD 720p 3
VGA
5 D-GIPS 2
2 3
Memory Access 1
Mpixels
0 0
2006 2007 2008 2009 2010 2011
Fig. 12.2 Multimedia processor performance requirements
form factor printed circuit boards (PCBs) to reduce the total size of the final
smartphone case. The same applies to the device thickness, which will need to be
kept to the very minimum to allow overall thickness reduction of the phone
application.
Mobile devices also need to have a very high mechanical reliability, which
requires the packaged devices to pass numerous tests, including “drop tests” to
ensure that they do not break or get disconnected from the PCB in case of shocks to
the phone case.
Battery life is a major competitive feature for a smartphone. The capacity of the
battery is very limited due to the small case volume and weight that can be allocated
to it. Complementary to a traditional low power design chip implementation, power
management techniques at system and device level need to be implemented to meet
the talk time, audio, and video playback targets. Microprocessor-controlled use
case management is used to switch entire functions, also called power islands, on
the chip on or off depending on their usage for a specific application or to limit the
frequency of the CPU or other critical circuitry when the full performance is not
required, a technique called DVFS (dynamic voltage and frequency scaling). The
silicon process needs to be optimized for device leakage to reduce the standby
current if the multimedia chip is idle and for low voltage operation to reduce
dynamic power. AVS (adaptive voltage scaling) methods are used to compensate
for silicon wafer process variations, by slowing down fast and leaky parts and
speeding up slow parts by applying transistor body biasing techniques. Chip I/O
(input/output) switching power, specifically on high-speed DRAM interfaces, still
account for a large amount of the overall power dissipation and can be significantly
reduced with the Wide-IO 3D memory attachment technology described in this
chapter. The material physics research promises the arrival of new battery
technologies with higher power densities and shorter battery charging times than
the currently established lithium ion technology, but none of these new
technologies is yet ready for mass production.
Another strong constraint for mobile devices is thermal integrity. The ultrahigh
performance of the application processor is pushing up the maximum power
dissipation of the chip to a level of largely above 2 W, which cannot be managed
without additional precautions at system, package, and PCB integration level. The
total thermal budget of the application is defined by the thermal conduction path
from the silicon to the application case. Constraints on the maximum silicon
temperature (Tj) are given by the silicon process reliability and the standby leakage
current, which increases exponentially at high temperatures. Common Tj maximum
values are in the range of 105–125 C. The application case maximum temperature
is around 40 C; above that temperature the device will be too hot for comfortable
handheld use. What is done in a PC or laptop by a chip- or case-mounted fan has to
be achieved in the mobile device by natural convection and thermal interface
materials to spread the heat from the SOC (system on chip) to the handheld case
surface.
Thermal system management needs to be implemented to ensure that the Tj and
Tcase temperatures will not exceed their specified limits. Software control can be
used to avoid launching several power-hungry applications in parallel. Thermal
sensors embedded close to the CPU or other high performance circuits are used to
monitor the temperature and to throttle the CPU frequency down in case of
imminent overheating.
12.1.3 Consumer Market Environment
The smartphone is situated at the high end of the mobile phone market and is
gaining strong momentum worldwide, quickly entering the mid range with a total
available market (TAM) of several hundreds of million phones per year in the next
few years. This leads to a very competitive environment in terms of product cost as
well as in terms of time-to-market pressure. To provide multimedia platform
solutions in this competitive environment it is inevitable to follow a very robust
and the same time flexible design process enabling the quick implementation and
verification of performance, power, and thermal requirements optimized for cost,
which are mainly driven by silicon area and package cost. Increasing software
complexity puts the software development on the critical path for the market
introduction of a new smartphone.
A high volume multimedia chipset platform production requires, owing to its
complexity, a very mature and efficient supply chain with robust operations and
testing to avoid quality and yield issues during production and to guarantee a high
level of reliability of the final shipped good.
256 G. Kimmich
12.2 3D Strategies
In the following section, we will explain the basic 3D-partitioning strategies that
can be considered for multimedia processors and their rationale.
12.2.1 Conventional 3D Packaging Techniques
3D device stacking has been common in the semiconductor industry for years.
Examples are shown in Fig. 12.3. The first applications using this technology were
DRAM and flash memory stacks, which have been in volume production for
numerous years using conventional wire bonding processes. Conventional die
stacking does not allow direct chip-to-chip connectivity and in the industry is
very often referred to as 2.5D packaging technology.
A similar process was used in some earlier application processor packages. This
die stacking technique helps to reduce the resistivity and capacitive load, which is
typical for the long trace line between the SOC and an external DRAM. It also helps
to reduce the overall PCB footprint area significantly. It was enabled by improved
assembly methods and the KGD (known-good die) memory test strategy and
business model. This packaging method presents some challenges to the supply
chain and because the chipset manufacturer takes the responsibility for the memory
test after assembly with the SOC.
Package on Package (PoP) Stacked Memory

DRAM,
Flash
Digital
Baseband
Stacked die in Package 2003 STM «world record»
DBB
NAND
DDR
Interposer
DDR
Memories
STn8815S22
Processor
Fanout Wafer Level
Package (FO-WLP)
Stud Bump
Flip Chip
Source: Prismark
Fig. 12.3 Conventional die stacking

The package-on-package (POP) stacking method combines the advantages of

the previously described stacked die in package technology with the traditional
supply chain and business model where the handset manufacturer undertakes the
final assembly of the product by stacking the standardized memory package, which
was previously tested, on top of the SOC package. POP stacking will remain
mainstream for the mid-range and high-end smartphone applications as long as
the memory bandwidth requirements can be satisfied with LPDDR DRAM
technology.
The fan-out wafer-level package (FO-WLP) technology, initially developed for
smaller die and package areas, provides a potentially competitive solution in the
near future. In the FO-WLP technology the Si die is placed into a reconstituted
plastic wafer. The die bump to package ball routing is performed with a back-end
metallization process, similar to that used on the silicon die. This process results in
improved routing density, reduced package height, and improved thermal
characteristics and provides a solution for memory stacking on the SOC die once
the dual side option becomes available at a competitive cost.
12.2.2 Wide-IO DRAM Interface
In most of the current smartphone designs the memory is stacked on top of the SOC
device via the POP technology described in the previous section. There are multiple
limitations inherent to the POP technology that will need to be overcome for the
next-generation smartphone and tablet applications. The main issue is that POP
limits the access bandwidth to the DRAM. The maximum bandwidth available with
LPDDR3 is limited by the interface frequency of 800 MHz, which cannot be
increased further without silicon process technology changes on the memory,
which would increase leakage and dynamic power. This frequency is also limited
by the interface signal integrity due to the parasitic external resistance and capaci-
tive load (RC) mainly induced by the package traces between the DRAM and SOC
device and their I/O buffers. The maximum data bus width that can reasonably be
implemented on a small mobile-compliant POP package with the typical size
of 12 12 and 14 14 mm2 are two 32-bit DRAM channels. This results in
a maximum memory bandwidth of maximum 12.8 GBps. Previous LPDDR2
configurations without POP limit the maximum bandwidth even further to about
3.2 GBps owing to the slower maximum interface frequency dictated by the
increased parasitic RC for external board traces between the memory and processor
chip. Another limitation for which the LPDDR concept cannot be further scaled up
is the power dissipation due to the switching of the I/O with relatively high
capacitive load in the range of 5 pF at the high frequency of up to 800 MHz.
Wide-IO 3D technology is based on the concept of massively enlarged (Wide)
data buses between the DRAM and the SOC. Since the total I/O count of the SOC is
already the determining factor for the package size and PCB footprint, a new
concept has been developed in which the memory die is directly stacked on top
258 G. Kimmich
Fig. 12.4 Wide-IO chip

stack
of the digital baseband (DBB) SOC using face-to-face or face-to-back die attach-
ment techniques, allowing thousands of interconnects between these chips as shown
in Fig. 12.4.
The Wide-IO memory interface is in the process of being standardized through
the JEDEC (Joint Electron Device Engineering Council) standardization body,
which was also the driver for the previously adopted memory interface standards
in the PC and mobile application world. This standardization will simplify the
cooperation between the chipset makers and DRAM memory suppliers because
they allow common electrical, timing, and physical specifications to be
incorporated in designs. This facilitates the compatibility between these devices
and allows second sourcing options between multiple memory vendors and
provides a healthy level of competition of memory suppliers on the same market
segment, which is required for high volume products. The physical locations of the
Wide-IO bumps are shown in Fig. 12.5.
The initial target of the Wide-IO standard memory bandwidth will be 12.8 GBps,
which will be reached with four memory channels running at 200 MHz each over a
128 bit wide data bus at single data rate (SDR). The hardware implementation of a
SDR interface is simpler, which means it will use less silicon area than the currently
used dual data rate (DDR) interface in which the data toggles at the rising and
falling edge of the clock signal.
In 2010, the single-die DRAM memory density available on the market is 1 Gbit.
Similar to logic chips, memory chips are following Moore’s law as set in [1] and
will continue to grow in density, which means the next-generation 4-Gbit devices
will become available in early 2012. The next-generation high-end smartphone
memory requirements for multimedia and application processing are in the range of
1–8 GB. To reach this high level of memory density – which is in the same range as
the PC market memory requirements today – engineers have come up with a new
idea, the stacking of multiple (2–4) memories via through-silicon via (TSV) and
Cu-pillar connections, as developed for the memory-to-SOC attachment. In the
memory industry this concept is also called “cubing”, as shown in Fig. 12.6. Data
on the maturity status, i.e., the ability to manufacture these cubes in high volumes
and at reasonable cost, are not yet publicly available. For that reason, system
architects are exploring alternative solutions to enable higher memory densities
for the next product generation. One identified solution to that problem is the
combined use of Wide-IO and the traditional LPDDR technique; The Wide-IO
will be used for the short latency, high bandwidth multimedia use cases, whereas
Fig. 12.5 Floorplan of wide-IO bottom die
Fig. 12.6 Wide-IO cube

stack
the slower and higher latency external LPDDR will be used for applications that do
not require high bandwidth. The drawback of this solution is the more complex
implementation of the application software, which needs to arbitrate the best use of
the memory resources.
Next-generation smartphones and tablet computers have memory access band-
width requirements that can no longer be satisfied with the LPDDR-standard. Ever
more sophisticated video gaming applications are pushing the graphics perfor-
mance to levels that require more than 15 GBps of memory bandwidth. These
graphic processors became available in 2010 but their full performance will only be
enabled if the memory interface can provide the necessary throughput. The rapid
260 G. Kimmich
market adoption of new tablet devices as well as the increased quality of TV home
displays is asking for very high video display resolutions on mobile chip sets either
integrated in or connected to these devices. The QXGA (Quad eXtended Graphics
Array) display standard defines a 3.1 Mpixel display, which together with the
requirement of an increased video frame rate of 60fps will almost saturate the
current LPDDR bandwidth, which has to be shared with other parallel-running
tasks on the application processor.
The Wide-IO technology is much more power-efficient than LPDDR, especially
at high bandwidth-use cases, as shown in [2]. A power dissipation comparison for
an external 16-bit DRAM interface with numerous Wide-IO data bus widths for a
video encode use case as shown in [3] is presented in Fig. 12.7. These simulations
show that even an eightfold increase in data bus width still results in lower total
power dissipation of the memory, due to the dramatically reduced RC of the bus
interface. The individual power management of the memory channels and banks
will further increase the advantage of the Wide-IO technology in an application
space where battery and thermal management is essential.
There are no immediately available alternatives on the market that can provide
the required DRAM bandwidth to the next-generation products. Scaling up the
LPDDR technology with POP is not an option due to pin count and power
considerations as shown previously. Nevertheless there are options to enhance the
standards by moving to low swing, single-ended, and terminated data on the DRAM
interface. Even if improvements can be made, there will still be a remaining gap to
reach the Wide-IO bandwidth at a reasonable power level. Other options on the
horizon are the use of serial interfaces between SOC and external memory. These
options are promising in terms of bandwidth potential, which can reach similar
levels as the Wide-IO concept. The drawback of the serial interface techniques are
on one hand the relatively higher switching power on the interface and on the other
hand the lack of standardization. Available serial proprietary technologies are not
taking off because of the business model, which creates too heavy dependencies on
the intellectual property (IP) and memory provider.
Fig. 12.7 Wide-IO power comparison

12.2.3 Technology Partitioning
Similar to the Wide-IO scenario, the idea of technology partitioning is to use the
best suited silicon technology per implemented function and then stack these silicon
dies together into one system. Instead of applying this concept only to the memory,
it can also be applied to the functions of power management and radio-frequency
(RF) transceiver technologies and others. The very expensive advanced CMOS Si
process technologies are often not very well suited for analog and high voltage
functions, which can be implemented at similar density in legacy CMOS process
nodes, which are much cheaper. Analog IPs hardly shrink beyond 40-nm process
nodes since they are very much dependent on the R, L, and C parameters, which do
not change significantly over high density CMOS silicon process node generations.
I/O area and electrostatic discharge (ESD) protection diode surfaces do not shrink
linearly with process technology advancements. The high voltage requirements for
the power management circuits and the legacy interfaces require thicker gate
oxides, which come for free with legacy process nodes and which need to be paid
for extra as additional gate oxide process steps in state-of-the-art CMOS processes.
Other than the potential cost optimization, there can be also some technological
advantages of such a 3D partitioning concept, in which, for example, the power
management unit (PMU) is closely connected via 3D to the circuitry that needs to be
supplied. Figure 12.8 shows the critical RLC path from the power management chip
to the processor. Shorter supply loops improve the power integrity and hence the
Problem Discription
Critical distance on printed circuit board (PCB). Can be very
long (several cm) when routing from backside to backside.
Using rule-of-thumb of 1 nH /1 mm a series inductor of up to
50nH can seriously impact PI.
Package (flipchip) Package (flipchip)

Die Ball Die
Pad Bump Ball Bump Pad
Vbat
DC/DC IP CPU
Ctrl
RDL PCB
RDL
Package
substrate
Fig. 12.8 Power integrity with conventional PMU/digital partitioning

262 G. Kimmich
Fig. 12.9 Silicon interposer
performance of these circuits. The implementation of decoupling capacitors to the

supply nets on the analog process node will help to improve the power integrity of
fast switching circuitry on the digital die. This split also allows better control of the
digital voltages, which can be lowered further to gain on dynamic and leakage power.
12.2.4 Other 3D Concepts
The 3D logic circuitry split is a concept that has been discussed for some time. The
underlying idea is the search for a breakthrough in further integration once Moore’s
law, described in [1], comes to its end at the boundary of the atomic limit, or before
that time to accelerate the integration by overlaying the 2D integration with another
dimension. There is currently no strong motivation to push for a logic 3D split for
the multimedia application processor space since the 3D split is more expensive
than the move to the next more advanced process node. Logic split could become
interesting for ultra high performance CPUs. One should not forget that this will
strongly increase the power density of these stacks, which will most likely require
new cooling techniques.
The silicon interposer or the silicon package, another, more conventional appli-
cation of 3D integration, is more likely to become interesting for the multimedia
application space. Silicon interposer designs are already available as prototypes.
Usually they consist of a relatively low cost substrate, which has a few levels of
metallization and integrated TSV connections, and multiple silicon components
placed side-by-side on top of and connected via the interposer. This concept allows
the same architectures as described in the Wide-IO or technology split but has the
advantage of better thermal performance with the drawback of increased package
size and cost. The silicon package concept shown in Fig. 12.9 takes it one step
further and replaces the organic package substrate altogether as shown. The cost of
silicon compared to the available organic substrates as well as the mechanical
properties of the silicon package are not yet competitive with the conventional
packages for the multimedia application processor space.
12.3 3D IC Technology
3D silicon stacking requires two fundamental key technology building bricks:

One for die-to-die attachment, which allows one chip to be connected vertically
to a second chip, and another one that is required to connect vertically through
Fig. 12.10 3D stack
a chip. New techniques are required to assemble the vertical chip stack
(Fig. 12.10).
12.3.1 Die-to-Die Attachment
We differentiate between two approaches: Face-to-face and face-to-back. In the

face-to-face approach the active sides of the two silicon dies to be attached are
facing each other. This approach has the advantage of a very low resistance and
capacitive load between the two silicon dies owing to very short connections. The
bottom die of the face-to-face stack can be connected to the package either via
traditional wire bonding techniques or as flip chip by adding vertical connectors
called through silicon vias (TSVs) through the silicon. The latter approach has the
advantage of being die size independent and it works also if the bottom die is
smaller than the top die. The flip chip approach also has major advantages for the
power distribution network implementation of the bottom die because the supply
pins of the PCB can be directly fed to the most power hungry circuitry on the
bottom die (Fig. 12.11).
The face-to-back approach consists in stacking the back side of the top die on the
active side of the bottom die. This is a more universal approach because it allows
more than two dies to be stacked on top of each other. The Wide-IO “cube”
technology uses the face-to-back approach.
Regardless of the die orientation approach, there is a need for an attachment
technology between the dies. The technology most commonly used in current
designs is called micro-bumping and uses copper pillars. Cu pillars are fine pitch
capable down to below 40 mm, which enables a very high interconnect density and
qualifies them as technology for the Wide-IO standard. Cu pillars have a high
264 G. Kimmich
Fig. 12.11 Cu pillar bump

cross section
current capacity and they allow various assembly strategies that guarantee a high
standoff height and thus simplify the underfill process required for the physical
attachment of the two dies. The Cu pillar technology is derived from the already
available die-to-package attachment methods, which are already in volume produc-
tion and can be considered as a mature technology.
Solder bumping, which is still very much in use for flip chip packaging, does not
allow a bump pitch below ~130 mm and as such does not qualify as high density 3D
interconnect technology.
The industry is looking into new technologies that are not yet ready for volume
production. Contact-less technologies, which are based either on capacitive or on
inductive coupling of the signals between the dies, should be mentioned. These
contact-less interconnect methods could provide an answer if high levels of vertical
stacking are required, but their maturity levels are still far from potential market
introduction with smartphones.
12.3.2 Through-Silicon Vias
3D silicon stacking requires a method to interconnect the dies vertically. This is

performed with TSVs. One can picture a TSV as a hole through the silicon die,
which is filled with metal and connected to the top and bottom metal routing layers
of the silicon. There are three established approaches to implement the TSVs in
silicon wafers. In the via-first process the etching is done before the FEOL (front
end of line), which signifies the transistor implementation. The via-middle process
implements the TSV after FEOL but before the metallization process (back end of
line, BEOL). In the via-last process the TSVs are etched as the final step of the
wafer process. The basic process steps for these methods are explained in
Fig. 12.12.
The via-first and via-middle process options are best suited for the high-density
interconnect requirements for the Wide-IO technology because they allow very
high via aspect ratios, which means the best trade-off between the via diameter,
which should be as small as possible, and the wafer thickness, which should be as
thick as possible for better manufacturability. Typical values for via-first and via-
Step #1 Step #2 Step #3 Step #4 Step #5 Step #6

BEOL 450 °C Thinning + De-Bonding
Via TSV Etch TSV Fill FEOL 1000 °C Backside prep
First
Vias are
made
before Handling carrier
CMOS Handling carrier
Thinning + De-Bonding
Via TSV Etch BEOL 450 °C Backside prep
FEOL 1000 °C TSV Fill
Middle
Vias are
made between
CMOS and Handling car
rier
BEOL Handling carrier
BEOL 450 °C Thinning TSV Fill + De-Bonding

FEOL 1000 °C TSV Etch Backside prep
Via
Last
Vias are
made after
BEOL Handling carrier
Handling carrier Handling carrier Handling carrier
Fig. 12.12 TSV process options (Courtesy of Yole)
middle are 10 mm diameter and 100 mm die thickness. The diameters will be further
reduced with increasing technology maturity. A very important parameter to moni-
tor is the parasitic resistance and capacitance value of the TSV, which is not only
linked to the aspect ratio, but also to the materials and the process used for the
implementation.
The via-last process variant remains advantageous for applications where the
diameter does not need to be below about 40 mm, because the process is robust and
cheaper to implement. Typical implementations of the via-last process are the
image sensors for camera modules shown in Fig. 12.14 used for small-form-factor
digital video and still camera sensor products, which have been in mass production
for several years.
12.3.3 Assembly
Depending on the application and multi-die-stacking requirements, the dies can be

assembled face-to-face or face-to-back as mentioned previously. For the 3D stack
assembly we also differentiate between the die-to-wafer and die-to-die attachment
techniques. Die-to-wafer attachment brings the advantage of more parallel
processing and the option of wafer testing of the stacked devices. The condition
to allow die-to-wafer stacking is to have a significantly smaller top die than the
bottom die. A more universal approach that is die size independent is the die-to-die
assembly process.
Multiple assembly methods for the Cu pillars, from solder to thermocompression
to ultrasonic bonding techniques, are available or under development. Most
common are conventional reflow techniques, which are being replaced by
266 G. Kimmich
Fig. 12.13 Die-to-die assembly flow

Fig. 12.14 Camera module
thermocompression and thermosonic methods, which allow finer pitches.

Figure 12.13 shows the assembly flow for a face-to-back, die-to-die scenario.
12.4 Design Automation
3D integration requires the development and use of new electronic design automa-
tion (EDA) tools. These tools need to be able to handle heterogeneous silicon
technologies and include models for the interconnect micro-bumps and the TSVs.
Exploration and planning tools are required to help to partition the top and
bottom die to optimize cost and performance and need to be capable of taking the
different performance, power, area, and cost properties of the involved silicon
process technologies into account. Electrical analysis tools are becoming more
important with increased integration, higher performance, and lower silicon process
voltages. Signal integrity simulation of multi-die systems for high speed interfaces
and power integrity analysis taking into account the complete RLC chain from PCB
to transistor, including the TSV and micro-bumps, are essential to guarantee the
required device performances. Improved thermal analysis tools and models are
necessary to ensure device reliability. A comprehensive thermal simulation for high
performance use cases will need to be performed at system level. This simulation
needs to include the thermal characteristics of the 3D multi-die interconnect and
the environment, specifically the package, PCB, and the case of the smartphone
application.
The 3D EDA tool chain is not in a very mature state yet since the R&D
investments of the EDA companies have been hesitant, based on the lack of 3D
industry adoption and hence the low return of investment on this type of tools. This
implies that for the first generations of 3D products there will still be some lack of
automation.
268 G. Kimmich
3D Tool User Interface 3D Thermal analysis

Techno 1 Techno 2 …
3D Thermo-mechanical
3D Partitioning
3D PI Analysis
3D netlist capture
3D SI Analysis
3D Stack definition
3D System assembly check

3D LVS check
3D visualization 2D Place & Route
3D floorplan Packaging
Courtesy of ST / CCDS
Fig. 12.15 Design automation for 3D design
Figure 12.15 shows the required elements and links for an integrated 3D design
framework and user interface.
12.5 Supply Chain Considerations
3D stacking is adding more integration process steps to the already complex process
of chip fabrication. This brings new challenges to meet the cost and time-to-market
requirements for these devices. During the manufacturing process the package
components will be processed by multiple companies in multiple locations. Before
the smartphone chipset manufacturer can delivery a final packed good to the phone
handset manufacturer, there will be process and test steps involving memory
suppliers for the stacked memory for Wide-IO products, the silicon foundry for
the SOC, the test site for the electrical wafer sort (EWS) process and outsourced
assembly and test (OSAT) companies for the wafer processing, assembly, and final
testing of the die stack involved. Figure 12.16 provides an overview of the process
steps and possible owners. New test, supply chain, and business models will need to
be worked out to cope with these new conditions.
12.5.1 Test and Failure Analysis
On top of the usual wafer test and final test steps of a system on chip, additional
steps are required for a 3D stack. First, each die will need to be tested separately
with a very high level of fault coverage to ensure a high yield of the separate
Fig. 12.16 Supply chain for Wide-IO stack
devices and to avoid the very costly aggregated yield loss after die assembly. The
combined value of multiple dies in one package can become significant; a failure of
one single element will cause failure of the complete system and overall impact the
business if the individual yield targets are not reached.
There are also new technical challenges that arise for testing of 3D circuits.
Ultra-thin-wafer handling is required since the TSV process requires wafer thin-
ning, which is also called back grinding, of the wafer into the range of 50–100 mm.
Wafers in this thickness range become very unstable and the equipment needs to be
adapted to be capable of achieving still fast testing without damaging the wafers.
Wafer thinning can also impact electrical parameters of the silicon and thus also the
performance.
Silicon-wafer probing is done with test needles that are placed on probe pads
which are on conventional silicon dies usually shared with the bumping locations
for flip-chip designs or wire-bond pads for package attachment designs. These
probe pads are placed in a pitch of typically 80 mm, arranged in only a few rows.
High-density 3D designs will connect through more than 1,000 pads in very dense
arrays of less then 50 mm. This will require the development of new more advanced
test equipment to enable fine-pitch probing. An intermediate solution to that issue is
the placement of probe pads with increased pitch in separate locations, which has
the drawback that the bump connectivity will not be tested before assembly.
For failure analysis of customer-returned failed parts, it is important to trace
back to the wafer lot, wafer number, and die location on the wafer of the failing die
270 G. Kimmich
Fig. 12.17 3D test flow
in the 3D stack. For conventional designs this is done by writing the main properties
of a die into a one time programmable memory (OTP), which can be read out after
assembly in the case of issues. In the 3D stacked configuration, there needs to be
also a way to identify the properties of the stacked devices; additional trace and
debug methods need to be implemented (Fig. 12.17).
Another challenge is the integration of heterogeneous technologies into one
package. Traditionally multimedia and application processing, power management,
RF transceivers, and DRAM functions are placed on separate chips, which are
tested separately. With the 3D technology some of these functions will be stacked
into one package and need to be tested together to ensure proper functioning of the
assembled system. The test equipment and test approach requirements for these
different technologies are not the same, which causes additional test steps and thus
costs. Improved, more versatile test equipment that is capable of running analog,
digital, and memory test sequences will need to be used. Other improvement steps
are to embed additional built-in-self-test circuitry (BIST) into the sSOC, which
enables complex memory tests run from the CPU of the application processor
through its memory interface. To enable specific direct access tests for the memory
on Wide-IO devices there are a number of reserved package pins, which are directly
connected to the memory through the TSV of the bottom SOC die, which can be
used by the memory supplier to perform some specific test sequences for production
testing as well as for failure analysis purposes.
Die-to-wafer assembly will bring an additional challenge to the wafer test since
the equipment will need to be capable of handling stacked wafers in case the chipset
maker would like to benefit from the possibility of an assembled wafer test prior to
the die singulation and packaging.
12.5.2 Business Model
In the case of mixed suppliers for the 3D stack, which is typical for the Wide-IO
systems, there is interest in new business models involving the memory suppliers,
chipset makers, and smartphone handset manufacturers involved in the design,
manufacturing, and testing of the chipset. Each company involved is interested in
optimizing its return on investment and in increasing the profit in its core business;
A memory supplier wants to sell memory; the chipset maker wants to sell the
system but does not want to buy and re-sell the memory because this represents a
big value that is not created inside its core business. Each process and assembly step
adds cost and value to the final good and each step includes risk of yield loss due to
lack of mature manufacturing processes or simply due to the complexity of the
processes. The preferred model of the memory suppliers and chipset manufacturers
requires the smartphone handset manufacturer to source all major components
similar to the conventional discrete chipsets as shown in Fig. 12.18.
In this model the handset maker accepts responsibility for the inventory man-
agement of all components. The chipset manufacturer, who is responsible for the
3D assembly, will draft the KGD DRAM memory wafers out of a consignment
cage, which can be managed either by the handset manufacturer or the memory
supplier, and assembles the final goods. The value chain has to be analyzed
carefully; the potential yield and resulting cost impact has to be quantified and
integrated into the KGD delivery contracts. The failure-analysis flow and liability
scenario in the case of customer returns of the finished goods has to be agreed by the
involved suppliers.
The KGD model was established for conventional 2.5D chip stacks. Known
good die means that each die on the wafer has been fully tested and failing devices
have been marked on the wafer.
The manufacturing, test, and assembly process steps are likely to take place in
several companies located in different continents and countries. For that reason, the
legal and tax framework has to be carefully studied to avoid issues with local tax
and custom regulations that could delay the assembly process significantly.
Fig. 12.18 Wide IO business

model
272 G. Kimmich
12.6 Outlook
Smartphones and tablet computers are still new on the consumer market, and many
generations of multimedia handheld devices are to come with the integration of
ever more features and the development countless software applications. New 4 G
modem technologies will provide download speeds of 100 Mbit/s or higher and
enable streaming of high quality video. Social networking software will become
more and more sophisticated and location-based services will be made possible by
the GPS and gyroscope circuitry integrated in smartphones. Augmented reality is
one of the upcoming applications in which graphics are overlaid on live video.
Head-mounted displays and virtual retinal displays (VRDs), in which the display is
projected onto the retina, will enable a whole new application space as well as new
advertisement techniques. Sensor and MEMS (micro-electro-mechanical systems)
integration will open entirely new doors to mobile medical applications and inter-
facing of wireless applications with the human body. Nobody can predict where this
development will end.
Very likely, the heterogeneous technology, memory, and logic stacking
techniques described in this chapter will be combined with the integration of
MEMS on one single silicon interposer to achieve the maximum performances
of all integrated functions.
All these applications will require more CPU processing power on a very small
area, which increases the power density further. The performance budget of
these future devices will be defined by the ability to keep the junction and case
temperatures of these devices down to guarantee lifetime reliability and user
comfort. New cooling techniques will need to be developed, based on new packag-
ing and thermal interface materials such as PCMs (phase-change materials) and
new packaging and product case design concepts.
Functional, electrical, and physical interface standardization and sound business
models are the key prerequisites for the high level of R&D efficiency required to
enable the take-off and evolution of the 3D silicon technology. Countless software
applications with as yet unimaginable functions will be available as soon as the
technology maturity allows these complex hardware designs to be built in high volumes.
References
1. Moore, G.: Cramming more components onto integrated circuits. Electron. Mag. 38(8), April 19
(1965)
2. Gu, S.Q., Marchal, P., Facchini, M., Wang, F., Suh, M., Lisk, D., Nowak, M.: Stackable
memory of 3D chip integration for mobile applications. IEEE IEDM (International Electron
Devices Management) 2008. doi: 10.1109/IEDM.2008.4796821
3. Facchini, M., Carlson, T., Vignon, A., Palkovic, M., Catthoor, F., Dehaene, W., Benini, L.,
Marchal, P.: System-level power/performance evaluation of 3D stacked DRAMs for mobile
applications. Design, Automation and Test in Europe 2009, DATE’09, pp. 923–928. IEEE,
Piscataway (2009)
Chapter 13
Technology for the Next-Generation-Mobile
User Experience
Greg Delagi
Abstract The current mobile-handset market is a vital and growing one, being
driven by technology advances, including increased bandwidth and processing
performance, as well as reduced power consumption and improved screen
technologies. The 3G/4G handsets of today are multimedia internet devices with
increased screen size, HD video and gaming, interactive touch screens, HD camera
and camcorders, as well as incredible social, entertainment, and productivity
applications.
While mobile-technology advancements to date have made us more social in
many ways, new advancements over the next decade will bring us to the next level,
allowing mobile users to experience new types of “virtual” social interactions with
all the senses. The mobile handsets of the future will be smart autonomous-lifestyle
devices with a multitude of incorporated sensors, applications and display options,
all designed to make your life easier and more productive!
With future display media, including 3D imaging, virtual interaction and con-
ferencing will be possible, making every call feel like you are in the same room,
providing an experience far beyond today’s video conferencing technology. 3D
touch-screen with integrated image-projection technologies will work in conjunc-
tion with gesturing to bring a new era of intuitive mobile device applications,
interaction, and information sharing.
Looking to the future, there are many challenges to be faced in delivering a smart
mobile companion device that will meet the user demands. One demand will be for
the availability of new and compelling services, and features on the “mobile
companion”. These mobile companions will be more than just Internet devices,
Reprinted by permission from “Greg Delagi: Harnessing Technology to Advance the Next-
Generation Mobile User Experience”, IEEE ISSCC 2010, Dig. Tech.Papers, pp.18–24, Feb.2010.
# IEEE 2010.
G. Delagi (*)
Texas Instruments, 12,500 TI Boulevard, MS 8723, Dallas, TX 75243, USA
e-mail: t-wright@ti.com

274 G. Delagi
and will function as on-the-go workstations, allowing users to function as if they

were sitting in front of their computer in the office or at home.
The massive amounts of data that will be transmitted through, to and from these
mobile companions will require immense improvements in system performance,
including specialized circuits, highly parallel architectures, and new packaging
design. Another concern of the smart-mobile-companion user will be that their
device is able to deliver an always-on, always-aware environment in a way that is
completely seamless and transparent. These handsets will automatically determine
the best and most appropriate modem link from the multiple choices on the device,
including WiFi, LTE, 5G, and mmWave, based on which link will optimize
performance, battery life, and network charges to deliver the best possible user
experience. In the future, adaptive connectivity will require many different
solutions, including the standard modem technologies of today, as well as new
machine-machine interfaces and body-area-networks.
All of the new and exciting applications and features of these mobile-companion
devices are going to require additional energy due to added computational
requirements. However, a gap in energy efficiency is quickly developing between
the energy that can be delivered by today’s battery technologies, and the energy
needed to deliver all-day operation or 2-day always-on standby without a recharge.
New innovations ranging from low-voltage digital and analog circuits, non-volatile
memory, and adaptive power management, to energy harvesting, will be needed to
further improve the battery life of these mobile companion devices.
Increased bandwidth combined with decreased latency, higher power efficiency,
energy harvesting, massive multimedia processing, and new interconnect
technologies will all work together to revolutionize how we interact with our
smart-companion devices.
The implementation challenges in bringing these technologies to market may
seem daunting and numerous at first, but with the strong collaboration in research
and development from universities, government agencies, and corporations, the
smart-mobile-companion devices of the future will likely become reality within 5
years!
13.1 Introduction
The mobile-handset market continues to be a dynamic and growing one, enabled by

technology advances that include increased bandwidth, greater processing perfor-
mance, increased power efficiency, and improved display technologies to deliver
compelling user experiences. We envision a world in 5 years where mobile devices
will offer new high-performance services and features, support always-on/always-
aware connectivity, and deliver battery life that will provide days of active-use
experience. These “smart-mobile-companion” devices of the future will be intelli-
gent autonomous systems with a multitude of incorporated sensors and display
options, all designed to make our lives easier and more productive (Table 13.1).
13 Technology for the Next-Generation-Mobile User Experience 275
Table 13.1 Features of the always-on, always-connected, always-aware mobile companion

• 3D display
• 3D imaging and gesture interface
• Integrated projection technology
• Object and face recognition
• Context awareness
• Noise cancellation
• Bio-sensing applications
• Adv. GPS location awareness
• “I see what you see” HD video-conferencing
• Adaptive connectivity
• Internet of Things
• The sixth Sense
• Brain–Machine Interface
Ten years ago, when mobile communication was voice-centric, who could have
imagined we could squeeze text and multi-media-centric mini-computers into a
mobile handset of 100 cm3?
Innovation drives new technologies in mobile handsets and is necessary to meet
the challenges of bringing the smart-mobile-companion device of the future to
reality. These numerous challenges include high performance, adaptive connectiv-
ity, and power efficiency.
The next generation of multimedia features, augmented with vision analytics and
3D displays, will enable users to control and interact with their mobile devices in a
more natural way. Mobile companions will be more than cell phones with internet
and multimedia capabilities. They will function as “work-anywhere” mobile
platforms integrated with all of the needed desktop applications. The massive
amounts of data processed by these mobile companions will require a 1,000
improvement in processing performance (power efficiency) from technology
advancements in device density, specialised circuits, highly parallel architectures,
and 3D integration.
The smart-mobile-companion device will deliver an always-on, always-
aware experience in a way that is completely seamless and transparent to the
user. These mobile companions will employ many different radio technologies on
the same device, including WiFi, GPS, BT, 3G, 4G, and mmWave. The handset will
automatically determine the best and most appropriate modem link to meet the
bandwidth requirements, adapt to the current environment, and provide context-
sensitive information.
Adaptive connectivity provides not only information links worldwide through
standardised wireless links, but also information connectivity to the mobile
companion’s immediate environment (Fig. 13.1).
This may include information about the user’s own body through a body-area
network, or data about the surrounding establishments on the street, in the building
or within the room, including other equipment and sensors. In a sense, we will be
connected to a few 1000 objects surrounding each of us in our daily lives, in a world
276 G. Delagi
Fig. 13.1 Evolving features of mobile handsets
connected through an “Internet of Things” [1, 2]. To meet the challenge of

delivering adaptive connectivity, technology advancements such as delivering
seamless connectivity, providing new machine–machine interfaces, and collecting
data from body-area networks will be required.
The demand of new and exciting applications and features of these smart-
mobile-companion devices will quickly drain the power available with today’s battery
technology [3]. In order to deliver days of operation and standby [4] without recharge,
technology advancements must meet the challenge of providing more energy-
efficient circuits and systems. New innovations, ranging from low-voltage digital
and analog circuits, zero-leakage standby, to energy harvesting, will be needed to
further improve the battery life of these smart-mobile-companion devices.
Research and innovation will be needed to meet these three main challenges:
– High-performance processing,
– Adaptive connectivity,
– Power efficiency.
This chapter examines the status of each challenge, and ascertains where we
need to be in the future, in order to deliver a compelling user experience within the
constraints of a smart-mobile-companion device.
13.2 The High-Performance Challenge
The smart-mobile-companion device of the future is not going to be the voice- and
text-centric device of the past, nor will it just be a multimedia and content-delivery
device of today. Instead, it will support exponentially-increasing computational
requirements for video, imaging, graphics, display controllers, and other media-
processing cores. Performance requirements for these functions will go beyond one
trillion (1012) floating-point operations per second (TeraFLOPS) or one trillion
operations per second (TeraOPS) in the next decade, as shown in Figs. 13.2 and
13.3, while power consumption for these functions must decrease.
As technology advances, a keyboard or even a touch-screen will no longer be
needed; you will use gesturing in the air to interface with your device. In the film
“Race to Witch Mountain” [5], 3D holographic user interfaces were shown as
control devices for a space-ship, allowing the children in the movie to fly the ship
without actually touching any part of it. In the future, we can look for mobile
devices to have similar 3D holographic controls, allowing the user to interact in 3D
space to manipulate data, play games, and communicate with others.
Yet another futuristic man–machine interface technology on the horizon is Honda’s
brain–machine interface (BMI) [6]. Using this technology, a computer is able to sense
what a person is thinking and translate that into action. In the future, users will be able to
Fig. 13.2 Performance trends in GOPS or GFLOPS over time

278 G. Delagi
Fig. 13.3 Relative complexity of a video system (MPEG4 coder/decoder, 30 frames/s), QCIF
(320 240 pixels) ¼ 1
interact with their mobile companion without touch or even gesturing. Combined with
3D-holographic displays, your smart-mobile-companion device will be able to sit on a
table, and you will be able to control everything about it using your brain alone.
Today, these man–machine interface concepts are moving from the silver screen to
the test labs. The future of the man–machine interface is rapidly evolving to the next
level with the help of technology advancements, and we expect that this interface will
become a standard means of communication between a user and the various
technologies, with which they interact. In particular, the BMIs discussed above
allow completely motion-free control of devices with direct brain-wave detection.
Voiceless communication is possible when sensors convert neuro-patterns into speech
[7]. But, there are other aspects of the man–machine interface as well:
The combination of free-space human interface using gesturing and projection
output onto a surface will enable the “view-and-interact” aspect of the future smart-
mobile companion to become much larger than the physical dimensions of the
device, allowing the always-aware mobile device in new ways. Together, these
technologies better enable users to get data into and out of the smart-mobile-
companion device, making it the center of their communications.
The future system-on-a-chip (SoC) must perform a wide range of computing and
multimedia-processing tasks and provide a true multi-tasking environment,
leveraging the concurrent operation of all the tasks associated with simultaneous
applications. The best approach for tackling these diverse, concurrent tasks is to
provide a range of general-purpose processors as well as multiple specialized compu-
tational elements for those tasks, where general-purpose processors cannot achieve
sufficient performance. The multimedia and computation functions must be mapped
onto the most appropriate processor to achieve the best performance/power ratios.
Complex-use cases will use many of these processors in parallel with data
flowing appropriately between the cores. Inter-processor communications and
protocols for efficiently distributing the processing among the available cores will
be needed. In some cases, dedicated hardware, to enable this computation “side-
loading”, will be beneficial for managing the real-time nature of the concurrent-use
cases, and for effectively abstracting the functions from the software operating
system running on the main processors. This functionality will be provided without
excessive frequency or excessive power being dissipated in any particular processor
or core. In addition, the future device must provide new features, capabilities and
services requiring significantly increased general-purpose computing power and
support for new peripherals, memory systems and data flow through the system.
With multiple input and output devices and 3D-holographic displays combined
with communication-intensive 3D graphics and imaging, as well as other
applications, the performance levels will continue to increase with future mobile-
device services and features.
The critical architectural choices and technology enablers will be
– Multi-core application processors,
– Specialised circuits providing parallel computations,
– 3D interconnects,
– Man–machine interfaces, and
– Managing energy and thermal limits.
13.2.1 Multi-core Application Processors
Increased general-purpose computing power will enable handsets to function

as our main computer at home or the office or as our main computer when we
are in a mobile scenario, achieving the “workstation-on-the-go” goal. Multi-
CPU processors will be required to provide the increased performance while simul-
taneously keeping the frequency at acceptable limits (Fig. 13.4). Today, most mobile
handsets are based on two processors operating in parallel at 4 DMIPS/MHz.
Fig. 13.4 Performance in MHz and MIPS

280 G. Delagi
13.2.2 Specialized Circuits Providing Parallel Computations
For demanding multimedia and other dedicated functions, specialized circuit

architectures and algorithms must be developed to perform computation in parallel
so that the exponentially increasing computational demand can be met within
achievable frequencies and acceptable power limits (Fig. 13.2). Figure 13.3
illustrates the relative complexity of a video core as it adapted to meet increasing
demands [18]. In this particular case, the core solution evolved over multiple
generations of mobile processors from a software-programmable DSP-only solu-
tion to a DSP-plus-hardware-acceleration solution and from there to a dedicated
video core with many computations performed in parallel.
Recently, extreme parallelism and specialisation have been applied in a video test
chip to demonstrate a video codec decoding a 720p/30fps video stream while
consuming only 2 mW of power [8]. Video is used here as an example, but new
services and features will require specialized circuits for other functions beyond
video. These will challenge designers and researchers to create specialized circuits
optimized to deliver the highest performance at the lowest possible power consump-
tion, meeting the consumer’s demand for high-performance services and features.
13.2.3 Thermal Limits
The hub of all this information exchange, compression and de-compression of

information, and information intelligence on the smart mobile companion is the
application processor. This device will continue to evolve to a solution that will
have multiple domains for specific functions. This type of partitioning will support
high-data-rate processing demands.
Within this design, the heterogeneous processors and cores will need to work
concurrently while having their own memory bandwidth, creating a need for
massive, low-latency, high-data-rate RAM for intermediate data and runtime stor-
age. Today, data rates between the application processors and external SRAM and
Flash are in the tens of Gb/s range. With increased demands for higher data rates,
we will see a more than 10 increase in 5 years to hundreds of Gb/s (Fig. 13.5).
It is also clear that, in addition to the external memory interface, on-chip data
communication among the multiple processors will require one or more intercon-
nection fabrics that can handle Terabits/s. Specialized low-latency, asynchronous
buses and low-power interconnection fabrics will be needed to achieve this 10
increase in data flow among the elements of heterogeneous systems.
A single-data-bit communication link between a processor and an external on-
board memory device at 1.5 V switches about 10 pF, resulting in 10 mW/Gb/s for such
a link. Similarly, using a serializer/deserializer (Ser/Des) driving a differential signal
over a 90 Ohm differential wire, also consumes about 10 mW/Gb/s including all
transmit (Tx) and receive (Rx) overheads in today’s systems. State-of-the-art
Fig. 13.5 Multiple connections to the application processor and associated data rates
published research shows that 2–3 mW/Gb/s is achievable when a number of parallel
links are designed together to share PLLs and other efficient supplies [9]. Yet, the goal
is to have another 10 reduction beyond today’s state-of-art technology (Fig. 13.6).
As we look into the future, we can expect that data will move at speeds of a Terabit/
s. Thus, the power spent per bit transmitted becomes a major challenge. Today, in a
typical smartphone with a volume of 100 cm3, the maximum power consumption,
using no mechanical air flow, is limited to about 3–5 W. Beyond that, the surface
temperature becomes uncomfortable for the device to be held in a hand or carried in a
pocket. With this power limit, the three key power-consumption areas are split
between the processor, the display, and full-power cellular and other wireless trans-
mission circuits. So each of these areas consumes anywhere from a fraction of 1 W to a
maximum of 2 W, depending on the use case. Specifically, for the processor, the data
transmission between different SoCs, memory, transceiver chips, and off-phone data
communications, consume 10–20% of the processor power budget. The above
calculations focus on peak-power budget, which is mainly due to temperature
limitations inside the smart-mobile-companion device. Reduction in power can be
achieved using circuit techniques, low-overhead efficient PLLs, and tuned processes.
Associated with the data communication on a mobile device is the complexity of
EMI (electromagnetic interference) radiation in the RF bands being used for
multiple wireless links. Use of optical connections for on-mobile communications
is another axis, where optics versus Cu can be used to reduce EMI side effects.
Supporting optical interfaces is another challenge to be met in delivering lower
power well below 1 mW/Gb/s.
282 G. Delagi
Fig. 13.6 Evolution of the communication power efficiency in mW/Gb/s. References [7] and [8]
in the figure refer to [9] and [19] in the text, respectively
13.2.4 3D Interconnect
Achieving the right level of performance is not just a circuit challenge, but also a
process and package challenge, amounting to a link-impedance increase and/or
voltage-swing reduction by 10 or more. Bringing external memories closer to the
processor will reduce the capacitance that needs to be driven, reduce noise levels, and
enable lower voltage swings. One way to achieve this is by using through-Si-vias
(TSVs) and die stacking for very wide data transmission between dies. Figure 13.7
illustrates the evolution of memory connectivity from wire-bonded stacked dies to
POP (package-on-package), to TSV stacking. The challenge will be to move 3D
interconnect from research into high-volume production. Chapter 12 provides further
guidance on 3D integration.
13.3 The Adaptive-Connectivity Challenge
Besides the high-performance communication needs, the smart mobile companion

of the future will be always-on and always-connected (Fig. 13.1). And, most
importantly, it will not only understand the user and the environment, but it will
also provide context intelligence, giving the right access to the right database at the
Fig. 13.7 Evolution of memory connectivity. Future 3D ICs will be a mix between different
memory processes, high-performance processors, low-leakage processors, and specialty interface
ICs (maybe optical) for a most optimal stack
right time. To achieve this adaptive connectivity goal, there will be at least a 10
increase in the next decade in the number of bits crossing various boundaries. Such
an increase in data flow poses a great challenge in terms of technologies used for
such data transmission, be it wireless, wired or optical.
Figure 13.5 shows the multiple data links that can be connected to the core
applications processor of the mobile companion. These links include cellular
wireless, home/office LAN wireless, point-point wireless, and Bluetooth, among
others. From the mobile companion to the cellular network, data rates will be
increasing from today’s few Mb/s to hundreds of Mb/s using LTE and other future
cellular-modem technologies. For comparison, within the confines of the office or
house, data rates using wireless local-area networks will increase from the tens of
Mb/s to Gb/s using mm-wave links.
13.3.1 Seamless Connectivity
As technology enables mobile devices to provide increased computational capa-

bility and connectivity to the Internet, advanced-application developers will have
284 G. Delagi
substantial MIPS and bandwidth at their disposal. However, for advanced func-
tionality, more than just increased performance is needed. Internet connections
will be enabled by a multitude of possible modem technologies on the mobile
device. The device will determine the most appropriate modem link to optimize
performance, battery life, network charges, and so on. This could switch between
3G, 4G, WiFi, mm-wave, and so on, at any time without the knowledge of the
handset user, making it completely seamless and autonomous. The mobile com-
panion will make connection decisions on its own based on the user’s preferences
and the mobile’s awareness of the context and the environment.
To enable seamless connectivity, where the mobile companion has context
awareness of the user’s behavior as well as environment awareness of where the
mobile is physically, will require multi-processing as well as low-power technology
to enable constant monitoring without draining the battery.
13.3.2 Machine–Machine Interface
For the smart-mobile-companion device to better serve its owner, it should know its
location, in what situation the owner is, and which services and features should be
provided automatically. Advances in the machine–machine interface, such as
vision analytics, can be a solution in the future. Using the phone’s cameras and
sensors, the mobile companion would automatically recognize where the user is.
The always-aware environment has already begun to be shown in film and in
research. For example, the sixth sense [10] was demonstrated in the film “The
Minority Report” [11] when Tom Cruise’s character was walking through a mall
and billboards, placed on the wall, recognized who he was and tailored each
advertisement specifically for him. In the future, this sixth-sense-like technology
will be used to make your mobile companion always aware of where you are, what
you are doing, and with whom you are interacting.
This sixth-sense system was demonstrated with real technology in March 2009
[10] with a wearable version. The demonstration showed how your handset will be
able to recognize people and give you information about them, as well as providing
you with a constant stream of data about things around you. While it may seem like
we will be in information overload, these mobile companions will be “smart”
devices that will learn about you, your likes and dislikes, so that all information
that comes to your attention will be relevant and needed.
To achieve these advanced machine-machine interfaces, research is needed on
different low-power sensors to measure and to communicate various body
measurements, such as the brain electrical activity using electroencephalography
(EEG) or a person’s head or eye direction. All of these sensor/camera/3D-imaging
devices can produce multiple data streams that can be interpreted into control
information.
13.3.3 Body-Area Networks (BAN)
The future smart-mobile-companion device could be used with body-area networks

(BAN) to interact with multiple biological sensors scattered over or within the
body. Depending on the device’s application, this biometric data may be used to
make appropriate and necessary suggestions to improve quality of life. The smart-
mobile-companion device can serve as the link between the biological sensors and
a wide-area network to automatically transmit health data to doctors or emergency
services. The enabling technology needed to deliver BAN will be ultra-low-power
sensors, ultra-low-power, low-bandwidth communication protocols and circuits,
and extreme system-security protocols.
13.4 The Power-Efficiency Challenge
As the amount and complexity of applications and features on the mobile compan-
ion increase in the future, the drain on the battery will be immense without the
“right” technology advances.
Over the past 35 years, technology scaling has halved the transistor area per
semiconductor-lithography node allowing the SoC designer to integrate more
features without increasing chip costs. With technology scaling, SoC designers
have also enjoyed the benefits of reducing the supply voltages and increasing drive
currents leading to reduced power and higher performance, respectively. In turn,
this has enhanced the user experience over time, with longer talk and standby times.
However, technology scaling comes with increasing process variations and
leakage currents. With smaller transistors, random doping fluctuations within the
channel affect transistor performance more and result in larger variations of tran-
sistor threshold voltage (Sect. 3.2) within the die as well as from die to die.
Traditionally, in older process nodes, the on-chip variation is accounted for with
either voltage or frequency margins, but in deep-submicron, the margins required
are becoming so large that they impact the ability to advance performance or lower
power. Additionally, with each process node comes an alarming relative increase in
leakage (¼ sub-threshold ¼ barrier-control) currents (Sect. 3.2, Fig. 3.14).
Technology-scaling benefits, coupled with architectural innovations in processor
design, result in the active-power dissipation per Mio. Instructions/s (MIPS)
continuing to decrease from mW/MIPS to mW/MIPS, that means by three orders
of magnitude, in 20 years, as shown in Fig. 13.8 for programmable digital signal
processors (DSPs).
During the period from 2000 to 2016, while the Li-battery technology is
increasing its energy density (Wh/cm3) by more than 40%, driven by package
and process optimisation, the conventional Li-ion technology will likely saturate
its energy density [4]. To reach the next level of energy-density gain will require
a shift to new chemistries, either with new materials or new technologies or both.
286 G. Delagi
Fig. 13.8 Power efficiency in mW/MIPS (nJ/instruction) for different processor architectures
Fig. 13.9 The energy gap between a 10 cm3 Li battery and the new mobile-companion energy
Even then, the net effect is a major energy gap (Fig. 13.9) for portable equipment.
This gap is the result of the linear improvement of the Li-battery technology and the
exponential growth of energy consumed in new consumer-use cases.
For mobile devices in the next 5–10 years, the goal will be to deliver at least a week of
operation from a single charge with battery efficiency increasing to deliver a capacity of
about 1–1.5 Wh. This will also require a more than 10 reduction in display power,
allowing the viewing of an HD movie. The new always-on, always-connected use
model will pose new challenges for standby operation. The energy consumed in
a conventional, voice-centric mobile phone is <0.75 Wh over 2 days. But for
a multimedia-centric device, it can consume 3.75 Wh of energy in the same duration
for just standby operation, an increase of 5 [3]. This is primarily due to the
exponentially increased leakage current as well as the complexity in logic gates
Fig. 13.10 Evolution of mobile-device capacity in minutes or hours, respectively
and on-chip memory for all the processing engines mentioned earlier. The smart-
mobile-companion device must also maintain the benchmark battery life in a
conventional, voice-centric talk-time and voice standby time (Fig. 13.10). With
the continued improvement in the quality of video and audio technologies, playback
time continues to increase, setting a new battery-life benchmark.
13.4.1 Adaptive Power Management
Among many innovations in the industry, SmartReflexTM power-and performance-

management technologies have been developed and have been in high-volume
production since the introduction of the 90nm OMAP2420 processor [12].
SmartReflex technologies include a broad range of intelligent and effective hard-
ware and software techniques that dynamically control voltage, frequency and
power, based on device activity, modes of operation, and temperature. SmartReflex
starts with co-optimisation between process and circuit technologies to reduce
memory-bit leakage by at least one order of magnitude. In the 65nm OMAP3430
multimedia-application processor [13], three major voltage domains are used, each
with associated 16 power domains. Together, these enable a critical trade-off of
frequency scaling as applications demand and power optimisation with fine-grain
voltage- and power-domain management. The net power reduction achieved at the
SoC level with autonomous software control is 1,200. With power management
turned to the SoC OFF mode, the SoC consumes only a few mA. Of the 1,200
power reduction,
– 4 comes from transistor-leakage optimisation,
– 300 comes from optimisation of circuits, SoC, and software.
288 G. Delagi
Details of the second generation of SmartReflex technology have been published

in [14] and [15], which focus on the active-mode power reduction. The RTA
memory achieves 4 active leakage reduction. When compared to the high-activity
work load with power management turned off, the sleep mode dissipates <0.1% of
the energy, achieving three orders-of-magnitude of energy reduction. Both
generations of the SmartReflex technologies increase the energy efficiency by
three orders of magnitude in both standby and active operations. Moving forward,
in order to close the energy gap (Fig. 13.9), another three orders of energy
efficiency boost will be required within the next decade.
A power reduction of 10 can be achieved through transitions to new process
technology including Hi-k metal gate. Another 100 will have to come from
additional research and innovations, including
– Ultra-low-voltage and -power circuits, logic, SRAM
– Non-volatile memory,
– High-efficiency DC–DC and A/D converters, and
– Parallel architectures.
13.4.2 Ultra-Low-Voltage Circuits
For the user-centric mobile companion of the future, we need to develop ultra-low-
voltage circuits in order to achieve orders-of-magnitude reduction in the power of
the key components. Operating at low voltage (VDD < 0.5 V) impacts both the
dynamic power and the leakage (DC) power. A hypothetical chip might include
low-voltage SRAM, NV(non-volatile)RAM (Chap. 11), ultra-low-voltage, ultra-
low power ADCs (analog–digital converters, Chap. 4), and distributed high-
efficiency DC–DC converters for each voltage domain on the chip. Properly
designed CMOS logic continues to operate, even when VDD is below threshold
(Sect. 3.3). However, the logic speed is very slow. VDD ¼ 0.5 V is a good compro-
mise between speed and ultra-low power. For 6T SRAM, operation at 0.5 V is also
possible (Sect. 11.1). For lower-voltage operation, 8T is an option.
On every interface to the world, there is a sensor or an antenna followed by
a signal path ending with either an ADC or a DAC (digital–analog converter).
Converters are fundamental blocks. Their resolution and power efficiency
determines the power and cost efficiency of that new sensor or wireless solution.
Fortunately, research is showing that, with the “right” converter architecture and
process advancements, ADCs are on a path for energy efficiency. As wireless
bandwidths extend to support increasing bit rates and more complex modulation
schemes, the power per conversion-step, an ADC figure-of-merit (FOM), must
continue on the reduction path to maintain a constant power per signal path, even
though the bandwidths and rates are increasing. The path shown in Fig. 13.11
indicates that, with the “right” choice of architecture, the FOM improves with
the process node, where today ADCs are well below 100 fJ per conversion step
(Chap. 4), while a 110db audio-quality DAC consumes <0.5 mW [16].
Fig. 13.11 Energy figure-of-merit (FOM) for analog–digital converters (ADCs) in pJ per conver-
sion step
In order to realise the savings in power that is enabled by 0.5 V operation,

a DC–DC converter is required to deliver ~10 mW of power with high efficiency.
LDOs, which are typically used for voltage conversion today, have poor efficiency
at low-output voltage. Switching regulators require an external inductor and can
have poor efficiency at low output power as well. Micropower management tech-
nology needs to be developed to deliver ~10 mA at an output voltage of ~0.5 V from
an input voltage of ~1.8 V with ~80% efficiency [17].
But low-voltage circuits by themselves are not enough to achieve the 100
power reduction that is needed for mobile devices of the future. Hardware
accelerators can also be used to perform high-throughput computation at low
voltage (Sect. 3.6). For mobile monitoring of medical diagnostics such as ECG
(electro-cardiogram), we need hardware accelerators for FT (Fourier transforma-
tion), FIR filtering, and CORDIC functions. Together, these have the potential to
reduce power by 100.
13.4.3 Zero-Leakage Standby
In many systems, SRAM-leakage power is a large fraction of the total power. For
many years, researchers have been searching for a practical non-volatile RAM
290 G. Delagi
(NVRAM) that dissipates zero-standby power because it retains its state when the
power is turned off. Magnetic RAM, Ovonic or phase-change RAM, or ferroelec-
tric RAM (FeRAM) are all candidates for NVRAM. As of today, none of these
technologies has proven to be scalable to very high densities. However, for
many ultra-low-power applications, ultra-high density is not required. For ultra-
low-power RAM in the range of 1–100 Mb, FeRAM has great potential.
Today’s newest technology is wireless power or inductive charging, which
allows mobile devices to be charged without plugging them into a wall. New
technologies in the future will allow true wireless power under certain conditions.
When the phone is in standby or monitoring mode, the power demands will
decrease significantly, and future devices will not depend on the battery, but will
use energy harvesting techniques (Chap. 19) to power the device. At these very low
power levels, it is possible to “harvest” energy from the environment to provide
power to ULP (ultra-low-power) circuits. For example, with energy harvesting, no
battery is required to make life-critical medical monitoring possible. There are
several sources, from which useful electrical power can be extracted, including
mechanical vibration, heat-flow from a temperature differential, RF energy in the
environment, and light energy. Figure 13.12 shows the practical power levels
that are achievable from these sources.
Energy levels can be realised from the mW range using thermal or piezoelectric
sources to the centiwatt cW level with electromagnetic sources (Chap. 19). With
the combination of energy harvesting and ULP ICs, future mobile devices could
have a near-limitless standby mode.
Fig. 13.12 Energy-harvesting power levels achieved by 2010

13.5 Conclusion
Users of future mobile-companion devices expect their devices to be always-on,

always-aware, to have extended battery-life, and to provide a highly adaptive
experience at very high performance levels of communication. This requires
broad innovation in technology. By using heterogeneous multi-processing, multi-
core processors, specialised circuits and parallel architectures along with 3D
interconnect and multi-chip partitioning, the high performance of these smart-
mobile-companion devices will be achievable while maintaining acceptable ther-
mal limits. An always-on, always-aware use case will require adaptive connectivity
that will be solved with machine-machine interfaces, seamless connectivity, and
body-area networks. While a 10 improvement in energy efficiency can be
provided by technology, another 100 improvement will have to come from
research and innovations in areas such as
– Ultra-low-voltage and –power circuits, logic, and SRAM (Chaps. 3 and 11),
– Non-volatile memory (Chap. 11),
– High-efficiency DC–DC and analog-digital converters (Chap. 4), and
– Parallel architectures.
New technologies, like TSV, optical interfaces, and efficient PLLs, will have to
be investigated to reduce the power level to well below 1 mW/Gb/s (Fig. 13.6), in
order to move around the massive amounts of data these new mobile companions
will require. The implementation challenges to bringing ultra-low-power and high-
throughput technologies to market may seem daunting and numerous at first,
but with the strong collaboration in research and development from universities,
government agencies, and corporations (Chap. 22), these smart-mobile-companion
devices will become reality within the next decade.
References
1. Santucci, G.: Internet of things – when your fridge orders your groceries. In: International
Conference on Future Trends of the Internet, Luxemburg, 28 Jan 2009
2. ITU Report on “Internet of Things” (2005)
3. Samms.: Evolving Li-Ion technology and implications for mobile phones, IWPC (2008)
4. Öfversten, J.: Mobile internet impact on battery life. In: IWPC (2008)
5. Walt Disney Pictures.: “Race to Witch Mountain”, 13 Mar 2009
6. Honda Research Institute Japan Co., Ltd.: Honda, ATR and Shimadzu jointly develop brain-
machine interface technology enabling control of a robot by human thought alone, News
Release, 31 Mar 2009
7. Texas Instruments Inc.: World’s first, live voiceless phone call made today at Texas
instruments developer conference, News Release, 26 Feb 2008
8. Sze, V., et al.: A low-power 0.7V H.264 720p video decoder. In: Asian Solid-State Circuits
Conference (A-SSCC), Fukuoka, Apr 2008
9. Palmer, R., et al.: A 14mW 6.25Gb/s transceiver in 90nm CMOS serial chip to chip commu-
nication, ISSCC Digest, pp. 440–441, Feb 2007
292 G. Delagi
10. Maes, P., Mistry, P.: Unveiling the “Sixth Sense”, Game-Changing Wearable Tech. In: TED
2009, Long Beach, Mar 2009
11. Dream Works/20th Century Fox Pictures.: “Minority Report”, 21 June 2002
12. Royannez, P., et al.: 90nm low leakage SoC design techniques for wireless applications,
ISSCC Digest, Feb 2005
13. Mair, H., et al.: A 65-nm mobile multimedia applications processor with an adaptive power
management scheme to compensate for variations, VLSI, June 2007
14. Gammie, G., et al.: A 45-nm 3.5G baseband and multimedia application processor using
adaptive body-bias and ultra-low power techniques, ISSCC Digest, Feb 2008
15. Carlson, B.: SmartReflex power and performance management technologies: reduced
power consumption, optimized performance, http://www.techonline.com/learning/techpaper/
209100225. Accessed Apr 2008.
16. Hezar., et al.: A 110db SNR and 0.5mW current-steering audio DAC implemented in 45nm
CMOS, ISSCC Digest, Feb 2010
17. Ramadass, Y., et al.: A 0.16mm2 completely on-chip switched-capacitor DC-DC converter
using digital capacitance modulation for LDO replacement in 45nm CMOS, ISSCC Digest,
Feb 2010
18. Meehan, J., et al.: Media processor architecture for video and imaging on camera phones.
In: ICASSP (2008), Las Vegas
19. Marcu, C., et al.: A 90nm CMOS low power 60GHz transceiver with integrated baseband
circuitry, ISSCC Digest, pp. 314–315, Feb 2010
Chapter 14
MEMS (Micro-Electro-Mechanical Systems)
for Automotive and Consumer Electronics
Jiri Marek and Udo-Martin Gómez
Abstract MEMS sensors gained over the last two decades an impressive width of
applications:
(a) ESP: A car is skidding and stabilizes itself without driver intervention
(b) Free-fall detection: A laptop falls to the floor and protects the hard drive by
parking the read/write drive head automatically before impact.
(c) Airbag: An airbag fires before the driver/occupant involved in an impending
automotive crash impacts the steering wheel, thereby significantly reducing
physical injury risk.
MEMS sensors are sensing the environmental conditions and are giving input to
electronic control systems. These crucial MEMS sensors are making system
reactions to human needs more intelligent, precise, and at much faster reaction
rates than humanly possible.
Important prerequisites for the success of sensors are their size, functionality,
power consumption, and costs. This technical progress in sensor development is
realized by micro-machining. The development of these processes was the break-
through to industrial mass-production for micro-electro-mechanical systems
(MEMS). Besides leading-edge micromechanical processes, innovative and robust
ASIC designs, thorough simulations of the electrical and mechanical behaviour, a
deep understanding of the interactions (mainly over temperature and lifetime) of the
package and the mechanical structures are needed. This was achieved over the last
20 years by intense and successful development activities combined with the
experience of volume production of billions of sensors.
This chapter gives an overview of current MEMS technology, its applications
and the market share. The MEMS processes are described, and the challenges of
MEMS, compared to standard IC fabrication, are discussed. The evolution of
J. Marek (*) • U.-M. Gómez

Robert Bosch GmbH, Automotive Electronics (AE/NE4), Postfach, 1,342, 72703 Reutlingen,
Germany
e-mail: Jiri.Marek@de.bosch.com; Udo-Martin.Gomez@de.bosch.com

294 J. Marek and U.-M. Gómez
MEMS requirements is presented, and a short survey of MEMS applications is

shown. Concepts of newest inertial sensors for ESP-systems are given with an
emphasis on the design concepts of the sensing element and the evaluation circuit
for achieving excellent noise performance. The chapter concludes with an outlook
on arising new MEMS applications such as energy harvester and micro fuel cells.
14.1 Introduction
Silicon is arguably the best-studied material in the history of the world to date.
Millions of man years of research have created a material with superior quality and
excellent control of electronic properties – this know-how is the basis of the
electronics industry, which drives the innovation of today’s world.
Besides the excellent electronic properties, silicon provides us with mechanical
properties barely found elsewhere:
– Three times the density of steel
– Four times higher yield strength
– Three times lower thermal expansion
Silicon is brittle; plastic deformation does not occur in monocrystalline silicon.
As a result, high-magnitude physical shocks less than the yield strength can be
applied to mechanical structures fabricated in silicon with no observable changes in
material properties. Overstress will destroy micro-electro-mechanical (MEMS)
structures; the destruction can be detected by a self-test.
As a result, silicon is the material of choice for critical safety applications prone
to high physical shock magnitudes, such as MEMS accelerometers used for crash
detection and subsequent automotive airbag deployment.
The combination of silicon’s superior electronic and mechanical properties
enabled the creation of completely new devices: micro-electro-mechanical systems,
or MEMS for short. Examples of MEMS devices (Fig. 14.1) include inkjet nozzles,
arrays of millions of gimballed micro-mirrors used for “digital light processing”
(DLP) projection systems, and the suspended spring-mass with comb-like capaci-
tive finger structures used to create an accelerometer. These seismic mass structures
are capable of measuring the rapid acceleration change experienced during an
automotive crash or by a laptop computer during a 1 g free-fall event.
Fig. 14.1 Acceleration sensor for airbag systems

14 MEMS (Micro-Electro-Mechanical Systems) 295
1967 1980 1980s

Resonant-gate MEMS Development of
transistor with scanning inkjet printers
metal-beam mirror (HP (Thinkjet),
micromachining (K. Petersen, Canon
500μm IBM) (Bubble Jet))
1970 1975 1980 1985 1990
First MEMS MEMS

sensors microphone
(Kulite, (G. Sessler,
Stanford, U Darmstadt,
Honeywell) Germany)
1970s 1980
1991 1994 1998 2008

Fully-integrated New DRIE- Volume First company
single-chip process production of with over 1
accelerometer (“Bosch- automotive billion sensors
(Analog process”) gyroscope produced
Devices) (Bosch) (Bosch) (Bosch)
1990 1995 2000 2005 2010
Prototypes of
“Digital Light
Tire-
Processing”
pressure
technology (TI)
sensor
1994
(Motorola)
Micromachined
silicon tuning- 2003
fork gyroscope
(Draper Labs)
1993
Fig. 14.2 History of MEMS
Initial MEMS micro-machining work was demonstrated in 1967 with singly

clamped metal cantilever beams serving as both an electrostatically excited beam
resonator and the gate electrode of an underlying silicon field-effect transistor [1].
Subsequently, the mechanical properties of silicon were intensely studied during
the 1970s, and researchers started to realize mechanical devices directly in silicon
[2, 3]. Figure 14.2 shows a short overview of MEMS history.
14.2 Today’s MEMS Market and Applications
The growth of the MEMS market and the market segmentation is shown in
Fig. 14.3. Today’s market size is around $6 billion with three main segments:
Other Emerging MEMS

10.000
Micro Fuel Cells
9.000 Flow Sensors
Thermopiles
8.000
Optical MEMS for Telecom
7.000 Microfluidic Chips
Millions of Dollars
RF MEMS Oscillators
6.000 MEMS Switches
5.000 BAW Filters

Microbolometers
4.000 MEMS Microphones
Optical MEMS for Displays
3.000
MEMS Wafer Probes
2.000 Gyroscopes
Pressure Sensors
1.000
lnkjet Print heads
0 Accelerometers
2006 2007 2008 2009 2010 2011 2012 2013
Fig. 14.3 MEMS market overview (iSuppli Corporation). BAW: bulk acoustic wave
Fig. 14.4 Detail of a digital light processing (DLP) chip (#Texas Instruments)
– Displays; dominated by TI’s DLP chip (Fig. 14.4)

– Ink jet nozzles
– Sensors (Fig. 14.1)
After the crisis in 2008/2009, the MEMS market is recovering very rapidly.
The driving force for sensors is the measurement of mechanical quantities for
(a) acceleration, (b) yaw rate/angular rate, (c) pressure, and (d) air flow.
The first wave of MEMS commercialization came in the 1980s and 1990s with
sensor applications in automotive electronic systems and with inkjet printers.
Around 2000, DLP chips joined in. Starting in 2005, the next wave arose with
new volume applications in the consumer-electronics market such as acceleration
sensors or microphones in mobile phones, laptops and game controllers.
According to iSuppli, the market segment of mobile and consumer electronics is
growing with a CAGR (compound annual growth rate) of 16.5% and will be the
largest one by 2013, accounting for 30% of the total MEMS market [4]. Main
applications are orientation/motion-sensitive user interfaces in mobile phones (e.g.
in Apple’s iPhone or Samsung’s Omnia) or in gaming stations (e.g. in Nintendo’s

Wii), vibration compensation in digital still-cameras and MEMS microphones.
14.3 The Challenges of MEMS Devices
14.3.1 Bulk Micromachining
MEMS technology is based on processes of the semiconductor industry. However,

new processes have to be added to produce mechanical structures in silicon. At the
same time, mechanical properties of silicon and the whole system have to be taken
into account, which are normally neglected by the semiconductor industry.
The first MEMS devices used bulk micromachining. Using anisotropic wet-
etching cavities, membranes and bridges can be etched into silicon (Fig. 14.5).
There are many different etchants with different etch rates along the various crystal-
lographic orientations of silicon. KOH is most widely used in the industry due to the
high etch rate as well as high anisotropy between the (100) and (111) directions.
Owing to the wafer thickness the device miniaturization is limited in bulk
micromachining since the complete thickness of the wafer has to be etched through.
14.3.2 Surface Micromachining
In order to shrink the MEMS devices even further, surface micromachining was
developed in the 1990s. The mechanically active layer is deposited onto the silicon
wafer covered with a sacrificial layer. The material is structured with high mecha-
nical precision; the deep reactive-ion etching (DRIE) process developed by Bosch
(also called the “Bosch process”) is capable of etching trenches with highly perpen-
dicular walls at a high etch rate. The sacrificial layer is removed, and the mecha-
nical structure is released (Figs. 14.6 and 14.7). For high reliability, these structures
are protected from particles using a second silicon wafer bonded as a cap-wafer
or by depositing a thin-film encapsulation.
Recently, we have invented the APSM (advanced porous silicon membrane)
process to create monocrystalline membranes for pressure sensors in a fully CMOS-
Fig. 14.5 Cross section of pressure sensor with KOH etched membrane
Fig. 14.6 Gyroscope fabricated with surface micromachining
Fig. 14.7 Details of the gyroscope structure
compatible process [5]. Monocrystalline silicon pressure sensor membranes have

been previously formed in most cases by anisotropic etching with an electrochemi-
cal etch-stop [6]. High-quality monocrystalline silicon membranes are produced
using the APSM process using the following process sequence: (a) anodic etching
of porous silicon in concentrated hydrofluoric (HF) acid, (b) sintering of porous
silicon at temperatures in excess of 400 C in a non-oxidizing environment [7],
and (c) subsequent epitaxial growth using the underlying rearranged porous silicon
as a single-crystal seed-layer.
14.3.3 Challenges of MEMS Production
MEMS requires the extension of the semiconductor technology into the world of
mechanical devices. Therefore several challenges arise:
(a) Special MEMS-processes: For the fabrication of MEMS additional dedicated

equipment is needed. Examples are tools for vapor etching with HF or ClF3 or
for DRIE.
(b) Contamination: Some MEMS processes constitute contamination risks to stan-
dard ICs, e.g. potassium originating from KOH can lead to severe drifts in
analog circuits. A stringent separation of standard CMOS production and
MEMS-specific production is therefore inevitable.
(c) Electromechanical system simulation: The MEMS devices need a completely
new design environment in order to simulate the mechanical structures in
silicon. Silicon has a low thermal expansion coefficient compared to the
materials used for assembly such as die-attach glue, mold plastics and die
pads. The simulation has to identify the influence of mechanical strain of
the package onto the MEMS device and optimize the device performance.
(d) Testing for mechanical parameters: A MEMS device has to be tested for its
mechanical performance: New test methods for pressure, acceleration or yaw
rate have to be developed for high volume manufacturing.
(e) Media protection: The devices must get in contact with various, partly aggres-
sive media in order to measure properties like pressure or mass flow. Despite
this exposure they have to work properly for years at high accuracy with failure
rates below the ppm range.
14.4 The Development of MEMS
Similar to the semiconductor industry, MEMS technology makes possible a

continuous shrinking of the devices. This development is demonstrated using
accelerometers for airbag application: Acceleration sensors decreased in size dur-
ing the last two decades by a factor of 5, enabling us to produce devices of smaller
and smaller package size; see Fig. 14.8.
Fig. 14.8 Evolution of accelerometer package size over time

1000 100
digital
# of analog transistors (thousands)

# of digital transistors (thousands)
800 analog 80
600 60
400 40
200 20
0 0
1995 2000 2005 2010 2015
Fig. 14.9 Evolution of ESP sensor ASICs (yaw rate + acceleration)
The evaluation circuitry evolved from bipolar to CMOS; nowadays more and
more digital signal processing is applied. The typical sensor ASIC contains only an
analog front end; the remaining signal evaluation is performed on digital level.
Figure 14.9 shows the evolution of the evaluation ASICs for an ESP inertial sensor
(yaw rate plus acceleration): the number of transistors in the analog part stayed
nearly constant at 10,000–15,000 but the number of transistors in the digital part
increased from 0 to ~1 million over a period of 10 years.
In the next two sections, we will describe two typical MEMS devices used
currently in automobiles at high equipment rates: yaw rate and pressure sensors.
14.5 Sensor Cluster (Yaw Rate and Acceleration)

for Electronic Stability Program
The vehicle dynamics control (VDC) system or Electronic Stability Program (ESP)
is one of the most effective active safety systems in the automotive domain,
assisting the driver to keep the vehicle on the intended path, thereby helping to
prevent accidents. A current description can be found in [8]. Since its introduction
in 1995, it has displayed a steadily growing installation rate in new vehicle
platforms, especially in Europe. Numerous studies show that this technology
drastically reduces the number of fatalities. In the USA, for example, a comprehen-
sive installation rate of ESP systems – as has been made mandatory by NHTSA
(National Highway Traffic Safety Administration) from 2012 on – could save up to
10,000 human lives and avoid up to 250,000 injuries every single year [9].
The first ESP systems used a macro-mechanical yaw rate sensor, which was
based on a piezoelectrically actuated, vibrating brass cylinder with piezoelectric
Fig. 14.10 First mechanical ESP yaw rate sensor
acceleration sensor
yaw rate sensor
Fig. 14.11 First micromechanical ESP yaw and acceleration sensor
material as a sensing element of the Coriolis force [10], for detection of the car’s
rotation along its vertical axis (Fig. 14.10). In addition, a mechanical single-axis
low-g accelerometer was applied to detect the vehicle’s dynamic state and for
plausibilisation of the yaw rate signal.
In 1998, as ESP systems were starting to gain a broader market share, Bosch
introduced its first silicon micromachined yaw rate sensor [11]. The sensing
elements were manufactured using a mixed bulk and surface micromachining
technology and were packaged in a metal can housing (Fig. 14.11).
Growing demand for new additional functions of ESP and of future vehicle
dynamics systems – such as Hill Hold Control (HHC), Roll Over Mitigation
(ROM), Electronic Active Steering and others – required the development of
improved inertial sensors with higher precision at lower manufacturing costs.
These goals have been achieved by the third-generation ESP sensors [12], a digital
inertial sensor platform based on cost-effective surface micromachining technology
(Fig. 14.12). Electromagnetic drive of the first generation has been replaced by
a simplified electrostatic actuation and sensing elements have been packaged in
Angular rate sensor

Accelerometer
μ-Controller
Fig. 14.12 The ESP sensor cluster MM3
sensing element
ASIC
Fig. 14.13 Yaw-rate sensor for sensor cluster MM3
a premolded, plastic package (Fig. 14.13). However, at that time the size of the
readout circuit prevented the combination of angular rate and acceleration sensors
in a small-size package.
Recent development at Bosch resulted in the first integrated inertial sensor
modules, combining different sensors (angular rate and low-g acceleration sensors)
and various sensing axes (x, y, and z) into one single standard mould package at
small size and footprint. In detail, the sensor consists of a combination of two
surface micromachined MEMS sensing chips – one for angular rate, one for two-
axis acceleration – stacked onto an application-specific readout circuit (ASIC) in
a SOIC16w package (Fig. 14.14).
14.5.1 Angular Rate Sensing Element
The angular rate sensing element is an electrostatically driven and sensed vibratory
gyroscope, which is manufactured in pure silicon surface micromachining, using
a slightly modified Bosch foundry process [13], with an 11-mm thick polysilicon
functional layer to form the movable parts of the micromachined structure
(Fig. 14.6).
Fig. 14.14 SMI540: The world’s first combined ESP inertial sensor (yaw rate and acceleration
sensor) in a mold package
Fig. 14.15 Operating principle of a yaw rate sensing element
The operating principle of the sensing element is depicted in Fig. 14.15. It

consists of two almost identical masses, connected by a coupling spring to ensure
common mode vibration shapes of the whole structure. Each part has a “drive”
frame on the outer circumference, followed by a “Coriolis” frame und ending at
a “detection” frame on the innermost position. All frames of each part are con-
nected via U-shaped springs, and the outer and innermost frames are also anchored
via U-shaped springs onto the substrate.
The drive frames are excited to a resonant vibration at approximately 15 kHz
by electrostatic comb drives anti-parallel along the x-axis. This drive motion is
Fig. 14.16 Block diagram of evaluation circuit of yaw-rate sensor
translated to the Coriolis frames by appropriately attached U-shaped springs,

whereas the detection frames are hardly affected by the drive motion, leading to
a decoupling of drive and detection motion at the detection frames. When an
external angular rate is applied around the z-axis, the Coriolis frames are driven
into a vibrating motion by the occurring Coriolis forces. This motion is translated to
the detection frames via U-shaped springs and picked up using parallel-plate
capacitor structures by their change in capacitance.
A scanning electron microscopy (SEM) image of the sensing element is shown
in Fig. 14.6. The application of U-shaped springs results in low mechanical
nonlinearities. The particular design with the feature of decoupled detection struc-
tures reduces mechanical and electrical crosstalk emerging from drive motion.
Mechanical crosstalk is additionally reduced by tightly controlled lithography and
trench etching, which are the critical MEMS manufacturing processes. Differential
signal evaluation and the high operating frequency of approximately 15 kHz lead
to sufficient insensitivity to external mechanical perturbation (vibration, shock)
usually present in automotive environments. High Q-factors of drive and detection
modes, due to the enclosed vacuum, lead to high bias stability.
14.5.2 Signal Evaluation Circuit
A block diagram of the signal evaluation circuit is shown in Fig. 14.16. It

consists of drive-control loop, detection loop, synchronous demodulator, signal
conditioning and self-test circuits. Except for the front ends, all of the signal
processing is performed digitally. The drive-control loop maintains a stable oscil-
lation of the sensing element by means of a phase-locked loop (PLL) to establish
the phase condition and an automatic gain control (AGC) to achieve constant
magnitude.
Measurement of the Coriolis acceleration is carried out by electromechanical
DS-modulators including force-balanced feedback loops. The analog front ends are
differential capacitance-to-voltage and flash A/D-converters. Early transition to the
digital domain provides for excellent offset stability. The micromachined sensing
element itself is part of the DS-modulators, giving a huge improvement to signal-
to-noise ratio (SNR) by applying noise-shaping techniques. Post-packaging
trimming of the angular-rate sensors is enabled by on-chip PROM and temperature
measurement. Angular-rate output is provided by a digital serial peripheral inter-
face (SPI).
Due to the demand of advanced vehicle-control-system safety and availability of
sensor signals, extensive internal safety and monitoring techniques have been
implemented in the sensor-system design. They consist of self-test and monitoring
internal features during start-up as well as operation; their status is presented at
the SPI interface. In addition, several internal parameters, such as temperature,
quadrature and current-control-loop quantities, are also output at the interface, thus
enabling the integration of safety and monitoring concepts into higher-level safety
and monitoring systems. In addition, the availability of the sensor’s internal signals
enables additional high-precision trimming.
14.5.3 Experimental Results: Sensor Performance
The sensor modules are designed to operate under usual automotive conditions.
This involves a temperature range from 40 C to þ120 C and external mechanical
and electrical perturbations (vibration and EMI).
The most important properties of inertial sensors and in particular of angular-rate
sensors are
– Noise-power spectral density, determining SNR and the limits of resolution,
– Offset drift of the sensor over operating temperature range and lifetime,
– Sensitivity of the sensor system to external mechanical perturbations (e.g.
vibrations).
Some typical measurements concerning these properties are discussed in the
following.
14.5.3.1 Noise Performance
The output-noise characteristic is shown in Fig. 14.17, measured on a typical

production sample. The inset displays the original time-domain data during
a measurement interval of 10 s at a sample rate of 1 kHz. Note that the noise
peaks are within approximately 0.1 /s. The main plot depicts the corresponding
Fig. 14.17 Output-noise characteristic of yaw-rate sensor
Sm , SDAC SOP , SkTC , SADC SQAD Φn,PLL
acor Rate
C/V- Digital Output
HBS ADC Adjustment
Converter Filter Filter
Coriolis
Evaluation Circuit
Sensing
Coriolis Detection Loop
Element
DAC
Fig. 14.18 Noise simulation model
noise-power spectral density. The characteristics of the sensor’s output low-pass filter
can easily be identified. The angular-rate signal bandwidth is from DC to approxi-
mately 60 Hz. Within this bandwidth, the output-noise density is typically about
0.004 ( /s)/√(Hz). At higher frequencies, the noise is attenuated by the output filter.
In order to determine the contribution of the particular noise sources, the design
of the ASIC and the sensing element, including parasitic effects and different noise
sources, is implemented in a simulation model according to Fig. 14.18. Using this
model, good agreement between experimental and simulated output noise could be
obtained, and the contributions of the different noise sources to the output noise
were evaluated as shown in Fig. 14.19.
Calculated relative contribution

Symbol Quantity
to the output noise
SOP Voltage noise of the C/V ~35%
converter
Sm Thermal noise of the ~20%
measuring element
SADC Quantisation noise of the ~20%
AD converter
SKTC KT/C noise of the C/V ~10%
converter
SQAD Quantisation noise of the ~10%
DAC converter
SDAC Noise of feedback force <1%
due to time-jitter and
amplitude variations
Φn, PLL Phase noise of the PLL < 1%
Fig. 14.19 Contribution of different noise sources
The simulation reveals three main noise sources:

SOP voltage noise of the C/V converter,
Sm thermal noise of the measuring element,
SADC quantization noise of the AD converter.
14.5.3.2 Offset Drift
Beside low noise and large signal bandwidth, high offset stability (over temperature
and time) is a prerequisite for the application of angular-rate sensors in advanced
vehicle-stabilization systems. Therefore, excellent offset stability is another key
feature of the sensor. A usual measure for the offset (in)stability over time is the root
Allan variance [14], which has been calculated from the measured rate-offset data
(Fig. 14.20, sample time 10 ms, measurement time 6 104 s, room temperature).
The slope at short cluster times corresponds to angle random walk (white noise).
The resolution noise floor is reached after ~600 s integration time with bias
instability of ~1.35 /h. Typical values are outstanding: <3 /h @ ~200–600 s. The
increase at larger cluster times (>1,000 s) indicates rate random walk (1/f noise
contribution) or an exponentially correlated process with a time-constant much
larger than 1,000 s. Furthermore, the rate-offset deviation over the whole tempera-
ture range is typically 0.4 /s.
The resolution limit <3 /h corresponds to amplitude changes in the micro-
mechanical structure as small as ~4 fm (10–15 m, about the size of the atomic
nucleus) or <2 zF (10–21 F). These outstanding numbers can only be achieved
by carefully controlling tolerances during the fabrication processes of the
Fig. 14.20 Root Allan variance of yaw rate sensor
micromechanical sensing element and by validating the capability and maturity of

the applied design concept.
The excellent noise and offset performance of these angular-rate sensors makes
them ideal choices, even for the highest demands of current and future VDC
systems.
14.6 Pressure Sensors Based on APSM Technology
New pressure sensor generations are based on a novel surface-micromachining

technology. Using porous silicon, the membrane fabrication can be monolithically
integrated with high synergy in an analog/digital semiconductor process suited
for high-volume production in an IC fab. Only two mask layers and one electro-
chemical etching step are added at the beginning of a standard IC process to
transform the epitaxial silicon layer used for IC circuits into a monocrystalline
membrane with a vacuum cavity underneath (Fig. 14.21).
14.6.1 Fabrication
The epitaxial layer is deposited on a porous silicon layer (Fig. 14.22). The cavity is
formed by subsequent thermal rearrangement of this porous silicon during the
epitaxial growth process and the following high-temperature-diffusion processes
[5]. The hydrogen enclosed inside the cavity during the epitaxial deposition diffuses
Fig. 14.21 Membrane fabrication technology: Top: Conventional bulk micromachining with
KOH etching and glass bonding to fabricate a membrane over a reference vacuum. Bottom: Porous
silicon technology: a monocrystalline membrane is formed by epitaxial-layer growth. The porous
silicon is converted into a vacuum cavity by sintering in H2 atmosphere
Fig. 14.22 Schematic fabrication steps: (a) substrate after anodic etching, (b) during sintering in
hydrogen atmosphere, and (c) after epitaxial growth
out, but no other gases can diffuse in. This leads to a good reference vacuum inside
the cavity (Fig. 14.23). The square membrane is deflected by 0.7 mm/bar. It is
mechanically robust against overload because of the “floor stopper” when it touches
the bottom of the cavity (Fig. 14.24).
A 1 bar sensor is overload-proof to more than 60 bar. The signal is sensed by a
piezoresistive Wheatstone bridge, which is a measurement principle with better
linearity than capacitive sensors. The piezoresistors are fabricated using the same
diffusion layers as the electronic circuit to achieve good matching properties. The
bridge signal of 55 mV/bar is amplified by the surrounding electronics. The mixed-
signal semiconductor process is suited for the automotive temperature range of
40 C to +140 C. The circuit is designed to be electrostatic discharge (ESD)-proof
and for electromagnetic compatibility (EMC).
membrane
vacuum cavity
substrate
Fig. 14.23 Pressure-sensor membrane with vacuum cavity under it and implanted piezoresistors
on top
Fig. 14.24 Sensor membrane with piezo resistors. The differential interference contrast of the
optical microscope makes the bending visible
14.7 Results
The sensor is digitally calibrated in four parameters: offset, sensitivity, temperature

coefficient of offset, and temperature coefficient of sensitivity. The calibration is
done in the packaged state to compensate for packaging stress (Fig. 14.25). The
digital calibration does not need openings in the chip-passivation layer (in contrast
to laser trimming), which makes it compatible with operation in a harsh environ-
ment. Results based on first engineering samples show that the sensor meets a total
accuracy of 1.0% over the pressure range of 200–1,200 hPa and from 0 C to
Fig. 14.25 Sensor chip wire-bonded in a premold package
Output voltage vs. pressure

4.8
+24.10 °C
4.5
+85.00 °C
4.0 +125.00 °C
+139.90 °C
Uout in V
3.5 +24.40 °C
–9.90 °C
3.0 –40.50 °C
2.5 +25.00 °C
2.0
1.5
1.0
0.6
180 400 600 800 1000 1200
pressure [hPa]
Difference to ideal curve vs. temperature

200
199.99 mbar
150 400.00 mbar
650.00 mbar
100 900.01 mbar
dUout in mV
1149.99 mbar
50
–50
–100
–150
–200
–40 0 50 100 150
temperature [°C]
Fig. 14.26 Measured output voltage versus pressure (top) and absolute error from 40 C to
140 C (bottom). The dashed line shows the 1.0% error band from 10 C to 90 C
110 C (Fig. 14.26). In the selected window for altimeter applications, the linearity
is significantly better.
14.8 Conclusion
The sensor chip is thin compared to KOH-etched pressure sensors because no

anodic bonding on the glass substrate is required. The piezoresistive pressure sensor
reported here looks like a standard IC chip. The easy packaging makes it suitable
for new consumer applications and tire pressure measurement systems, besides
the known automotive applications manifold-air-pressure and barometric-pressure
sensors.
14.9 The Future: MEMS Everywhere!
After automotive applications, MEMS is conquering currently the area of consumer

electronics. Acceleration sensors and MEMS microphones are contained in many
laptops, mobile phones and game controllers; gyros are following soon. In the near
future, we will see magnetic sensors for navigation, small projectors with MEMS
mirrors embedded in mobile phones or cameras (Fig. 14.27), or inertial units
containing acceleration sensors and gyros, including the associated algorithms.
Further on, MEMS energy harvesters, (Fig. 14.28) and micro-fuel cells for energy
storage, to provide wireless autonomous sensor nodes, will be launched. MEMS
sensors will enable applications never imagined before.
Fig. 14.27 Microprojector (#Microvision)

Fig. 14.28 Energy harvester (#IMEC)
Acknowledgments Many thanks to Gary O’Brien and Frank Sch€afer who contributed to the
preparation of this chapter.
References
1. Nathanson, H.C., Newell, W.E., Wickstrom, R.A., Davis Jr., J.R.: The resonant gate transistor.
IEEE Trans. Elec. Dev. 14(3), 117–133 (1967)
2. Angell, J.B., et al.: Silicon micromechanical devices. Sci. Am. 248(4), 44 (1983)
3. Petersen, K.E.: Silicon as a mechanical material. Proc. IEEE. 70, 420 (1982)
4. Global MEMS market suffers first-ever decline in 2008. iSuppli Corporation, 2009 (http://
www.isuppli.com/News/Pages/Global-MEMS-Market-Suffers-First-Ever-Decline-in-2008.
aspx)
5. Armbruster, S. et al.: A novel micromachining process for the fabrication of monocrystalline
Si-membranes using porous silicon. TRANSDUCERS 2003. In: 12th International Conference
on Solid-State Sensors, Actuators and Microsystems, Vol. 1, pp. 246–249. DOI: 10.1109/
SENSOR.2003.1215299
6. Kress, H.J., Bantien, F., Marek, J., Willmann, M.: Silicon pressure sensor with integrated
CMOS signal conditioning circuit and compensation of temperature coefficient. Sens.
Actuators A 25(1–3), 21–26 (1990)
7. Herino, R., Perio, A., Barla, K., Bomchil, G.: Microstructure of porous silicon and its evolution
with temperature. Mater. Lett. 2, 519–523 (1984)
8. E.K. Liebemann, K. Meder, J. Schuh, G. Nenninger: Safety and performance enhancement:
The Bosch Electronic Stability Control (ESP). SAE Convergence on Transportation Electron-
ics 2004, Detroit, MI, October 2004, Technical Paper 2004-21-0060. http://papers.sae.org/
2004-21-0060/
9. Electronic Stability Control (ESC) and Electronic Stability Program (ESP) automotive
applications. http://en.wikipedia.org/wiki/Electronic_stability_control. Accessed April 2011
10. Reppich, A., Willig, R.: Yaw rate sensor for vehicle dynamics control system. SAE Technical
Paper 950537. http://papers.sae.org/950537 (1995)
11. Lutz, M., Golderer, W., Gerstenmeier, J., Marek, J., Maih€ofer, B., Mahler, S., M€unzel, H.,
Bischof, U.: A precision yaw rate sensor in silicon micromachining. In: Proceedings of
Transducers’97, Chicago, IL, June 1997, pp. 847–850
12. U.-M. Gómez et al.: New surface micromachined angular rate sensor for vehicle stabilizing
systems in automotive applications”. In: Transducers’05: 13th International Conference on
Solid-State Sensors, Actuators and Microsystems, Seoul, June 2005, Digest of Technical
Papers, pp. 184–187. DOI: 10.1109/SENSOR.2005.1496389
13. Offenberg, M., L€armer, F., Elsner, B., M€ unzel, H., Riethm€uller, W.: Novel process for
a monolithic integrated accelerometer. Technical Digest Transducers’95 and Eurosensors
IX, Stockholm, June 25–29, 1995, Vol. 1, pp. 589–592. DOI: 10.1109/SENSOR.1995.717293
14. IEEE Std 952–1997, IEEE Standard Specification Format Guide and Test Procedure for
Single-Axis Interferometric Fiber Optic Gyros
Chapter 15
Vision Sensors and Cameras
Bernd Hoefflinger
Abstract Silicon charge-coupled-device (CCD) imagers have been and are a

specialty market ruled by a few companies for decades. Based on CMOS techno-
logies, active-pixel sensors (APS) began to appear in 1990 at the 1 mm technology
node. These pixels allow random access, global shutters, and they are compatible
with focal-plane imaging systems combining sensing and first-level image
processing. The progress towards smaller features and towards ultra-low leakage
currents has provided reduced dark currents and mm-size pixels. All chips offer
Mega-pixel resolution, and many have very high sensitivities equivalent to ASA
12.800. As a result, HDTV video cameras will become a commodity. Because
charge-integration sensors suffer from a limited dynamic range, significant
processing effort is spent on multiple exposure and piece-wise analog-digital con-
version to reach ranges >10,000:1. The fundamental alternative is log-converting
pixels with an eye-like response. This offers a range of almost a million to 1,
constant contrast sensitivity and constant colors, important features in professional,
technical and medical applications.
3D retino-morphic stacking of sensing and processing on top of each other is
being revisited with sub-100 nm CMOS circuits and with TSV technology. With
sensor outputs directly on top of neurons, neural focal-plane processing will regain
momentum, and new levels of intelligent vision will be achieved. The industry push
towards thinned wafers and TSV enables backside-illuminated and other pixels
with a 100% fill-factor. 3D vision, which relies on stereo or on time-of-flight, high-
speed circuitry, will also benefit from scaled-down CMOS technologies both
because of their size as well as their higher speed.
B. Hoefflinger (*)

316 B. Hoefflinger
15.1 Vision and Imaging Fundamentals
Visual information provided by images and video has advanced the most due to the
ever increasing performance of nanoelectronics. The first decade of the new century
offered a jump by three orders of magnitude in the product of pixels frame-rate to
the mobile user of camera-phones or digital cameras: Resolution went from 100 k
pixels to 10 M pixels and the frame rate from 3 fps (frames/s) to 30 fps.
It is interesting to note that the pixel size on camera chips followed a roadmap
with the same slope as that of its associated technology node with a magnifier of
about 20 (Fig. 15.1). We will see what enabled this progress, which subtle trade-
offs in expected pixel performance have been accepted to realize this scaling
strategy, and what the future challenges and opportunities will be.
As Dr. Theuwissen states convincingly in [1], there are attractive scaling rules
associated with the pixel size (pixel edge) p:
1. Pixel area. ~ p2
2. Chip area ~ p2
3. Chip cost ~ p2
4. Read-out energy/pixel ~ p2
5. Lens volume ~ p3
6. Camera volume ~ p3
7. Camera weight ~ p3
8. Sensitivity ~p2
9. Dynamic range (bright–dark) ~p2
As is typical for a chip-dominated electronic product, these economic factors

have led to more pixels/camera, increasing market opportunities to the incredible
volume of almost one billion camera-phones sold in 2010. However, there have
been subtle physical penalties regarding sensitivity and dynamic range.
The weaker performance of recent phone-cameras and compact cameras in
darker scenes can be noticed by more frequent activation of flash and/or double
exposure to compensate motion blur. Is this indicative of an inflection point in the
Fig. 15.1 The pixel size of

CMOS image-sensor chips,
their associated technology
node, and the minimum half-
pitch according to the ITRS
[1] (# 2010 IEEE)
15 Vision Sensors and Cameras 317
scaling of digital imagers? We study some basics to find out where an inflection
point for simple scaling may occur, and we discuss the key future opportunities for
advanced imaging and vision systems for everyone.
The basis for the success story of solid-state image sensors, replacing earlier
vacuum-tube diodes, is the Si p-n junction diode and its near-perfect photosen-
sitivity (Fig. 15.2). We see that in this photodiode, in spite of the heavy processing
overhead for an advanced CMOS process, every second impinging photon
generates an electron – hole pair, which can be recorded as an electronic charge
or current (50% quantum efficiency) over most of the visible spectrum as well as
into the near infrared (NIR) and into the ultraviolet (UV). How many electrons does
the photometric flux density, the illuminance, generate? To be quantitative:
1 greenð555 nmÞ1x ¼ 4 1011 photons cm2 s1

¼ 2 103 e mm2 s1 ðquantum efficiency 50%Þ:
Fig. 15.2 Quantum efficiency and spectral sensitivity of a 40-mm2 photodiode in a CMOS chip
versus photon wavelength [2] (#Springer)
318 B. Hoefflinger
Table 15.1 Sensitivity standards and photoelectrons in a Si photodiode (quantum efficiency 50%,
exposure time 25 ms)
ASA 100 200 400 800 1,600 3,200 6,400 12,800
DIN 21 24 27 30 33 36 39 42
mlx s 8 4 2 1 0.5 0.25 0.12 0.06
Signal Smin [e mm2] 16 8 4 2 1 0.5 0.25 0.12
Signal [e mm2 s1) 640 320 160 80 40 20 10 5
This sensitivity of Si basically can result in a minimum detectable illuminance

better than that of the human eye and considerably better than film. We can
calibrate a photodiode against the well-known sensitivity standards ASA and DIN
(Table 15.1). To gauge ourselves: Standard film is 100ASA with high-sensitivity
film reaching 800ASA.
Table 15.1 tells us what the minimum detectable signal Smin would be for a given
sensitivity. This signal has to be as large as the noise level, so that any extra electron
would make a just noticeable difference (JND). The dark limit is the leakage current
of the p-n junction, which has a dominant peripheral component around the edges
of the metallurgical junction against the environment and an area component,
which only shows up at large reverse bias voltages close to the breakdown of the
diode. In the context of photodiodes, this diode reverse current is called the dark
current. It is obviously the most critical parasitic component of a photodiode and
the real treasure in the proprietary know-how of a manufacturer.
The fundamental noise component in the photodiode is that of the variance of the
dark-current electrons ND [e mm1]. As in any observation of small numbers, the
standard deviation of ND is (ND)1/2, also called the dark-current shot noise. We can
determine the size L L of a photodiode to reach a certain sensitivity as follows:
Smin L2 ¼ ðND 4LÞ1=2 ;

2 2=3 1=3
L¼ ND :
Smin
This is the photodiode size L required for a given sensitivity Smin in a technology
with dark charge ND. The inverse is the sensitivity achievable with a given diode
size:
1=Smin ¼ 2ðND Þ1=2 L3=2 :
We show this relationship in Fig. 15.3 for three levels of dark electrons: 50, 5,
and 0.5 e/mm2.
Several lessons can be learned from Fig. 15.3:
– Minimize the dark current.
– Since dark current depends on perimeter, the penalty for scaling down the diode
size L is severe.
min. detectable
sensitivity dark current (e/mm/s) photo-current
2
(ASA) 20 200 2 103 (e/mm /s)
12800 5
6400 x 10
3200 x 20
1600 x 40
800 x 80
400 x 160
200 x 320
100 640
0 2 4 6 8 10 12 14 16
photodiode size (mm)
Fig. 15.3 Sensitivity levels in ASA and minimum signal charge [e mm–2] versus photodiode sizes
L and three levels of dark electrons
– Sensitivity can be increased with larger diode sizes, because the signal increases
with the diode area ~ L2, while the dark noise increases as only L1/2.
– The binning of n pixels, i.e., combining their signals into one, improves the
sensitivity only by n1/2.
– The binning of 4 pixels improves the sensitivity only by 2 (1 f-stop or
f-number), while a photodiode 4 larger improves the sensitivity 8 (3 f-stops).
The sheer drive for more megapixels in the product sheet by scaling down pixel
size severely reduces the pixel sensitivity. The customer should be told: Ask for
sensitivity first, then for the number of pixels! Furthermore, the resolution of lenses
poses a lower limit on useful pixel sizes. This resolution can be identified either by
the diameter of the image dot resulting from an original spot, the so-called Airy
disk, or the minimum resolvable pitch of line-pairs [3]. These characteristics are
listed in Table 15.2 for the f-stops of lenses.
How do we sense the electrons generated in the photodiode? In most sensors, we
integrate the charges for a certain integration time onto a capacitor, and we measure
the resulting voltage change on this capacitor. This is accomplished basically with a
source-follower transistor, serving as a voltage-to-current (or impedance) con-
verter, as shown in Fig. 15.4.
320 B. Hoefflinger
Table 15.2 Image resolution of lenses (line resolution for 9% contrast)

Lens Airy disk diam. [mm] Line pairs per mm Line-pair pitch [mm]
f/1.0 1.3 1,490 0.65
f/1.4 1.9 1,065 0.95
f/2.0 2.7 745 1.35
f/2.8 3.8 532 1.88
f/4.0 5.4 373 2.68
Fig. 15.4 The fundamental CMOS active pixel [2] (#Springer)
This active pixel basically consists of the photodiode, three transistors, and a
capacitor. The so-called linear charge-integration mode (Fig. 15.4 center) has two
phases: In phase 1, the reset, the first transistor is turned on, and the capacitor is
charged to VDD, irrespective of the photocurrent. In phase 2, the first transistor is
turned off, the capacitor is discharged for a certain time by the (photo-plus dark-)
current of the diode, and the resulting voltage is detected as the source voltage on
the source-follower through the third transistor, the select transistor. The dynamic
range of signals, which can be detected in this mode, is given by
(a) The dark-charge voltage loss in the absence of photo charge and
(b) The maximum charge, CVDD, on the capacitor.
It is on the order of 800:1.
In CMOS area-array sensors with rolling shutters, all pixels in a selected row are
reset simultaneously, integrated for the same (exposure) time, and, with all row-
select transistors turned on, the column-sense lines are charged by the source-
followers to their individual signal levels. All recent sensors have per-column
analog/digital (A/D) converters so that digital outputs are available for the whole
row at the same time. We illustrate the characteristics of this linear-type CMOS
sensor with an advanced 2010 state-of-the-art sensor [4] in Table 15.3.
If a 100% fill-factor is assumed, the photodiode area of this sensor is 2.7 mm2.
The sensitivity reported by Wakabayashi et al.then means that 1 photoelectron/mm2
is generated by 0.27 mlx s, a signal just exceeding the random noise of 0.6 e/mm2 so
that we would rate this sensor close to ASA3200. A SEM cross-section of this
sensor is shown in Fig. 15.5.
Table 15.3 2010 advanced CMOS sensor for imaging and video [4]
Pixel-array size 3,720 (H) 2,780 (V)
Technology 140 nm
Pixel size 1.65 mm 1.65 mm
4-pixel clusters, 1.5Tr./pixel, back-illuminated
Supply voltage 2.7 V/1.8 V
Saturation signal 9130e
Dark current 3e/s 60 C
RMS random noise 1.7e
Sensitivity 9.9e/(mlx s)
Dynamic range 71 dB
Output: 10b serial LVDS
Clock: 72 MHz
Pixel rates 10.3 Mpix at 22–50 fps
6.3 Mpix at 60 fps
3.2 Mpix at 120 fps
Power consumption 375 mW HD video
Fig. 15.5 Cross-section of an

advanced CMOS video sensor
with backside-illuminated
(BI) photodiode (PD) and
microlenses (MLs) [4]
(# 2010 IEEE)
Pinpointing an important future direction is the formation of the photodiode deep

inside the Si active layer and removing the backside of the wafer after completion
of the full CMOS process. This way, the photodiode can fill the whole pixel area
(~100% fill-factor), and the optical components such as color filters (CFs) and
microlenses (MLs) become most effective. The viewing angle of the photodiode
also is much larger than that for diodes inside the deep canyons of a multi-level
interconnect topography (Fig. 15.6). The other advantage of this structure is that the
back-end metal structures (at the bottom of Fig. 15.5) are ideal for the 3D intercon-
nect to chip stacks with further processing and memory capacity.
322 B. Hoefflinger
7, 0 μm 7, 0 μm 5, 6 μm 5, 6 μm
Thin Film
α - Si:H
4,0 μm Photodiodes
metal 4
metal 3
metallization
metal 2
layers
metal 1
Poly
active silicon active silicon
Standard CMOS Thin-Film-on-CMOS

30 percent fill factor, 90 percent fill factor,
diode area16 μm2 diode area 30 μm2
Fig. 15.6 Cross section of a standard CMOS active-pixel sensor (APS, left) and a 3D configura-
tion (right) with an amorphous Si photodiode on top of a finished CMOS structure [5]
An alternative that offers 100% fill-factor and the optimization of the diode is the
3D integration of an amorphous-Si photodiode layer on top of a CMOS wafer as
shown on the right in Fig. 15.6 [5]. Both topographies, back-illuminated and
3D-photodiode-on-CMOS, are indicative of future imager strategies.
15.2 Natural-Response and High-Dynamic-Range Sensors
The charge-integration-based sensors have a linear optoelectronic conversion func-

tion (OECF), as shown in Fig. 2.2 of [2]. Their fundamental problems are
– Limited dynamic range
– Poor contrast resolution in the dark
– White saturation in bright scenes
– Complex image processing because of luminance dependence of shapes
– Color distortion because of local luminance dependence
– Unnecessarily large bit rates
Many of these problems have been solved in still-cameras, and even in their
Video Graphics Array (VGA)-resolution video modes, with truly high-speed, high-
performance image processing on-board. A key example is dual or multiple expo-
sure to increase the dynamic range, to improve resolution in the dark, to avoid white
saturation, and to estimate correct colors. The adoption of certain routines and
standards has helped this a posteriori image improvement to impressive results and
to a well-established science [6]. The sophisticated approach is to take seven
exposures in order to cover a dynamic range of 24 f-stops. In the case of motion,
two or three exposures have to suffice, while the steady progress in motion
Fig. 15.7 The acquisition of high-dynamic-range images
compensation and multi-resolution recording/processing promises further improve-

ments along this avenue. Figure 15.7 illustrates this approach.
Natural and technical scenes can have a dynamic range of eight orders of
magnitude or 28 f-stops from a clear night sky at 1 mlx to direct sunlight at
100 klx and beyond, and our eyes manage this with a logarithmic response to
illuminance, like many other natural senses. This natural logarithmic-response
function has been adopted in digital voice codecs since the 1950s with the so-called
A-law or m-law for coders, which have several fundamental features:
– Constant contrast resolution independent of voice level, which means a constant
signal-to-noise ratio from very low voice-levels to very high levels,
– Natural bit-rate compression by compressing a 12-bit dynamic range to an 8-bit
signal.
The human eye has this logarithmic response (like our ears), with a high range of
over five orders of magnitude in its spontaneous response and another three orders
of magnitude in long-time adaptation (Fig. 15.8). A HDRC (High-Dynamic-Range
CMOS) log-response sensor with an input dynamic range of 24 bit and a 10 bit
output provides a natural bit-rate compression of 2.4:1.
In a comparison with the multi-exposure approach, which requires at least 7
8 ¼ 56 bit, the advantage of the natural (log)-response sensor is significant for
future developments, particularly video.
The inner pixel of the HDRC sensor is shown in Fig. 15.4 on the right. Gate and
drain of the transistor are connected, VGS ¼ VDS, so that this two-terminal transistor
works in the sub-threshold (barrier-controlled) mode with the current–voltage
characteristic (3.20)
ID ¼ I0 expf½ðA BÞVGS AVT =nVt g;

324 B. Hoefflinger
Fig. 15.8 The contrast-sensitivity function (CSF) of the human eye, indicating that a difference in
a local gray level of 1% is just noticeable. The contrast sensitivity of a HDRC sensor with a
dynamic range of 7 orders of magnitude (or 24 f-stops) and a digital 10-bit output is also shown to
have minimum detectable differences of 1.6% [2] (#Springer)
where ID is the sum of the photocurrent and dark current of the photodiode. The
voltage V on the capacitor follows the current over more than five decades as
V / ln ID ;
except at very low levels, where it settles at the dark-current level so that the
optoelectronic conversion function (OECF) has the form
y½DN ¼ a lnð1 þ xÞ;
and the contrast sensitivity function (CSF), the minimum resolvable difference in
percent, is
1x
DC ¼ :
ax
Here, y is given as a digital number (DN), and x is the illuminance relative to its
3-dB value, the illuminance where its CSF has increased to twice its optimum
value. Figure 15.9 presents the OECF and the CSF for a ¼ 100.
The contrast sensitivity, as a measure of the minimum detectable difference in
gray levels, establishes the fundamental quality of how uniform a field of pixel
images can become under a homogeneous illuminance signal S, and it enters as
quantization noise into the total flat-field noise. Other noise contributions are the
dark-current shot noise, discussed before, the signal shot noise, the fixed-pattern
noise, and the reset noise in charge-integration sensors (for a detailed treatment see
[3] and Chap.3 in [2]). The resulting signal-to-noise ratio is plotted in Fig. 15.10.
Fig. 15.9 The OECF (a) and the CSF (b) for an eye-like sensor with a ¼ 100
326 B. Hoefflinger
S/N dB
102 40
HDRC
10 20
Piecewise Linear
HDRC Linear
0
1
10–3 10–2 10–1 1 10 102 103 104 Lux
Fig. 15.10 Flat-field signal-to-noise ratio for a linear, a piecewise linear, and a HDRC sensor as a
function of the illuminance
The linear sensor at low levels is plagued by quantization and reset noise and at
higher levels by white saturation, resulting in a very limited dynamic range (300:1
in our example). For this reason, many linear sensors operate with a two-(our
example) or multi-step gain [1] in order to achieve a piecewise linear approxima-
tion to a log-response OECF. Even then, they remain limited in dynamic range by
white saturation (full-well electron capacity), 2,000:1 in our example. The HDRC
sensor starts with the best contrast resolution and, as a continuous-current sensor, it
is purely shot-noise-limited at low levels, free of reset noise. It maintains a natural,
constant S/N ratio of 70:1 with a 10-bit output over more than four orders of
magnitude.
The HDRC sensor appeared as a radical innovation in 1992 [7] with the
fundamental features of a natural, biomorphic vision system:
– Logarithmic response with unparalleled dynamic range (>140 dB at room
temperature), unlimited on the bright side except for lens limitations.
– A photometric continuous sensor, which means:
– No aperture adjustment needed,
– No shutter and no integration-time adjustment needed,
– Separation of luminance and (object) reflectance, making object and pattern
recognition very simple,
– Color constancy, superb for contrast and color management as well as tone
mapping,
– Gain control and white balance are offsets rather than multiplies,
– Very-high-speed read-out of regions of interest.
All these capabilities remain as challenges for any alternative future develop-
ment either as a continuing refinement of the HDRC sensors and cameras [2] or as a
radical alternative. A critical-size European research effort in this direction was

started in 2008 with the project HiDRaLoN [8].
For future high-performance focal-plane imaging and, even more for video
processing and compression, logarithmic sensor data offer significant advantages
(see Chaps. 11–14 in [2]), and the reduction of digital multiplies to the addition of
logarithms remains an all-time goal.
15.3 Stereo and 3D Sensors
The formidable repertory of image registration programs and of low-cost sensors

has helped in the design of digital stereo cameras both for consumer and profes-
sional use so that such systems including polarization glasses have appeared on the
market for the production of 3D movies as well as for photographers. Highly
compact consumer cameras are being designed using neighboring chip-pairs
directly off the wafer, which are automatically perfectly aligned, and prism-optics
to establish the proper stereo basis.
For the precise determination of depths in a scene, a base plate with three
cameras is being used [9]. A triangular arrangement with sensor distances of
~25 cm allows ranges of ~8 m with a depth accuracy of <10 cm. 400kpixel
HDRC sensors with global shutters are used because of their photometric, high-
dynamic-range capabilities, including disregard of luminance and safe edge detec-
tion as well as straightforward global synchronization, because there are no
concerns for individual integration times.
The generation of 3D depth or range maps is achieved with artificial lighting and
array sensors, in which the pixels resolve the delays of local reflections. Clearly, the
gains in switching speed achieved by scaling down CMOS to sub-nanosecond
responses have made various such concepts practical since the late 1990s. We
describe two such systems, one because it demonstrates the versatility of Si
technology, and the other because of its high-speed, high-performance capabilities.
The first system is based on a Si photodiode with a robust response to extremely
low signals, namely the arrival of single photons. The principle, which has been
used for a long time, is the multiplication of a primary photoelectron by secondary
electron – hole pairs resulting from the impact ionization of Si atoms in high-field
regions close to the avalanche breakdown of the diode. The naturally large variance
of the multiplication factors can be averaged in the case of pulsed lighting by
correlated detection and repetition of the measurement 100 or 104. A 32 32
pixel CMOS 3D camera with a 25-V single-photon avalanche diode (SPAD) in
each 5 T pixel was reported 2004 [10]. Owing to the high multiplication factor, the
detected signals could be routed directly to an external time-to-digital converter,
which had a resolution of 120 ps with a standard deviation of 110 ps. An
uncollimated laser source provided an illumination cone with a rate of 50 MHz, a
pulse length of 150 ps, a peak power of 100 mW, and an average optical power of
750 mW. For a life-size object at 3 m, and 104 repetitions, the range uncertainty was
328 B. Hoefflinger
Fig. 15.11 A 2D image and the 3D depth map and profile (in millimeters), taken with a 32 32
pixel time-of-flight (TOF) camera in 2004 [10] (#2004 IEEE)
1.8 mm or 0.06%. This mode would offer 5 frames/s. Figure 15.11 shows a 2D gray
image as well as a 3D depth map and profile taken with the same camera.
These 3D imagers are very important for the understanding and generation of
high-fidelity 3D scenarios as well as for intelligent operations in such scenarios, be
it people, health care, traffic, and, in particular, the interaction between man and
machines. The pressure for higher map resolution and higher frame rates will
continue, and our second system [11] is an example for the rate of progress: With
a resolution of 256 256 pixels, it generates 14,000 range maps/s using a sheet-
beam illumination and the so-called light-section method. Together with a row-
parallel search architecture, the high speed and a high accuracy was achieved.
Taking 10 samples/map, a maximum range error of 0.5 mm was obtained for an
object distance of 40 cm, a relative range accuracy of 0.12%.
The field of 3D vision will see exceptional growth for several reasons:
– Ever more powerful sensors
– High-density, power-efficient focal plane coding and processing
– 3D integration of optics, sensing, and processing
– The eternal fascination with the most powerful sense of our world
15.4 Integrated Camera Systems
The sensor is only a means to an end, which is perfect vision encompassing the
camera (or several cameras) as well as encoding, storing, and transmitting the
needed image information in the most effective way. At the one end of these system
requirements is the effective manufacturing of miniature cameras in a batch process
Fig. 15.12 3D wafer-level

integration miniature cameras
[12] (#2010 IEEE)
Fig. 15.13 Comparison of a video frame in standard definition video and in HDTV, and their bit
rates [13] (#KAPPA)
Fig. 15.14 One frame in a

high-speed video of a fire
breather [2] (#Springer)
including lens, sensor, and all necessary focal-plane operations. 3D wafer-level

integration of these elements is illustrated in Fig. 15.12, certainly the next big thing
for the cameras in the billions of mobile phones, but also for all minimally invasive
cameras in medical, technical-inspection, and surveillance applications.
At the other end, in the world of high-definition video (Fig. 15.13), we face the
big challenge of compressing the large bit rates, 1.5 Gb/s in our example or 2 to
330 B. Hoefflinger
3 higher in 3D high-definition, to rates that are compatible with future energy and
bandwidth limitations.
15.5 Conclusion
The availability and quality of visual information and sensation is the single
biggest roadblock or driver of our future digital world, both in wireless and
in the internet backbone (Chaps. 5, 13, and 20). The figure-of-merit will be
video quality/bandwidth. An attractive direction to pursue is photometric-log-
compression sensing and processing, producing frames as shown in Fig. 15.14.
References
1. Theuwissen, A.J.P.: Better pictures through physics. IEEE Solid-St. Circ. 2(2), 22 (2010)
2. Hoefflinger, B. (ed.): High-Dynamic-Range (HDR) Vision. Springer, Berlin/Heidelberg
(2007)
3. Janesick, J.R.: SPIE course “Introduction to CCD and CMOS imaging sensors and
applications”, SC504. SPIE Education Services, San Diego (2003)
4. Wakabayashi, H., et al.: A 1/2.3-inch 10.3Mpixel 50frame/s back-illuminated CMOS image
sensor. IEEE ISSCC (International Solid-State Circuits Conference) 2010, Digest Technical
Papers, pp. 410–411
5. www.ims-chips.de
6. Debevec, P., Patanaik, S., Ward, G., Reinhard, E.: High Dynamic Range Imaging. Elsevier,
San Francisco (2006)
7. Hoefflinger, B., Seger, U., Landgraf, M.E.: Image cell for an image recorder chip, US patent
5608204, filed 03-23-1993, issued 03-04-1997
8. www.hidralon.eu
9. www.pilz.com/safetyeye
10. Niclass, C., Rochas, A., Besse, P.A., Charbon, E.: A CMOS 3D camera with millimetric
depth resolution. In: Proceedings IEEE Custom Integrated Circuits Conference (CICC),
pp. 705–708. (2004)
11. Mandai, S., Ikeda, M., Asada, K.: A 256 256 14 k range maps/s 3-D range-finding image
sensor using row-parallel embedded binary search tree and address encoder. IEEE ISSCC
(International Solid-State Circuits Conference) 2010, Digest Technical Papers, pp. 404–405
(2010)
12. Jaffard, J.L.: Chip scale camera module using through silicon via technology. IEEE ISSCC
(International Solid-State Circuits Conference), Forum F1: Silicon 3D-Integration Technology
and Systems (2010)
13. www.kappa.de
Chapter 16
Digital Neural Networks for New Media
Lambert Spaanenburg and Suleyman Malki
Abstract Neural Networks perform computationally intensive tasks offering smart

solutions for many new media applications. A number of analog and mixed digital/
analog implementations have been proposed to smooth the algorithmic gap. But
gradually, the digital implementation has become feasible, and the dedicated neural
processor is on the horizon. A notable example is the Cellular Neural Network
(CNN). The analog direction has matured for low-power, smart vision sensors; the
digital direction is gradually being shaped into an IP-core for algorithm accelera-
tion, especially for use in FPGA-based high-performance systems. The chapter
discusses the next step towards a flexible and scalable multi-core engine using
Application-Specific Integrated Processors (ASIP). This topographic engine can
serve many new media tasks, as illustrated by novel applications in Homeland
Security. We conclude with a view on the CNN kaleidoscope for the year 2020.
16.1 Introduction
Computer architecture has been ruled for decades by instruction pipelining to serve
the art of scientific computing [1]. Software was accelerated by decreasing the
number of clocks per instruction, while microelectronic advances raised the speed
of the clocks (the “Moore” effect). However, algorithms became also more complex,
and eventually the “processor-to-memory” bottleneck was only solved by revealing
a more severe “algorithm-to-processor” bottleneck, a “more than Moore” effect [2].
The algorithmic complexity was first raised in communications, and recently we see
a further need in new media tasks. As the microelectronic advance has bumped into
L. Spaanenburg (*) • S. Malki

Department of Electrical & Information Technology, Lund University, P.O. Box 118, 22100
Lund, Sweden
e-mail: Lambert.Spaanenburg@eit.lth.se; Suleyman.Malki@eit.lth.se

332 L. Spaanenburg and S. Malki
the “power wall,” where merely speeding up the clock makes little sense, new ways
to accommodate the ever-growing algorithmic needs are demanded.
A typical example of a complex algorithm is the neural network. Though
originally inspiring von Neumann to computing structures, the neural network
had little life for decades [3]. It saw rejuvenation in the mid 1980s to demonstrate
the power of the upcoming supercomputers. Large-scale correlation on physical
datasets demanded a “more than Moore” acceleration to overcome the principle
problems in the highly interconnected computations. Going from the urge to
partition the software for parallelization on the underlying computational platform
to the inherent parallelism of ASICs was then just a small step.
A cellular neural network (CNN) is a mathematical template with interesting
properties. Especially, it allows nonlinear dynamic behavior to be modeled. A well-
known example is in the Game of Life (GoL) [4]. But the apparent simplicity is
deceptive for such examples where the simple binary cellular GoL topology is
sufficient. Where realistic examples with reaction/diffusion behavior come in, one
sees that the local connectivity brings a tide of data hazards, producing a surplus of
processor stalls and greatly reducing the potential speed of a classical central
processing unit (CPU). Specialized fine-grain hardware seems to be the remedy.
However, in [5] it is argued that the performance deficit remains due to the
presence of global control. The computing is limited to the single cell, and many
cells will make the task faster in proportion. Global control will destroy this perfect
scalability by the need for long control lines. Long lines are slow and system
performance is dominated by the longest lines. It seems that the cellular network
will have a close-to-constant performance at best, when scaling.
As a typical media application, vision offers interesting applications because
the basic neural equation fits nicely with low-level pixel operations. Analog
realizations allow a focal-plane processor: a tight integration of light sensors with
on-chip low-level operations to reduce the off-chip bandwidth. Parallelism between
many small nodes is a given, but programmability is still needed. This led
first to analog [6] and then to digital realizations. With scrutiny of number
representations and on-chip network communication, the digital realization has
slowly evolved from a simple pipelined CPU to a tiled application-specific integ-
rated processor (ASIP) [4].
The many-core architecture with thousands of tiles was developed, but it showed
quickly that the classical compilers are failing by a margin. In the early days, the
sequential programming paradigm was expanded with partition and allocation
methods, but gradually the awareness is growing that other programming
paradigms on different abstraction levels are required. It is not even clear what
algorithms are to be preferred, and how such algorithms should be implemented.
Therefore, we will look here into the scaling potential of cellular-neural-network
applications to establish a better understanding of the new phenomena.
The chapter discusses the history of CNNs as a gradual development to
a many-core implementation. From a review of the basic neural design equation,
it first reviews some of the analog and acceleration systems. Hence we show the
communication needs in a digital implementation. Subsequently, we introduce the
16 Digital Neural Networks for New Media 333
application-specific CNN processor and treat the need for new technology to
create the CNN platform for new media applications.
16.2 Some Basic Theory
A neural network is, in its most general format, a system of connected elementary
processing units (PUs), which for historical reasons are called neurons. Each neuron
is associated with a real numerical value, which is the state of the neuron.
Connections between neurons exist so that the state of one neuron can influence
the state of others. The system operation can be programmed by weighting each
connection to set the degree of mutual influence. Each processing unit performs the
following operations: (a) the incoming signals x are collected after weighting, (b) an
offset y is added, and (c) a nonlinear function ’ is applied to compute the state [7].
A single neuron is depicted in Fig. 16.1. It states the series of computations when
going from left to right. The offset y is modeled as the weighted contribution of
a source with value 1 and therefore not explicitly shown in the remainder of
this chapter. Creating a network of neurons can be done in numerous ways, with
different architecture for different applications. The system gets a specific function
from the structure and the weight settings: the structure allows for general func-
tionality, while the weights set the specific function.
The neural architecture represents the network function globally and graphi-
cally. The network function (network transfer function, mapping function, or
“answer of the network” are equivalent terms) accomplishes the main task of the
network: to adapt to associate the questions posed to the network with their
X1
θi
W1
Xj
Wj
∑W1:j
Y
F(∑k, θi) ϕ(F, λ)
∑Wl:n
Wl
Xl
Wn
Xn
Fig. 16.1 The operation of a single neuron

answers. “Questions” and “answers” are terms borrowed from the human associa-
tion process. As widely used alternatives, the terms input and output examples are
used: examples are shown to the network of how the problem to be solved behaves;
the input example (the stimulus) is shown together with the adequate reaction to it –
the response or the output example. This process is performed in the forward pass
through the network: the outputs are obtained as a response of the neural system to
the input (question) stimulation. Thus the forward pass through the network
evaluates the equation that expresses the outputs as a function of the inputs, the
network architecture, the nodal transfer, and the parameters so that, during the
backward pass, the learning algorithm can adapt the connection strengths.
For the multilayer feed-forward neural network (Fig. 16.2), the input vector x
and the scalar output y are connected via the network function f, as described by
! !!
X X
f ð~ ~Þ ¼ ’
x; w wji ’ ’ wik xk (16.1)
i k
where ’ is the nodal transfer and wmn denote the different weight connections
within the network architecture with indices according to the direction of informa-
tion transport. The nested structure of the formula represents the steps that the feed-
forward track of the learning algorithm has to pass.
The nature of the nonlinear neural transfer function has not been specified so far.
Basically it could be anything, from a straight threshold to a complicated expres-
sion. As (16.1) already indicates, it is even possible to wrap an entire neural
network to be used in another network as a single neuron with a complex transfer
[8]. We will refrain from that and apply only an S-shaped type of transfer,
especially a sigmoid function. More specifically, we will use the logistic and the
zero-centered variety: two sigmoid transfer functions with an exemplary symmetry,
the effects of which on the overall learning performance are nicely suited to display
the behavior as studied here. Their practical importance relies on the fact that the
logistic sigmoid has found major application in digital circuitry, while the zero-
centered sigmoid appears foremost in analog circuitry. The specification of these
functions is shown in Table 16.1.
x1
x2
y = f(x,w)
xp
Fig. 16.2 A multi-layer feed-forward network

Table 16.1 Specification of two sigmoid transfer functions

Logistic Zero-centered
Function ’ 1/(1 + ex) (1ex)/(1 + ex)
Output range (0, +1) (1, +1)
First-order derivative ex/(1 + ex)2 2e2x/(1 + ex)2
0
’ ¼ f(’) ’0 ¼ ’ (1’) ’0 ¼ 0.5(1’)2
The artificial neural networks (ANNs) adapt themselves to map the event–
consequence, evidence of behavior of the problem to solve, on the known problem
areas. In human terms, it can be said that the network learns the problem. This
learning is an optimization process, accomplished in many of the networks by
a gradient optimization method. The gradient methods can be described as a hill-
climbing (descending) procedure, which ends by finding an optimum (extreme)
state in which the network models the problem best. At this state it is expected that
every question from the problem area that is presented to the network will bring the
right reaction (answer) at the output of the network.
Any optimization problem can be defined by a set of inequalities and a single
cost function. In neural networks, this cost function is usually based on a measure of
the difference between the actual network response and the desired one. Its values
define the landscape in which an extremum must be found. By deliberately adapting
the weights in the network, both responses are brought together, which defines the
desired extremum. In other words, a measure of the difference in responses for one
or more inputs is used in an algorithm to let the network learn the desired weight
settings: the learning algorithm. It is a fundamental network design issue to
determine just the right value [9].
16.2.1 The Cellular Neural Network
Cellular automata have always created a scientific attraction by displaying complex

behavior on the basis of a regular computing structure with simple rules [40]. From
the observation that a cellular automaton can both imitate biology and work as
a Turing machine, the understanding has grown that they can also create auto-
nomous behavior. Many cellular automata implementations have been made in
software, but this approach will focus on a specific variation, the cellular neural
network, as it has a number of promising hardware realizations [10].
After the introduction of the Chua and Yang network [11], a large number of
CNN models appeared in literature. Like a cellular automaton, a CNN is made of
a regularly spaced grid of processing units (cells) that only communicate directly
with cells in the immediate neighborhood. Cellular neural networks [12] are widely
used with real-time image processing applications. Such systems can be efficiently
realized in analog techniques. Early attempts to realize the CNN functionality in
macro-enriched field programmable gate-arrays (FPGAs) have shown impressive
potential [13].
Fig. 16.3 Block diagram of a DT-CNN cell
A DT-CNN (discrete-time CNN), introduced by Harrer and Nossek [14], is

a regular multi-dimensional grid of locally connected cells. Each cell c communi-
cates directly with its r neighbors, i.e., a set of cells within a certain distance r to c,
where r 0. For example, if r ¼ 1 we have a 3 3 neighborhood and if r ¼ 2 we
have a 5 5 neighborhood. A cell can still communicate with other cells outside its
neighborhood owing to the network propagation effect.
The state of a cell c, denoted xc, depends mainly on two factors: the time-
independent input ud to its neighbors d and the time-variant output yd(k) of its
neighborhood. The neighborhood always includes c itself. This dependency in
a discrete time k is described by
X X
xc ðkÞ ¼ acd yd ðkÞþ bcd ud þ ic ; (16.2)
d2Nr ðcÞ d2Nr ðcÞ
while Fig. 16.3 gives an illustration.

Spatially invariant CNNs are specified by a control template A containing acd ,
a feedback template B containing bcd , and the cell bias i ¼ ic . Node template,
T ¼ <A, B, i>, determines together with input image u and an initial output y(0)
completely the dynamic behavior of the DT-CNN. Convergence to a stable state
will be achieved after a number of iterations n, i.e., at time k ¼ n.
Although neighborhoods of any size are allowed in DT-CNNs, it is almost
impossible to realize large templates. Limited interconnectivity imposed by current
VLSI technology demands that communication between cells is only local. Let
us restrict ourselves to the use of 3 3 neighborhoods, where templates A and B
are 3 3 matrices of real-valued coefficients. Additionally, the input range of
a DT-CNN cell is restricted to [1, þ1].
16.2.2 The Analog-Mostly Network ACE
Since the introduction in 1992, the concept of CNN-UM [41] (universal machine)
has been considered to realize the most complex CNN applications electronically,
owing to its universality and implementation capability. Several realizations

have seen the light of day, some focusing on analog-only or mixed-signal imple-
mentation in CMOS while others follow the footprints of the predecessor emulators
CNN-HAC [42] (hardware accelerator) and ACE. In this section, the mixed-signal
type is considered with focus on a specific chip series mainly developed by
a group at Centro Nacional de Microelectrónica at the University of Seville in
Spain [6, 15–20].
These CNN implementations have a number of drawbacks (Table 16.2). One
has to do with the difficulty of electrical cell design due to the various ranges
for the internal voltages and currents. These ranges have to be considered to
reduce the MOS transistor nonlinearities. Another issue is that input signals are
Table 16.2 Comparison of mixed-signal full-custom CNN universal chips. All chips use
a modified CNN model, i.e., the FSR model
CNNUC1a ACE400 [15] ACE4k [17, 18] ACE16k [19]
[6]
CMOS technology 1 mm 0.8 mm 0.5 mm 0.35 mm
Density [cells/mm2] 33 27.5 82 180
Array size 32 32 20 22 64 64 128 128
Input Type Binary Binary Binary and Gray Binary and Gray
scale scale
Optical ✓ ✓ ✓ ✓
Electrical ✓ ✓ ✓
Output Type Binary Binary
Electrical ✓ ✓
Dig. external control ✓
Global instr. memory Static Dynamic Static Dynamic
# Ana. instructions 8 8 32 32
# Dig. instructions 0 64 64 64
Local Type Digital Digital Dig.&Ana.
memory Dynamic ✓ ✓
Amount 4 Binary 4 Binary 4 Binary 2 Binary
(1-bit)
4 Gray 8 Gray
Ana. Acc. Input 8 bits 8 bits
7 bits + sign 7 bits + sign 7 bits 7 bits + sign
Bias 7 bits + sign 8 bits + sign N/A 7 bits + sign

Ana. circuit area/cell N/A 70% N/A
Cell array area/chip N/A 53% 58%
Cell area 180 170 190 190 120 102.2 73.3 75.7
mm2 mm2 mm2 mm2
Power Entire N/A 1.1W @ 5V 1.2W @ 3.3V < 4W@ 3.3V [20]
chip
Per cell N/A N/A 370 mW 180 mW
a
This architecture was not given a name in [6], but is called CNNUC1 here to emphasize that it was
the first universal chip of the series
always voltages while internal signals may be voltages or currents. This is crucial
in focal plane architectures where sensors provide the signals in the form of
currents. Incorporation of the sensory and processing circuitry on the same semi-
conductor substrate is pretty common [21] as CMOS technologies offer good
photo-transduction devices [22]. A conversion into voltages is then needed,
which complicates CNN interface design. Finally, the combination of internal
voltage and current signals leads to internal high-impedance nodes and, hence,
large time constants. This results in a lower operation speed than desired.
The full signal range (FSR) model has been introduced to overcome these
limitations [23, 24]. Here, all variables are in the form of currents, thus eliminating
the need of current-to-voltage conversion. The main difference compared to con-
tinuous-time (CT) and DT-CNNs is found in the way state variables evolve. State
variables have the same variation range as input and output variables, i.e., indepen-
dently of the application (16.3).
dxc X
t¼ gðxc ðtÞÞ þ dc þ facd yd þ bcd yd g;
dt d2Nr ðcÞ
8 (16.3)
< mðx þ 1Þ 1 x< 1
where gðxÞ ¼ lim x j xj 1
m!1 :
mðx 1Þ þ 1 x>1
This results in a reduced cell complexity for both CT and DT cases and thus
reduces area and power consumption in VLSI implementations. Stability and
convergence properties are guaranteed and proven to be similar to the original
models. It is further shown that uniform variations of the coefficients of the cloning
template affect only the time constant of the network.
The flexibility and generality of the CNN-UM lies in the ability to freely
reprogram the system using distinct analogic parameters, i.e., different cloning
templates and logic functions. This is guaranteed in [6] through a synergy of analog
and digital programmability. Internally, all cells are equipped with an analog-
programmable multiplier, while digital control signals are provided externally,
i.e., outside of the cell array. A specific interface circuitry is required to generate
the internal weights from the corresponding external digital signals. The interface is
located at the periphery of the cell array and behaves as a nonlinear D/A converter.
The analog weights are gradually adapted to the desired level and then used to
control the analog multiplier within the cells in the array. Each peripheral weight
tuning stage consists of an analog-controlled multiplier and a digital-controlled
multiplier connected in a feedback loop through an integrator. The conceptual archi-
tecture of the ACE16k, also called VSoC as being a clear advance in a roadmap
towards flexible vision systems on chips, is shown in Fig. 16.4.
Notable improvements in performance are further reached by pre-processing the
pixels locally to decrease the communication load [25]. By integrating more than
only low-level pixel processing, it aims to reduce the communication load with the
other system parts. This is clearly a next step in the development of smart sensors.
Load/upload
128 A /D & 128 D/A I/O image
control
Done
InstBus [53:0]
DDBUS [31:0]
24 D/A
Convolution Masks &
References Memory
GetPrg
Memory
Control
Done
Instruction
Memory
Fig. 16.4 Conceptual architecture of ACE16k [19]
The digital feature bus brings the advantage of intelligent networking of parts that
mask their analog nature but retain the low power and small area.
16.2.3 Pipelined CNN Architecture
In contrast to the analog focal-plane processor, the digital version will assume
a separate device to extract the image and to make the pixels available in a distinct
digital memory. This allows any image size, but cannot handle all pixels simulta-
neously. Large images are divided into sub-frames that are handled in sequence.
Consequently the operation of a DT-CNN on images covers many dimensions. The
local operation is performed in a two-dimensional plane (width and length) and
iterates in time (Fig. 16.5). Owing to the limited capacity of the CNN implementa-
tion, this has to be repeated over image slices and iterates over the surface to handle
potential wave propagation. Finally, the state-flow architecture is performed on
F low
a ge
Im
y
ra
e Ar
lin
pe
Pi
Fig. 16.5 The ILVA pipeline
sequences of images. This can be implemented in many ways, compromising

between speed and resource usage.
The CASTLE architecture [26] is a representative of the class of fully digital
emulators. The architecture is capable of performing 500 CNN iterations using
3 3 templates on a video stream with frequency of 25 fps taking 240 320
pixels each. This is valid for a system with 24 processing units (PEs) with
a precision of 12 bits. CASTLE makes use of the FSR model where the absolute
value of the state variable is never allowed to exceed the value of þ1.
Loading input pixels on-the-fly from an external memory into the processing
array constitutes a performance bottleneck. On the other hand, storing the entire
image on chip is impossible due to the limited resources. Instead, the image is
divided into a number of belts with a height of 2r pixels, where r represents the
neighborhood. Each belt is then fed to a PE (Fig. 16.6, right). In this case, the I/O
requirements of the PE, i.e., the cell, are reduced to two inputs and two outputs
per cell update. Each pair consists of one state value and one constant value,
corresponding to the combined contribution of control template together with the
bias (Fig. 16.6, left). The main memory unit in the PE consists of three layers of
equally sized circular shift-register arrays for the state input and two layers for each
of the constant and template selection inputs. Inputs from left and right neighboring
PEs are directly connected to the corresponding ends of the shift-register arrays.
In line with the proposed approach in the CNN-UM, the functionality of the
CASTLE architecture is ruled by means of a global control unit. One of the most
important features of this unit is the selection of the employed precision. Data
precision is variable and can be set to 1, 6, or 12 bits. The lower the accuracy, the
faster the system.
An important issue is the amount of logic occupied by the register arrays
constituting the internal memory units (IMUs, Fig. 16.6, left). In the first
StateIn ConstantIn TemplSelIn
LeftOut RightOut
Memory Unit
LeftIn RightIn
Timing Template
& Ctrl Memory
Arithmetic Unit
StateOut ConstantOut TemplSelOut
Fig. 16.6 Left: A schematic view of the processing unit in CASTLE, where dashed lines represent
control signals and continuous lines show the data path. Right: The belt of pixels stored on-chip for
1-neighborhood, where the black-outlined square indicates the current position of the convolution
operation [27]
experimental chip that has been fabricated in 0.35 mm CMOS [27] technology,
about 50% of the total area of a single PE is allocated to register arrays, while the
arithmetic block occupies not more than 21% of the area. Furthermore, experiments
show that a CASTLE emulator with 24 processors outperforms the digital signal
processor (DSP)-based ACE engine (Sect. 16.2.2) only when the rate of logic
operations is high enough [27].
The CASTLE architecture suffers from a number of drawbacks. One has to do
with the inability of emulating complex dynamic systems where operative parallel-
ism is a key feature. The single-layered architecture handles only one operation
at time. Other drawbacks include the limited template size, cell array size, and
accuracy. Hence, a new architecture called FALCON has been developed to
provide higher flexibility and to allow for multi-layer accommodation [28]. The
implementation is based on the FSR model with discretized state equations. In
contrast to all CNN-UM-inspired implementation discussed so far, the design is
hosted on a Xilinx Virtex series FPGA. This increases the ability of reconfiguration,
brings down developing time cycle, and decreases the overall cost.
The arithmetic uses a fixed-point representation, where possible value widths are
2–64 bits. Configurability is essential to allow accommodation of flexible precision
when needed. But for the highest possible precision the cell array will consist of not
more than four processing cells! The configuration is unfortunately not dynamic but
the entire design has to be re-synthesized and loaded on the FPGA every time a new
configuration is required! Apparently, for algorithms with alternating operations of
low and high precision the FPGA has to be reconfigured several times in order to
provide accurate results.
To overcome this drawback, spatial and temporal elements are mixed by
interweaving three pipelines corresponding to a row of three pixels (Fig. 16.7).
This reduces the latency and makes better utilization of the available resources. The
nodes are grouped in columns where each column holds a scan-line in the image
Scan-lines CNN topology Timing & Control
Iteration 1
Iteration 2
Fig. 16.7 Mixed spatial-temporal state-flow architecture operating directly on the pixel pipeline
stripe. The columns will then form iterations performed on the image. In this way,
one dimension (width or length) of the image frame together with the number of
iterations are implemented as columns of nodes while the other dimension of the
frame is handled by slicing as illustrated in Fig. 16.7. One of the resulting
realizations is a design called ILVA [4].
The underlying idea is that a two-dimensional computation of the local cell is
flattened into a series of one-dimensional computations by dropping intermediate
results on the computational path. In this way, the requirement of each node to have
data from eight neighbors for finding the output is met. In other words, we let every
node in the network contain image data from three pixels, i.e., pixel values for the
cell itself and for its left and right neighbors are stored in each node. A direct
connection with the two nodes above and below completes the communication
between a node and its neighborhood. In short, one node contains three pixels and
calculates the new value for one pixel and one iteration.
The prescheduled broadcasting in ILVA keeps the communication interface
at minimum, which allows for a large number of nodes on chip. The performance
is high as the system directly follows the line accessing speed, but the design suffers
from a number of weaknesses. The iterations are flattened on the pipeline, one
iteration per pipeline stage, making the number of possible iterations not only
restricted due to the availability of logic, but also fixed. Operations that require
a single iteration only still have to go through all pipeline stages. Output data has
to be fed back to the pipelined system in order to perform additional iterations,
making it far from trivial to handle larger iterations without accessing the external
image memory. This requires additional logic for loading and uploading pixel data
and therefore adds overhead for timing control and thereby severely slows down
the system.
16.2.4 In Short
Neural networks are programmed by templates that modify the connectivity of

many small but identical processing elements. A suitable technology realizes such
a PE by keeping it small. Analog design does this by suitable circuit design to
answer some reasonable precision requirements. Digital design has struggled with
the size of the multiplication operator. Therefore a number of specialized algorithm
implementations have been developed to limit the dependency on multiplier-size by
going for bit-level implementations [29].
Where size does not matter anymore, speed becomes of interest. But a faster
element does not make a fast system. Getting the data in and out of the system is
a major concern. This already plays a role in the matrix interconnect of the focal-
plane architecture, where early-on the light sensors are buffered to separate
the image extraction from the pixel reading. The concern about system-level
communication dominates the further developments towards the current Eye-RIS
vision system [30].
In the digital domain, the state-flow architecture, in its desire to exploit
pipelining for compromising speed and resource usage, gets a streaming character-
istic that loses simultaneously the efficiency of two-dimensional diffusion-reaction
behavior. This limits the attraction of CNN algorithms to provide more than just
low-level pixel processing.
Moreover, the FALCON architecture comes with no possibility of algorithmic
control on chip. All algorithmic steps, as well as local logical operations and
programs, are executed on a host PC. This reveals that the system cannot stand
alone but is always dependent on the host PC! Obviously, all the benefits of
performing complex tasks on the CNNs are lost.
To remedy these problems, the architecture is extended with a global control unit
GAPU (global analogic programming unit) [31] in line with the conceptual CNN-UM.
In addition to on-chip memories and some peripheral blocks, the GAPU is built up
using an embedded MicroBlaze processor core with 32-bit RISC architecture. Most
modern FPGAs provide at least one of these processor cores on chip. The extended
FALCON architecture is implemented on a Xilinx Virtex-II 3000 FPGA. Apart from
the embedded processor core, the GAPU occupies about 10% of the available logic,
which can be compared to the area of a single CNN processor that requires about 2.8%
of the logic. It is worth mentioning that the GAPU runs on lower clock frequency than
the processing units (PEs), thus setting a higher limit of the overall speed.
16.3 Many-Core CNN Architecture
The sequential evaluation order of CNN equations is closely related to the software
evaluation. Pixel lines are consumed immediately and subsequently the iterations
are pipelined, the time-ordered style. The alternative is the spatial approach, based
w
e Flo
ag
Im
1
g ine
En
e2
gin
En
Fig. 16.8 The n-dimensional Caballero architecture
on nodes which each relate to a CNN cell. In this situation, the pixel lines come into
the FIFO (first-in-first-out) until it is fully filled. Then these values are copied into
the CNN nodes, which subsequently start computing and communicating. Mean-
while new pixel lines come in over the FIFO. When the FIFO is filled again and the
CNN nodes have completed all local iterations, the results are exchanged with the
new inputs (Fig. 16.8). This leaves the CNN nodes with fresh information to work
on and the FIFO can take new pixel lines while moving the results out.
16.3.1 Architectural Variety
Another communication technique is involved with the network itself. The CNN
equation is unrolled not in time but in space, and the nodes retain the result of the
equation evaluation so that next iterations do not involve access to the external data
memory. In this state-scan architecture, the neighborhood is actively scanned for
the input values, known as Caballero [4]. The schedule is predetermined, but
computation and communication needs are decoupled by splitting the simple
node into a processor and a router. The nodes can theoretically transfer their values
within the neighborhood in parallel.
The need for explicit scheduling on nodal activities works out differently for
different CNN-to-network mappings. Two main categories can be distinguished:
– The consumer node is fully in accordance with (16.1). The discriminated output
of a cell is also the cell output and broadcasted to all connected nodes, where it is
a b
Legend: Circuit Node Multiplier Summing + Discrimination
Fig. 16.9 Consumer (a) and producer (b) cell-to-node mapping
weighted with the coefficients of the applied template before the combined
effect is determined through summation (Fig. 16.9a).
– The producer node discriminates the already weighted inputs and passes to each
connected node a separate value that corresponds to the cell output but weighted
according to the applied template (Fig. 16.9b).
Ideally all nodes are directly coupled and therefore bandwidth is maximal. In
practice, the space is limited and the value transfer has to be sequenced over a more
limited bandwidth. This problem kicks in first with the producer-type of network,
where we have 2n connections for n neighbors. The network-on-chip (NoC)
approach is meant to solve such problems. However, as the CNN is a special case
for such networks, being fully symmetric in the structure and identical in the nodal
function, such a NoC comes in various disguises.
In the consumer architecture, scheduling is needed to more optimally use the
limited communication bandwidth. Switches are inserted to handle the incoming
values one-by-one. To identify the origin of each value, one can either schedule this
hard to local controllers that simply assume the origins from the local state of the
scheduler (circuit switching, Fig. 16.10a), or provide the source address as part
of the message (packet switching, Fig. 16.10b). The former technique is simple.
It gives a guaranteed performance as the symmetry of the system allows for
an analytical solution of the scheduling mechanism. The latter is more complicated,
but allows also for best effort.
The counterpart of consumption is production. Every node produces values that
have to broadcast to all the neighbors. Again, where the communication has
a limited bandwidth, we need to sequence the broadcast and this can be done in
the same way as for the value consumption (Fig. 16.11). A typical example, where
each node can only handle one value at a time, is discussed in [4]. Such a scheme
suffers from a large latency overhead.
a b
Legend: Circuit Node Multiplier Summing + Discrimination Switch
Fig. 16.10 Value routing by multiplexing in space (a) and in time (b)
a b
Legend: Circuit Node Multiplier Summing + Discrimination Switch
Fig. 16.11 Value routing by multiplexing in space (a) and in time (b)
In the case of producer architectures, the nodal output is already differentiated

for the different target nodes. Each target node will combine such signals to a single
contribution. This combining network is an adder tree that will reduce the n values
to 1 in a pipeline fashion. Consequently, this tree can also be distributed, allowing
for a spatial reduction in bandwidth.
The overall processing scheme as shown in Fig. 16.12 is then similar to what has
been discussed for the consumer architecture. The main difference is that the
communicated values will be larger than the represented products and are therefore
of double length. Whereas the consumer architecture is characterized by “transfer and
calculate,” the producer architecture is more “calculate and transfer”. Furthermore,
a b
Legend: Circuit Node Multiplier Summing + Discrimination Adder
Fig. 16.12 Adder trees combine the network in the production architecture
they both rely on a strict sequencing of the communication, simultaneously losing

much of the principle advantage of having a cellular structure.
16.3.2 Numbers and Operators
Digital neural networks are like digital filters in the sense that they take a lot of
additions and multiplications. Actually they take mostly only multiplications.
As digital multiplications take a lot of space, one has always had an open eye for
serial multiplication or even “multiplier-less” algorithms.
16.3.2.1 Precision Needs
Moving from floating-point to fixed-point numbers is not merely a question of

reducing the value scope, but a careful arbitration between precision and accuracy.
Precision is addressed by the smallest step in the value space, such that small
variations in the value space have little significance in the problem space. Accuracy
on the other hand has to do with the degree by which the computation can achieve
the desired result. Figure 16.13 visualizes the difference between these two notions.
We see the difference between precision and accuracy back at the moment when
we move from floating- to fixed-point notation. The fixed-point notation requires an
explicit scaling, which is implicit in the floating-point counterpart. Let us assume
the decimal point in the right-hand location of the number. When we move the
decimal point to the left, the absolute size of the number decreases while we add
a fractional part. What happens when we move the decimal point to the left is that
we make the smallest step in the value space smaller, such that variations in the
Fig. 16.13 Precision versus Crisping to physical units

accuracy
Not accurate Accurate
Number Space
Not precise
Precise
Fig. 16.14 Precision needs in number representation
lower-order bits will have a reduced meaning. In other words, we gain precision,
but the absolute meaning of the value has decreased, giving a smaller range for the
number, and therefore chances are that we lose accuracy at the same time.
In the case of a DT-CNN, input and output values of every node are scaled
between 1 and +1 by definition. However, this is not true for the inputs to the
system. In the case of template values for image processing, such inputs have
a dynamic range that is close to that of the internal values. For other nonlinear
systems, this is not the case: they have a larger dynamic value range and/or that
range lies far away from the internal values (Fig. 16.14).
When scaling is properly performed, pruning can be used for further optimiza-
tion. In earlier work [32] it has already been shown that for 8-bit inputs the internal
representation does not have to allow for the full arithmetic precision of 21 bits
because the CNN structure is largely permissive for statistically random rounding
errors. The compensatory behavior is principally due to the feedback structure. The
approach performs gradual adjustment of template values, which allows pruning of
internal numbers down to the minimum with retained accuracy.
Throughout the experiment, the same fixed boundary condition has been used.
However, the choice of boundary conditions has a direct impact on the complete
stability of a certain category of CNNs, i.e., those with opposite-sign templates [33].
The instability of these CNNs depends far more on the boundary conditions than
on the template. Simple tests of, e.g., hole filling, have shown that gradual refine-
ment of the boundary conditions leads to noticeable deviations in convergence in
terms of functionality and speed [4]. Furthermore, robustness is also sensitive
to input data [34]. Hence, derived templates must be checked for robustness for
the “most difficult input image,” as this guarantees correct functionality. Further
modification in the pruning approach shows that the number of robust templates is
about 15% less for an image containing two holes, a complete hole positioned in the
center of the frame, while an incomplete hole is placed on one of the edges [4]. It is
therefore mandatory to have as many nodes as possible on a single chip.
16.3.2.2 The Case of the Multiplier
The high throughput of the system is due to the accommodation of three parallel
multipliers performing, in parallel, three multiplications that use pixels and
corresponding template coefficients as operands. Multiplication results are shifted
1-bit in the least significant bit (LSB) direction before they are forwarded to a tree
of adders to accumulate the results with the previous intermediate result. In order to
improve the accuracy, rounding units are introduced between the shifters and the
following adders. A limiter unit brings the final sum into the operational region.
Figure 16.15 depicts the structural architecture of the arithmetic unit. It is obvious
that the reduction of communication demands comes at the cost of larger arithmetic
units with more functional blocks.
In the serial approach, we pass all values simultaneously but in bit sequence.
For every step, the coefficient bits multiply the single-bit input from each neighbor;
the results are added and accumulated. As usual in bit-serial logic, the data-path
Pixel 1 Templ 1 Pixel 2 Templ 2 Pixel 3 Templ 3 Constant
Multiplier Multiplier Multiplier
Shift & Shift & Shift &

round round round
Adder Adder
Adder
Register Accumulator
Limiter
Fig. 16.15 The arithmetic unit in CASTLE [27]

becomes small but at the expense of a more complicated control. Furthermore we

can expect a higher latency, as more clock ticks are needed to get to the result. But
that is only true for the single node. The basic ten clock cycles for a single node
have to be repeated for five neighboring nodes for reasons of bus contention. It does
not seem likely that a serial solution that eliminates such bus contention problems
will need more. As the small serial node allows for a larger network to be imple-
mented on a single chip, it is worthwhile evaluating its potential.
A fully serial node computation is shown in Fig. 16.16. It requires the
coefficients to be locally available in a ring-buffer. This is not as bad as it seems,
because a serial shifter can be implemented in a single look-up table (LUT) per
4 bits. For longer coefficients one may consider building the ring-buffer in the
BlockRAM, but usually coefficients are not long. Together with the bit-register for
the input and the bit multiplier, a 4-bit base unit takes just a slice. The outputs are
tree-wise added and give a 4-bit result to be added to the shifting accumulator.
This final addition has to be in parallel because long carry propagation may occur.
Also the result has to be available in parallel, because a final table lookup is needed
for the output discrimination.
All values are notated in 2-complement numbers and range over [1. . .0]. For
such 2:x numbers we take x as 7 for the inputs and for the template coefficients. This
brings an internal representation of 2:21. Because all the values are identically
1
x 1
1
+
2
1
+
3
+
5 v.w+3
+ +
+
3
+
Control scheme:
0 w
2
+ 0
1 v-1
1
x
1 1 cycle
Fig. 16.16 Serial architecture for bit-serial communication

aligned, there is no real problem in finding the binary point in the final results.
Addressing the discrimination function by looking up the table implies trimming
the internal representation to the size of the external one.
16.3.2.3 The Series/Parallel Approach
This overview of alternative designs shows a rich variety of compromises between

speed and area. Starting from the bit-serial structure, even more alternatives
can be created by logic transformation. A typical example of such a derivative
implementation is in series/parallel computation (Fig. 16.17). Every single input
bit multiplies the entire coefficient. The outputs are tree-wise added and give
a w + 3-bit result to be added to the shifting accumulator. This reduces the latency
and the control significantly, but at the expense of wider adders. Connected to this
comes the implementation of buffering. Where in the pure bit-serial approach, the
buffers are directly created in hardware; in the derivatives it becomes worthwhile
implementing the buffers in the local memory.
We introduce here a very little known third alternative that has been introduced
in the past to achieve the best of both worlds: good addition and good multiplica-
tion. It is based on fixed-point arithmetic; hence addition and subtraction deserve no
further discussion. Multiplication and division are accomplished by an in-line
1
x w
w
+
w+1
1
+
w+2
+
w+3 v.w+3
+ +
+
+
Control scheme:
w+1 0 v-1
+
1 w 1 cycle
x
w
Fig. 16.17 Series/parallel architecture for bit-serial communication

interpretation (rather than off-line conversion) into the logarithmic domain in

a seemingly similar way as performed in FP.
The basic idea is that DIGILOG multiplication (Sect. 3.6 and [35–37]) can
be transformed to addition in the logarithmic domain. A binary logarithmic
number, A ¼ 2nan þ 2n1an1 þ . . . þ 21a1 þ 20a0 can be rewritten as log2A ¼
j þ log2(1 þ Ar/2j) in the binary logarithmic system; here j is the leading “1”s
position of A, and defined by j|aj ¼ 1 and ai ¼ 0 for all i > j. Ar is the remainder,
2j1aj1 þ . . . þ 21a1 þ 20a0. Therefore, in the DIGILOG system, two n-bit
binary numbers, A and B, can be written as
A ¼ 2n an þ 2n1 an1 þ þ 21 a1 þ 20 a0
and
B ¼ 2n bn þ 2n1 bn1 þ þ 21 b1 þ 20 b0 :
These two numbers have a leading “1” at the j and k positions, respectively. If all
the leading “0”s bits higher than the j in A or k in B are neglected, these two n-bit
binary numbers also can be written as a combination of the leading “1” and
a remainder:
1
A ¼ 2j aj þ 2j aj 1 þ þ 21 a1 þ 20 a0 ¼ 2j þ Ar
and
1
B ¼ 2k bk þ 2k bk1 þ þ 21 b1 þ 20 b0 ¼ 2k þ Br :
Then the multiplication of the two binary numbers A and B becomes
A B ¼ 2j 2k þ 2j Br þ 2k Ar þ Ar Br :
In this formula, the three terms 2j, 2k, 2jBr, and 2kAr can be easily implemented as
shift-and-add on the hardware level. In the DIGILOG multiplication, the rest term
Ar Br can be realized through iteration until Ar or Br becomes zero, rather
than that Br becomes zero in the conventional FP (floating point) multiplication.
Therefore the iteration time in DIGILOG only depends on the total number of “1”s
in A or B, whichever has the smaller number. In the FP, the iteration time of
multiplication depends on the total numbers of “1”s in the argument of choice.
16.3.3 Broadcast Mechanism
Also here, we have to look at the way values are broadcast. In contrast to the
consumer architecture, we have as many output values as there are neighbors. This
makes for an identical situation and no additional measures are needed, except for
the fact that we are not be able to generate all the different products at the same and
the sequencing issue pops up again.
The elimination of the global controller will speed up the operation of a CNN.
It has to synchronize all the local nodal controllers to ensure that no value transfer is
missed. The presence of such a lock step mechanism will undoubtedly decrease the
operational speed of the network. Therefore it has been investigated whether such
a lock step method is needed. It is proposed here to replace the local controller
that actively scans the inputs by an administration that is not dependent on the order
of arrival. This allows the operation of individual nodes to be speeded up without
harming the overall behavior.
Because of the regularity of the CNN network model we can combine the switch
and the network interface into a single unit that administrates per node the status
of all local communication per nodal computation. For each virtual connection,
we have (Fig. 16.18)
– A status field, showing whether the value has been rewritten;
– A redirection field showing whether the value has reached the final destination or
should still be forwarded;
– A value field, where the content of the last transmission is stored.
If a node receives a new value, this will be passed to the virtual switch and this
unit will in turn try to transmit this value to its neighbors. As all nodes may be thus
activated at the same time, infinite bandwidth will be needed. As mentioned earlier,
the bandwidth will be limited to handle one or two values at a time for most
practical applications. An arbitration mechanism is required to handle multiple
requests that can potentially create deadlocks. In the existing implementations, this
is done by enforcing a strict sequencing of all nodes, which can be easily shown to
be correct.
Local control is supported by the use of the forward field, which is meant
to pass values between the switches until eventually all final destinations are
reached. We have limited the forwarding to such switches that are also final
destinations. By stretching the communication cycles of a 1-neighborhood so
that it overlaps with the sequence of operations, routing demands are reduced and
controller is simplified.
overwrite forward value

LOCAL NODE
Fig. 16.18 The virtual

≠
switch model
From this basic concept, we can see an easy simplification. Where the values
come in, they can be immediately absorbed by the node. Performing the CNN
equation evaluation on-the-fly only the most current input value has to be saved if
– All values arrive in some order and
– Within this period no duplication occurs.
What remains is therefore an input register, an output register, and flags for all
virtual connections.
The moral of the serialized broadcasting approach is that the transfer needs to be
sequenced when the communication bandwidth is limited due to area restrictions.
We have also found that doing so by state machines leads not only to architectural
rigidity but also to degraded performance. For instance, 15% of the utilized area in
the serial broadcasting scheme is occupied by the state machine (Fig. 16.19). One
way to eliminate the need for the nodal controller, at least partially, is by transfer-
ring all values in a source-addressed packet. The original data-only packet used
previously is padded with a small header containing the position of the source node
in the grid. Hence, the packets carry their addressing information in the header,
which can be exploited in two different ways.
A better approach will make use of the intrinsic positioning information carried
in the header to address the local template memory of the current node. The nodal
equation is then performed in the manner the packets are received. The logic
required for the addressing of the value/coefficient pairs is greatly reduced through
the use of mirrored binary numbers of both the rows and the columns. In the case of
a 1-neighborhood only 2 bits for the row and 2 bits for the column address are
required. In general we need only 2(r + 1) bits, where r is the neighborhood.
A second remark is on the need for full buffering in the network interface. In
serial broadcasting, we can consume every incoming value immediately, as long
as we can guarantee monotonicity. For this, however, we need only the flags, not the
values themselves. Hence, the presence of the overwrite and forward flags allows us
to consume and pass on-the fly, reducing the need for buffering to only a read and
a write register.
Flip-flops LUTs Slices

200
150
100
50
0
Router Ctrl Proc. Discr. FIFO Node
Fig. 16.19 Area utilization of the different components with the serial broadcasting scheme
We are now left with a simple and scalable network interface, taking:
– Overwrite/forward flags for all cells in the neighborhood;
– A single node communication register for temporary storage of the incoming
values;
– At least one CNN state value register
– At least one CNN input register
As the status of each realized node takes only two registers, it can implement
a number of CNN cells. This raises the capacity for the multi-layer and multi-frame
situation. Add to this the elimination of the “control & forward bottleneck,” which
not only savesg clock cycles but also increases clock speed.
16.3.4 In Short
The many-core architecture provides a complete implementation including the

reaction/diffusion behavior. This enables the occurrence of nonlinear dynamics,
which uses more accurate template values. This seems to suggest the use of a
floating-point number representation. On a closer look, the precision has not
changed, mostly because of the precision of the nonlinear smashing function.
Consequently, a block floating-point representation is usually sufficient. This can
be accommodated simply on fixed-point numbers by the use of a scaling factor that
comes to each individual template.
A major architectural issue is the balance between the data-flow rate and the
computational complexity. The many-core architecture has special advantages
where each template requires much iteration, as is usually the case in nonlinear
dynamics. It has less of an advantage for low-level pixel operations for image
processing, where a pipelined architecture seems superior. In that light, the
seeming disadvantage of serial processing (on word-level or on bit-level) is not
severe, where otherwise the computation needs to be halted to wait for the data to
flow in.
16.4 The CNN Processor
The CNN Instruction-Set Architecture (ISA) defines the exterior of the CNN
Image Processor in terms of signals and visible memory locations. The overall
CNN ISA is depicted in Fig. 16.20. Overall, we find four modes of operation and
their respective instructions using two separate bus systems: the Image-Memory
Bus (IMB) and the Host-Interface Bus (HIB), both with a R/W signal and strobed
address and data bus.
Fig. 16.20 External view of IMB

the CNN architecture
Image Management Unit
(1)
Host Interface Unit

HIB
(3)
(2)
ISU Control Unit
16.4.1 Instruction Set
The Instruction Set covers four domains of activity: (a) the window on the image,
(b) the configuration of the network, (c) the running of the network, and (d) the
debug of the CNN software.
The window operations influence the image-management unit only. It converts
physical into virtual pixels and will autonomously fill the CNN with pixel informa-
tion with respect to the designated Region of Interest (RoI) for any frame format
using the IMB. Using the window settings it is possible to repeat the CNN program
on a steadily smaller part of the image while increasing the resolution.
– Frame Size: the width and height of a frame in pixels.
– Center coordinate: the non-sampled center of the first frame to be handled.
The internal operation is governed by a number of tables, downloaded over the
HIB. They all start with a preamble that gives the general table information and then
subsequently provides the table entries. The template and discrimination table will
be distributed to all nodes, while the program table is saved in the Instruction Store
Unit (ISU).
– Discrimination: table for discrimination function
– Program: Instruction Store (opt.)
– Template: label and content of a template
The discrimination function lists the transformation from internal node status to
external data result. The length of the table is therefore given by the table size
divided by the table step. The program tells the successive applications of pixel
operations that can be either templates or hard-coded linear instructions. It implic-
itly relates the use of various layers and how they are combined either in time or in
space. A template gives each CNN function. Templates can be downloaded and
stored in every CNN node for use later on. The pixel operations can be selected
from a number of linear (hardwired) and nonlinear (downloadable) options. The
instructions will be placed into a separate ISU.
– Logical: NOT, AND, OR, EXOR
– Arithmetic: Sum, Minus per pixel or horizontal or vertical
– CNN: refers to downloaded templates
The program execution is controlled by:
– Run: none, per clock, per iteration, per template till a specified breakpoint in the
program.
– Boundary: the boundary conditions as stated in the templates can be overwritten
for debug purposes.
– Sample Size: the amount of physical pixels represented by one virtual (CNN
internal) pixel as implied by the window can be overwritten for debug purposes.
– Mode: only this window, or a stripe of the entire image.
The ISA makes the CNN network architecture invisible to the host program and
therefore allows a late binding of the actual structure to an application at hand.
More often than not, the development network is different from the production
network. Starting from a MATLAB model with values represented in a double
floating-point format, a gradual conversion into fixed-point numbers is needed [6].
The length of the internal words is application dependent, though accuracy can be
easily guaranteed by block-based scaling with a factor derived by inspection of the
templates. In practice we have not seen the need for more than 8-bit precision, but
for simple templates a smaller length can be accepted.
In line with this, we have inserted a number of in-line debug facilities. The
system can be run in various time step sizes, inspected for network data, while
allowing the network status to be overwritten and to continue from the existing
status. In our reference system we assume that the network is configured separately
from the rest. Consequently we have to ensure that the system components can
handle appropriate network architectures.
16.4.2 Blocks
The processor largely comprises three blocks: (a) the Host Interface Unit, (b) the
Image Management Unit, and (c) the Control Unit.
A host must be able to control the overall functionality of the system by sending
instructions and cloning templates and by setting a number of configuration
parameters. The communication is handled by the HIU, which receives the requests
from the host over the HIB and forwards them to the system using a wishbone bus.
The HIU is also responsible for data delivery to the host. Two different FIFOs are
used, one for acquiring host requests and one for putting out data to the host.
R/W Virt. Addr. Data

bit 24 23 16 15 0
Fig. 16.21 A host request is subdivided into flag, address, and data fields
A host request is 25 bits long and is divided into three fields: a Read/Write flag
that determines the type of the request, a virtual address field, and a data field that is
of interest only in write-requests (Fig. 16.21). Once a request is captured by the
FIFO, the virtual address is translated into a system memory address by the I/O
memory mapping unit (IOMMU). This address will serve as a base address for all
incoming data as long as the virtual address field in the subsequently received
requests remains unchanged. The bus master acts partially as a direct memory
access (DMA); it generates the proper addresses from the base address and puts it
on the address port of the wishbone bus. In the case of a read request, once data are
available, the wishbone bus raises an ACK(nowledgment) signal notifying the bus
master that reads the data and puts it on the output FIFO. Write requests are handled
similarly. Here the ACK signal notifies the bus master that the writing of data is
accomplished so the next pair of address/data can be handled.
The camera captures images and stores them in an external memory. The 8-bit
gray-scale pixels are then retrieved and converted by the IMU to a signed fixed-
point notation with a precision of 7 bits for the fractional part. One of the main
operations of the IMU is the windowing operation. As the size of the network is far
smaller than the processed image frame, a gradual zooming toward the RoI is
required. At the beginning the RoI covers the entire frame, where each CNN
node on the chip is mapped onto a virtual pixel that corresponds to a group
of real pixels in the image. The virtual pixel is suitably obtained through
a conventional averaging of all pixels in the corresponding group. In the next
round the RoI covers a smaller part of the frame, depending on the output of the
previous round.
The unit has direct communication to the CNN core and the HIU through
wishbone buses. It is built with the concept of pipelining in mind and consists of
two main components: Instruction Fetch and a Controller (acts as instruction
decoder). The naming convention is somehow misleading as the former pipelining
stage generates two additional signals; control (used by the Controller pipeline
stage) and iteration; in addition to the instruction that is fetched from a dual-
port RAM. The controller consists of two major components. One is the actual
instruction decoder and provides the proper template, while the other generates
CNN-enable and instruction-done signals, depending on the number of iterations
and whether equilibrium is reached or not.
The instruction memory (ISU) is arranged as shown in Fig. 16.22 with a space for
64 instructions. Taking a Xilinx Virtex-II 6000 as reference, which accommodates
34000 slices, we find that the overhead incurred by turning a CNN network into a
system ranges from 1% for a limited edition to 5% for a complete one with large
buffers.
Fig. 16.22 Memory address 143 16 15 1 0 bytes

space as used by the
control unit
Reserved
Iter.Step
Control
Instructions Configuration
Figure
Reconstruction
Bifurcation Logical
Detection AND
False
Image Logical
Feature
Preprocessing OR
Elimination
Ending Logical
Detection And
Figure
Reconstruction
Fig. 16.23 Algorithmic structure of hand vein recognition
16.4.3 Multi-level Implementation
The CNN computations can be formally identified within morphological algebra.

The primitives of low-level pixel operations are the three set-operations “not”, “or”,
and “and”, supplemented with directions by iterative operations on “dilation” and
“erosion.” It has been demonstrated how any morphological expression can be
written out into a flat list using only primitives, such that this flat list can be
procedurally assembled into an optimal CNN structure [12].
The dataflow graph of the CNN computation displays a number of parallel paths.
Each parallelism runs on each instance of the basic CNN structure, which can be
either a separate level or a recursive structure over a separate memory domain.
A typical example is in hand vein recognition, where the same image is analyzed for
different features (bifurcation and ending) to achieve a collective annotation of the
reconstructed overall image (Fig. 16.23).
Fig. 16.24 Motion extraction
A similar situation is where the same CNN computation operates on subsequent

images. In Fig. 16.24, it is shown how motion can be extracted from subsequent
images. Where the expression is evaluated before the next frame is available, it is
feasible to use the memory domain without loss of system speed. Nevertheless,
the fundamental issue involves state explosion, that is strictly reduced by time-
multiplexing, where image extraction is less accelerated than by microelectronic speed.
The integration of multiple images helps not only to analyze in time, but also to
analyze in space. The objects on a single image, but at a different distance from the
sensor, are nonlinearly and therefore non-constantly scaled. Finding the scaling is
important for 3D reconstruction. A similar technique can be used for stitching
together images of a single object from sensors at different angles [38].
16.4.4 In Short
Branch handling is not only a concern for a conventional processor, but also for
a CNN. In the pipelined architecture, a branch condition will stop the operation until
the current cycles are performed while new input is halted, similar to flushing. For the
many-core architecture, this is not an issue, but the question arises how the instruction
fetch is realized. In our CNN processor, the CNN area has the CNN program
inclusive branch conditions, and it will pre-fetch. This shows the control is local,
while the attached processor is not involved with the details of the CNN operation
but simply performs the additional administration. In other words, the CNN part
autonomously handles a complete algorithm: no additional DSP is required.
The ingredients of the CNN architecture are close to current tiled architectures of
the many-core type. This similarity can be exploited by the evaluation of many-core
compilers. The (optimal) CNN many-core solution is at least understood, even if
not solved. Therefore, it provides a benchmark for compilers by evaluating how
well the compiled code exploits the symmetry issues in the CNN algorithm. The
similarity is not complete, as the CNN allows a cellular network without global
control, whereas the compiler technology will usually have problems taking the
Wave Computing effect of nonlinear dynamics into account.
Clock synchronization is not part of the mathematical concept of cellular
networks. On closer inspection, the always “natural” addition of clock synchroni-
zation makes it easier to understand but at the same time takes the performance
profit away. The analogic realization does not need a clock, because it reaches,
by nature, the more speed- and power-efficient solution. Boxing the CNN as a
programmable digital unit is a consequence of the need for a generally applicable
software development environment, but the obvious detrimental effects of achieving
generality in loosing specificity may be worth to spend new thoughts.
For the digital realization, the clock synchronization paradigm has always
been taken for granted. As a consequence, a number of architectural experiments
have broken down in scaling as the global control lines became dominant for the
system speed. Architectures that can run for many cycles without being controlled
have clearly provided a far better operation.
Apparently, the actual problem is not in the clock, but in the global control that is
usually been designed in a synchronized way. Having a small instruction stack in
each node will provide some remedy, but the effect of the small control spread will
still not become visible if the nodes are globally synchronized. A more attractive
solution comes from the area of asynchronous design. If a node can autonomously
determine that it has no activity, it can go into a “no operation” mode. If a node sees
that all surrounding nodes are in “no operation,” it can fetch the next nodes.
This simple solution is not sufficient, as it cannot be guaranteed that all nodes in the
network are inactive by simply looking at the neighbors. Where the CNN creates
system activity by doing only local interaction, a mirror mechanism is clearly needed.
16.5 New Media Applications
The naive image application of CNNs has largely been involved with low-level
pixel computations, such as edge detection. It takes the image from the camera and
makes it a better one. But that is drastically changed with the advent of smart
cameras. Adding high-level image computations, the camera starts to determine the
parts of the scene. Then it is just a small step towards an understanding of what
appears in the scene. This comes together with a further complication of the camera.
To handle the algorithmic complexity, the platform requires a processor with
a high processing power, typically a graphics processor with the potential of
a supercomputing node.
This is invoking a counter movement in an algorithmic swap. The camera is not
the encapsulation of the functionality, but the function will encapsulate vision. And
even more! The overall system will operate on information that is real-time
extracted from the embedding system. As the embedding system reacts to this, it
will enforce the sensor, and the time-sequenced interaction creates the desired
knowledge. This vision-in-the-loop is the area of New Media. In a similar fashion,
audio and video can be handled. The hardware platform discussed here is only
restricted to any topological map.
The cellular hardware is especially efficient in video extraction. It is known from
the continuous MPEG improvements that 90% of the resources are exploited by
movement extraction. There is an interesting alternative, when each object with
interesting movement has a small signal added to it outside the natural light
spectrum. Pulsation brings a further identification. Consequently, the signal samples
have to be stored.
It is only a small step from the pulsating signal to audio. And again, we have to
consider the increased order of the state explosion. The window on an audio signal
is small from a suitable selected pulse, but in general the window will be much
larger. Using modularity the window size is still limited, but it will be too large for
comfort [39].
The multi-level CNN is handled by time-multiplexing: the same cells are run but
by new parameters from the local memory. For the Fourier transform, we cannot
use the same cells, but we can time-multiplex new cells that reconfigure a number
of CNN cells. This is a straightforward extension of the previous concept,
as illustrated in Fig. 16.25.
So far we have discussed exploiting CNN operation to its theoretical potential.
The added problem is to raise that potential. For the simple one-level CNN, all the
computations will be local and therefore short and fast. For more complex systems,
and notably for new media applications, we need multi-level and multi-mode.
As the planar microelectronic technology is essentially two-dimensional, we have
used multi-levels through memory structures, but this technique is limited.
The use of 3D technology is an interesting concept from the 1980s, but has
recently become more viable. Especially by looking at the adder tree structures in
the nodes, it seems that stacking transistors can provide a clear decrease of size and
therefore a structured method to shorten the signal path in computational logic.
With new vigor in stacking CMOS technology, smaller cells could be realized
within the coming decade, but it does not solve all of the problems.
The other technology need is illustrated by the figures that raised the increasing
modes. Instead of time-multiplexing over the memory structures, it seems of
advantage to stack wafers. Again, the idea is not new. In the 1990s, research was
proposed to stack an analog, a mixed-signal, and a fully digital wafer for complex,
new media applications. Most of these projects come from a desire to imitate
low
g eF
Ima
e1
gin
En
e2
gin
En
Fig. 16.25 The heterogeneous CNN-based system
biology. Recently, the idea has been renewed, though initially as an accelerator for
bio-morphic systems.
Overall, we see that the CNN was inspired in the late 1980s, and it has gradually
matured in theory and application, with both analog and digital architectures. The
new interest in scaling this paradigm to larger systems, where few or no other
technologies can solve the complex problems, matches nicely with recent techno-
logical progress in different areas. This makes the basic ingredients of a disruptive,
promising development.
References
1. Hennessy, J.L., Patterson, D.A.: Computer Architecture: A Quantitative Approach, 4th edn.
Morgan Kaufmann, San Francisco (2007)
2. Tummala, R.R.: Moore’s law meets its match. IEEE Spectr. 43(6), 44 (2006)
3. von Neumann, J.: The general and logical theory of automata. In: Newman, J.R. (ed.) The
World of Mathematics, pp. 2070–2089. Simon and Schuster, New York (1954). Reprinted by
Dover Publications (2000)
4. Malki, S.: On hardware implementation of discrete-time cellular neural networks. Ph.D.
Thesis, Lund University, Sweden (2008)
5. Zhirnov, V., Cavin, R., Leeming, G., Galatsis, K.: An assessment of integrated digital cellular
automata architectures. IEEE Comput. 41(1), 38 (2008)
6. Domı́nguez-Castro, R., Espejo, S., Rodrı́guez-Vázquez, A., Carmona, R.: A CNN universal
chip in CMOS technology. In: Proceedings 3rd IEEE International Workshop on Cellular
Neural Networks and their Applications (CNNA-94), Rome, pp. 91–96, December 1994
7. Haykin, S.: Neural Networks: A Comprehensive Foundation. Macmillan, New York (1994)
8. Keegstra, H., Jansen, W.J., Nijhuis, J.A.G., Spaanenburg, L., Stevens, J.H., Udding, J.T.:
Exploiting network redundancy for lowest–cost neural network realizations. In: Proceedings
IEEE International Conference on Neural Networks, ICNN ’96, vol. 2, Washington, DC,
pp. 951–955 (1996)
9. Hornik, K., Stinchcombe, M., White, H.: Multilayer feedforward networks are universal
approximators. Neural Netw. 2, 359 (1989)
10. Spaanenburg, L., Malki, S.: Artificial life goes ‘in-silico’. In: Proceedings 2005 IEEE Interna-
tional Conference on Computational Intelligence for Measurement Systems and Applications,
CIMSA 2005, Giardini Naxos, Taormina, Sicily, Italy, pp. 267–272, 20–22 July 2005
11. Chua, L.O., Yang, L.: Cellular neural networks: theory. IEEE Trans. Circuits Syst. 35,
1257–1273 (1988)
12. ter Brugge, M.H.: Morphological design of discrete-time cellular neural networks. Ph.D.
Thesis, Rijksuniversiteit Groningen, Netherlands (2004)
13. Uchimoto, T., Hjime, H., Tanji, Y., Tanaka, M.: Design of DTCNN image processing
(Japanese). Transactions of the Institute of Electronics, Information and Communication
Engineers J84-D-2, 1464–1471 (2001)
14. Harrer, H., Nossek, J.A.: Discrete-time cellular neural networks. Int. J. Circuit Theory Appl.
20, 453 (1992)
15. Domı́nguez-Castro, R., et al.: A 0.8 mm CMOS two-dimensional programmable mixed-signal
focal-plane array processor with on-chip binary imaging and instructions storage. IEEE
J. Solid-St. Circ. 32, 1013 (1997)
16. Espejo, S., Domı́nguez-Castro, R., Liñán, G., Rodrı́guez-Vázquez, A.: A 64 64 CNN
universal chip with analog and digital I/O. In: Proceedings 1998 IEEE International Confer-
ence on Electronics, Circuits and Systems (ICECS98), Lisbon, pp. 203–206 (1998)
17. Liñán, G., Espejo, S., Domı́nguez-Castro, R., Roca, E., Rodrı́guez-Vázquez, A.: CNNUC3:
a mixed-signal 64 64 CNN universal chip. In: Proceedings 7th International Conference on
Microelectronics for Neural, Fuzzy and Bio-Inspired Systems (MicroNeuro-99), IEEE,
Piscataway, NJ, pp. 61–68 (1999)
18. Liñán, G., Espejo, S., Domı́nguez-Castro, R., Rodrı́guez-Vázquez, A.: ACE4k: an analog I/O
64 64 visual microprocessor chip with 7-bit analog accuracy. Int. J. Circuit Theory Appl.
30, 89 (2002)
19. Rodrı́guez-Vázquez, A., Liñán-Cembrano, G., Carranza, L., Roca-Moreno, E., Carmona-
Galan, R., Jimenez-Garrido, F., Domı́nguez-Castro, R., Espejo Meana, S.: ACE16k: the
third generation of mixed-signal SIMD-CNN ACE chips toward VSoCs. IEEE Trans. Circuits
Syst., I: Regul. Pap. 51, 851 (2004)
20. Liñán, G., Rodrı́guez-Vázquez, A., Espejo, S., Domı́nguez-Castro, R.: ACE16k: A 128 128
focal plane analog processor with digital I/O. In: Proceedings of the 7th IEEE International
Workshop on Cellular Neural Networks and their Applications (CNNA 2002), Frankfurt,
Germany, pp.132–139, July 2002
21. Liñán, G., Foldesy, P., Espejo, S., Domı́nguez-Castro, R., Rodrı́guez-Vázquez, A.: A 0.5 mm
CMOS 106 transistors analog programmable array processor for real-time image processing.
In: Proceedings 25th European Solid-State Circuits Conference ESSCIRC ’99, Stresa (Italy),
pp. 358–361
22. Delbr€uck, T., Mead, C.A.: Analog VLSI phototransduction by continuous-time, adaptive,
logarithmic photoreceptor circuits. California Institute Technology, Computation and Neural
Systems Program, Tech. Rep. CNS Memo no. 30, Pasadena, May (1994)
23. Rodrı́guez-Vázquez, A., Espejo, S., Domı́nguez-Castro, R., Huertas, J.L., Sanchez-Sinencio, E.:
Current-mode techniques for the implementation of continuous- and discrete-time cellular neural
networks. IEEE Trans. Circuits Syst. II: Analog Dig. Signal Process. 40, 132 (1993)
24. Espejo, S., Rodrı́guez-Vázquez, A., Domı́nguez-Castro, R., Carmona, R.: Convergence and
stability of the FSR CNN model. In: Proceedings 3rd IEEE International Workshop on Cellular
Neural Networks and their Applications (CNNA-94), Rome, pp. 411–416 (1994)
25. Jimenez-Marrufo, A., Mendizdbal, A., Morillas-Castillo, S., Dominguez-Castro, R., Espejo, S.,
Romay-Judrez, R., Rodriguez-Vazquez, A.: Data matrix code recognition using the Eye-RIS
vision system. In: Proceedings IEEE International Symposium on Circuits and Systems,
ISCAS 2007, New Orleans, p. 1214
26. Zarándy, Á., Keresztes, P., Roska, T., Szolgay, P.: An emulated digital architecture
implementing the CNN Universal Machine. In: Proceedings 5th IEEE International Workshop
on Cellular Neural Networks and their Applications (CNNA-98), London, pp. 249–252 (1998)
27. Keresztes, P., et al.: An emulated digital CNN implementation. J. VLSI Signal Process 23,
291–303 (1999)
28. Nagy, Z., Szolgay, P.: Configurable multilayer CNN-UM emulator on FPGA. IEEE Trans.
Circuits Syst. I: Fundam. Theory Appl. 50, 774 (2003)
29. Lopich, A., Dudek, P.: An 80 80 general-purpose digital vision chip in 0.18 mm CMOS
technology. In: Proceedings IEEE International Symposium on Circuits and Systems, ISCAS
2010, Paris, pp. 4257–4260, May 2010
30. Alba, L., et al.: New visual sensors and processors. In: Arena, P., Patanè, L. (eds.) Spatial
Temporal Patterns for Action-Oriented Perception in Roving Robots. Cognitive Systems
Monographs, pp. 351–370. Springer, Berlin/Heidelberg (2009)
31. V€or€osházi, Z., Nagy, Z., Kiss, A., Szolgay, P.: An embedded CNN-UM global analogic
programming unit implementation on FPGA. In: Proceedings 10th IEEE International
Workshop on Cellular Neural Networks and Their Applications, CNNA ’06, Istanbul, Turkey,
Aug 2006. doi: 10.1109/CNNA.2006.341652
32. Fang, W., Wang, C., Spaanenburg, L.: In search of a robust digital CNN system.
In: Proceedings 10th IEEE International Workshop on Cellular Neural Networks and their
Applications, CNNA ’06, Istanbul, Turkey, Aug 2006. doi: 10.1109/CNNA.2006.341654
33. Thiran, P.: Influence of boundary conditions on the behavior of cellular neural networks. IEEE
Trans. Circuits and Syst.– I: Fundam. Theory Appl. 40, 207 (1993)
34. Mirzai, B., Lı́m, D., Moschytz, G.S.: Robust CNN templates: theory and simulations. In: 4th
IEEE International Workshop on Cellular Neural Networks and their Applications, CNNA-96,
Seville, Spain, pp. 393–398 (1996)
35. Grube, R., Dudek, V., Hoefflinger, B., Schau, M.: 0.5 Volt CMOS Logic Delivering 25 Million
16 16 Multiplications/s at 400 fJ on a 100 nm T-Gate SOI Technology. Best-Paper Award.
IEEE Computer Elements Workshop, Mesa, (2000)
36. van Drunen, R., Spaanenburg, L., Lucassen, P.G., Nijuis, J.A.G., Udding, J.T.: Arithmetic for
relative accuracy. In: Proceedings IEEE Symposium on Computer Arithmetic, ARITH 1995,
Bath, England, pp. 239–250 (July 1995)
37. Spaanenburg, L., Hoefflinger, B, Neusser, S., Nijhuis, J.A.G., Siggelkow, A.: A multiplier-less
digital neural network. In: Digest 2nd Int. Conference on Microelectronics of Neural
Networks, MicroNeuro 1991, Muenchen, Germany, pp. 281–290 (October 1991)
38. Zhang, L., Malki, S., Spaanenburg, L.: Intelligent camera cloud computing. In: IEEE Inter-
national Symposium on Circuits and Systems, ISCAS 2009, Taipei, Taiwan, pp. 1209–1212,
May 2009
39. Chalermsuk, K., Spaanenburg, R.H., Spaanenburg, L., Seuter, M., Stoorvogel, H.: Flexible-
length fast Fourier transform for COFDM. In: Proceedings 15th IEEE International Confer-
ence on Electronics, Circuits and Systems, ICECS 2008, Malta, pp. 534–537, Aug 2008
40. Kari, J.: Theory of cellular automata: a survey. Elsevier Theoretical Computer Science 334,
3–33 (2005)
41. Chua, L.O., Roska T.: The CNN universal machine – Part I: The architecture. In: Proceedings
2nd IEEE Int. workshop on Cellular Neural Networks and their Applications, CNNA-92,
Muenich, Germany, pp. 110 (1992)
42. Roska, T et al.: A hardware accelerator board for Cellular Neural Networks: CNN-HAC.
In: Proceedings IEEE Int. workshop on Cellular Neural Networks and their Applications,
CNNA-90, Budapest, Hungary, pp. 160–168 (1990)
Chapter 17
Retinal Implants for Blind Patients
Albrecht Rothermel
Abstract Recently, very promising results have been obtained in clinical trials
with eye-prostheses for the blind. There is a chance that advances in surgical
techniques, microelectronics design, and material science may lead to the first
really useful applications of retinal implants in the near future. This chapter will
focus on the actual status of subretinal surgery and implant technologies.
Opportunities and limitations of the different technologies will be discussed in
terms of patients benefit and technological challenges.
Finally, a vision on how the devices may work and look like in the future will be
given.
17.1 Introduction: Causes of Blindness and Overview

of Projects
Blindness is one of the most severe handicaps we can imagine. However, here we
have first to distinguish between people who are blind from birth, on the one hand,
and those who became blind by disease or accident.
People who are blind from birth completely lack any training of their visual
system, and that means that probably there will never be any artificial vision for this
group. However, people who have never seen in their life typically do not miss this
ability too much.
Today’s approaches for prostheses for the blind work exclusively with patients
whose blindness is a result of diseases. Among them are age-related macular
degeneration (AMD) and retinitis pigmentosa (RP), which are both related to
photoreceptor degeneration [1, 2]. An important characteristic of these degenera-
tive diseases lies in the fact that the optical nerve and most cell layers in the retina
A. Rothermel (*)
University of Ulm, Institute of Microelectronics, Albert-Einstein-Allee 43, Ulm 89081, Germany
e-mail: info@albrecht-rothermel.de

368 A. Rothermel
Fig. 17.1 The human visual system [3]
are still alive and functional. Another motivation for the development of electronic
appliances to help RP patients derives from the complete lack of any medical
treatment to stop the photoreceptor degeneration. A restoration of even rudimentary
vision would be of significant benefit to these patients.
Figure 17.1 shows the main nerve layers in the retina of humans. Only the rods
and cones are missing when RP patients have gone totally blind; the neurons, which
include bipolar cells, horizontal cells, and ganglion cells, are still alive and are able
to generate spike train action potentials.
Experiments have been carried out at the Natural and Medical Sciences Institute
at the University of Tuebingen by Stett and Gerhardt. The retina was extracted from
a Royal College of Surgeons (RCS) rat with inherited retinal dystrophy, which is a
widely studied animal model of retinal degeneration. Samples of the extracted
retina were placed on a microelectrode array. Pulse trains could be measured on
ganglion cells as a response to stimulation pulses applied to the electrodes [4].
Figure 17.2 shows measured pulse trains for various stimulation pulse durations.
This kind of experiments verified that direct stimulation of retinal cells could
generate so called phosphenes (visual perceptions) in patients [5]. With qualitative
and quantitative results about the sensitivity of the retinal nerve cell layers, it was
possible to design experiments on blind humans.
Four different locations of the stimulation electrodes are being investigated by
different teams throughout the world.
17 Retinal Implants for Blind Patients 369
Fig. 17.2 Action potentials measured on a RP rat retina in vitro [4]
1. Epiretinal: Looking from the direction of the vitreous body in the eyeball, we
approach the retinal layer from the ganglion cell side. The blood vessels and
axons that form the optical nerve are located here. The epiretinal approach
places the stimulating electrodes into the space of the vitreous body close to
the ganglion cell layer. These devices typically separate the stimulation elec-
tronics from the electrodes themselves. The advantage is a high flexibility in
electrode design and placement, the drawback the wiring required for each and
every electrode. One very active team in this area is the company Second Sight
[6]. In Europe, the same approach is followed by the company IMI [7] and the
consortium Epiret. Figure 17.3 shows an example of such an approach. The
picture information to control the electrodes is captured by an external camera,
which allows for extensive image processing to optimize the stimulation signals.
If the eye movements of the patient are to influence the stimulated pattern in
a “natural” way, eye tracking is required.
2. Subretinal: Placing the electrodes at exactly the position where the photo-
receptors were leads to the so-called subretinal approach. Most active in this
area is a German group with the company Retina Implant AG driven by Zrenner
[8, 9]. Also a group at the MIT follows this principle [10]. The advantage is that
the stimulating electrodes are as close as possible to the location of the natural
photoreceptors (Fig. 17.4). This arrangement promises the best stimulation
sensitivity and best spatial resolution. The natural pressure of the retinal layer
on the choroid guarantees the best proximity of electrodes and nerve cells.
The MIT group [10] combines this approach with an independent design of
370 A. Rothermel
Fig. 17.3 Epiretinal stimulator design
Fig. 17.4 Subretinal stimulation device in the human eye [11]
electrodes and electronics, allowing forcomplex electronics and high flexibility

in the design, with the drawbacks of the wiring costs being proportional to
the number of electrodes and the requirement of an external camera.
The German group based in Tuebingen [8] is the only one to the best of our
knowledge that combines photosensors and electrodes in close proximity to
use the natural optical path of the eye. The subretinal chip is a camera, signal
processor, and stimulator all in one piece of silicon. Wiring this way is indepen-
dent of the number of electrodes. In fact, the electrode number could be very
high and is limited in practice by the limited spatial resolution and charge-
transfer characteristics of the electrode material. The major drawbacks are the
limited picture signal processing capabilities and limited flexibility of signal
processing. Also the surgical procedure might be the most ambitious of the ones
discussed here as the retinal layer, especially of RP patients, is a very delicate
structure. However, retina manipulation is a quite common eye treatment
surgery these days.
3. Subchoroidal: A little bit further towards the sclera is the position that an
Australian consortium [12] is planning to use. The bionic eye project with
a volume of Au$42 million uses an external camera again and places the
electrodes between the choroid and the sclera. As far as we understand, the
advantage lies in the fact that the eyeball does not have to be opened for this kind
of surgery.
4. Suprachoroidal transretinal (STS): Going one step further, a Japanese group
follows the STS approach, which means that the electrodes are attached to the
sclera and again the eyeball does not have to be opened [13].
17.2 Organization of the Subretina Chip
The following sections mainly deal with a German project of the Tuebingen group.
Driver of the project is Prof. Zrenner from the eye hospital in Tuebingen [14].
Commercial aspects are covered by the company Retina Implant AG. Research in
this area started more than 10 years ago [15] together with the Natural and Medical
Sciences Institute at the University of Tuebingen and the Institute for Microelec-
tronics in Stuttgart.
The subretinal implant device as of today is wirelessly supplied by inductive
energy and control signal transfer through the skin. A sealed ceramic box is used as
the receiving device, similar to today’s technology available for cochlea implants.
Highly flexible wiring connects the subretinal chip with the receiver box
(Fig. 17.5). This approach has been proved by the company Second Sight to
allow a very satisfactory lifetime of several years [16].
With the German design, just the transmitting coil is attached to the outside of
the patient’s head, typically behind his ear, and magnetically secured. No other
devices, no camera, and no glasses are essential for the operation of this type of
device, which opens up chances for a very natural appearance of patients with this
kind of prosthesis.
Two different chip designs are utilized for the subretinal stimulator. The earlier
generation chip [17] was arranged in an array of about 1,500 pixels, each equipped
with their own photosensor, amplifier, and stimulator. A more recent generation
372 A. Rothermel
Fig. 17.5 Subretinal implant device
Fig. 17.6 Block diagram of subretinal chip [18]
[18] has been designed together with the University of Ulm, Germany. Figure 17.6
shows the block diagram. To improve the lifetime of the delicate structure, all
terminals of the chip have been designed to work in a DC-free way. Control signals
are supplied by control currents with an input voltage level clamped to GND. The
supply voltages, however, are also DC-free. VH and VL are symmetrically oriented
piecewise linear square wave voltages with peak-to-peak values of just 2 V. For
internal operation, of course a rectifier is incorporated, however, to drive the
stimulating electrodes, the polarity change of VH is used directly at the output to
generate a biphasic output stimulating current and voltage.
Not all output electrodes have to be addressed at the same time; they can be fired
sequentially by a pattern selection signal to reduce peak currents in the supply and
in the tissue.
Fig. 17.7 Chip micrograph
Figure 17.7 shows a micrograph of the subretinal chip with 40 40 pixels

arranged on a 3 3 mm2 surface. Three bond pads of usual size are available for
every terminal in parallel. Although no DC supply voltages are available (except
the GND potential) electrostatic discharge (ESD) protection of all input terminals
is supported.
17.3 Function and Characteristics of the Pixel (Active Receptor)
Figure 17.8 shows the pixel cell schematic. A logarithmic photosensor is used [17]
to obtain a maximized dynamic range of approximately seven decades. Although
this kind of image sensor is not the first choice for lowest fixed pattern noise, the
quality of the sensor is not the bottleneck in the overall stimulation homogeneity.
The electrode properties and the interface to the tissue also cause a variation in the
amount of charge transferred. In addition, it is known that the visual system itself
has a temporal high-pass characteristic, which is assumed to cancel out local
sensitivity variations to a certain extent. Power and space limitations make the
mentioned photosensor technology the best choice for this kind of application.
Sensor output is amplified by a differential amplifier, which was adjusted by
a reference voltage VGL supplied externally in the earlier generation of implants.
374 A. Rothermel
Fig. 17.8 Schematic of pixel cell
Fig. 17.9 Schematic of automatic brightness adjustment
The maximum stimulation current also is controlled externally and adjusted to the
individual patient’s sensitivity.
Addressing signals allow selective addressing of pixels as mentioned earlier.
Transistors M8 to M14 form the output-inverting driver. As VH and VL change
polarity, the output driver polarity is automatically inverted. If VH is positive, the
dotted lines show how the output current of the differential amplifier is mirrored to
a positive output current of the pixel cell. If VH is negative and VL positive, the solid
arrows indicate the current mirror delivering a negative output stimulation polarity.
The most recent chip generation includes an automatic ambient-brightness

adaptation, which is shown in Fig. 17.9 [19]. The signal of every fourth pixel in
both spatial directions is read out, which gives 100 measuring points on the chip. An
analog differential difference amplifier is incorporated to still allow fine adjustment
of the sensitivity of the logarithmic sensor.
17.4 The Pixel–Retina Contact: Electronics and Physiology

of the Interface Between Silicon and Nerve Cells
An advantageous property of the subretinal approach lies in the fact that the natural
adhesion of the retinal tissue to the sclera is used to form the contact of the nerve
cells to the stimulating electrodes (see Fig. 17.4). No additional mechanical
measures are required. The surgical procedure to place the device beneath the
retina requires specialized know-how, however, if the placement is successfully
carried out, proximity of the tissue to the stimulator is very close.
The electrical contact between the electronics and the nerve cells is established
by conducting ceramics such as titanium nitride or iridium oxide. Compared to
approaches using a silicon dioxide interface [20], which is chemically inactive, the
porous surface of the ceramics allows a much larger capacitance per unit area,
which is essential for achieving a reasonable spatial resolution for the stimulation
(Fig. 17.10) [9].
Fig. 17.10 Porous electrode for subretinal stimulation

376 A. Rothermel
Fig. 17.11 Optical coherence tomography cross section showing the retinal layer in close
proximity to the subretinal chip [11]
Experiments with in vitro retina from rats [4] and acute human experiments [8]
have proven that with this kind of electrodes and stimulation voltages below 2 V
visual sensations, called phosphenes, can be reliably generated.
To guarantee a long lifetime of the electrode material, charge balancing is an
important electronic challenge. Many proposals have been made to control the
charge transferred to the nerve tissue [21]. These include active charge balancing by
measuring the residual voltage on the electrodes, controlling the injected charge,
and active discharge of the electrodes after stimulation. As long as the electrode
works in the safe voltage limits, it behaves like a capacitance and no ions are
exchanged with the tissue layers.
In subretinal implants, it is essential that the device placed beneath the retina is
as thin as possible. As can be seen in Fig. 17.11, the retina attaches itself closely to
the implanted chip. No additional measures for holding the implant close to the
nerve cells are required. The electrodes have to be flat therefore. However,
maximizing the charge transfer calls for a rough structure, as depicted in Fig. 17.10.
17.5 The Surgical Procedure
Among the implantations of retinal prostheses, the subretinal surgery is possibly the
most delicate one. Nevertheless, a series of 14 implantations have been carried out
by the German team in Tuebingen without the occurrence of adverse effects.
Therefore, the surgery can be considered to be safe and reproducible.
Figure 17.12 (from [22]) shows the major steps. First, the vitreous body has to be
removed from the eyeball. The extraction of the vitreous body is a standard
procedure in eye surgery manipulating the retina for various reasons. In a second
Fig. 17.12 Surgical procedure for subretinal implantation [22]
step, the retinal layer is detached from the background of the eyeball. Next, the
subretinal implant can be safely inserted into the gap between the retina and the
sclera. Finally, the retina is attached to the background of the eyeball again. Please
note that most of the steps carried out here are surgical practices developed for
different retinal diseases.
The ribbon that supplies the subretinal device is firmly attached to the outer part
of the eyeball in such a way that no mechanical stress is introduced. The retina
attaches the complete subretinal part of the implant to the background of the eyeball
in a very natural way.
17.6 Results Achieved To-Date
A large variety of results have been achieved today with patients with subretinal
implants. This includes the responses of blind patients to different amplitudes of
stimulation, different pulse patterns, and different pulse frequencies. Generally, the
378 A. Rothermel
perception is described as white to yellow in color. The intensity can reliably be

varied, and maximum intensity is sometimes described to be very bright.
Different intensities, experimentally represented by sheets with a different gray
color, can be distinguished very well. The orientation of a stripe pattern can be
reliably recognized, as well as motion of stripes and other patterns.
One patient, who was blind for more than 10 years, demonstrated an extraordi-
narily high spatial resolution. Figure 17.13 shows examples of objects he was able
to distinguish. He was even able to read the letters shown in the figure [8, 9].
Besides mug and plate, he could distinguish spoon and knife.
The only other group that consistently reports about visual perceptions in
implanted patients, to the best of the author’s knowledge, is Second Sight [16].
They show patients walking along a wide white line drawn on the floor, or they also
show the perception of letters (Fig. 17.14).
Fig. 17.13 Blind patient with subretinal implant
Fig. 17.14 Letter recognition reported by the epiretinal approach (Second Sight Inc.)
17.7 Future Developments: What Can We Expect

by the Year 2020?
Today, there are mainly two successful approaches, namely the epiretinal and
subretinal approaches. Which of the two approaches will be a product in the future
is difficult to predict, especially for somebody who is working for one of the two.
There is also a certain chance that we may have two products that have different
benefits for the different medical conditions of the patients.
It is very likely that we will have a fully self-contained implant that is
completely located on the eyeball. Wireless inductive power and data transmission
directly to the eyeball is available already today, e.g. in the Argus II prototype by
Second Sight Inc. This type of power transmission requires a transmitting coil,
which may be located on the frame of a pair of glasses the patient is wearing.
Epiretinal implants in all cases work with an external camera. To adjust the
picture that is stimulated on the retinal ganglion cells to the gaze of the patient, eye
tracking is required. Also for eye tracking, it will be necessary for the patient to
wear a pair of glasses. A small camera tracking the eye position will be mounted on
the frame. This means, in conclusion, that the epiretinal approach will always rely
on the patient wearing a pair of glasses, and therefore it is very likely that a wireless
transmission to the eyeball will be the solution for the future even in 2020.
For the subretinal approach, a similar setup is highly probable. The problem of
power and data transmission is even relaxed, because only control data have to be
transmitted, as the video data are generated by the active pixel array in the eye
itself. In addition, the power consumption of subretinal approaches is relatively
small up to now.
But there are additional, cosmetic considerations to be taken into account. For
the subretinal approach, a pair of glasses is not essential and it might be discarded
for some patients. The natural path of the light into the eyeball is still in use with
subretinal prosthesis. Optical correction is not necessarily an issue, because typi-
cally the lens of retinitis pigmentosa patients is replaced anyway, as most of them
suffer from cataract.
In conclusion, the subretinal approach has the potential of giving visual
sensations to patients without them wearing any artifacts in front of their eyes. In
that case, it might be of interest to supply the subretinal implant in the same way as
is done today, by a wire from a receiving box located behind the ear. The trans-
cutaneous transmission to that receiving box uses the well proven technology
applied for the hundreds of thousands of cochlear implants in use today.
The experience of the subretinal group is that the wire does not cause any
uncomfortable sensations to the patients. It is interesting, however, that also the
epiretinal group with Second Sight Inc. has good long-term experience with the
same approach with Argus I. At least two patients with such an implant have had no
problems with the wired connection for 7 years now. Also the approach has proven
to give a reliable long-term lifetime.
380 A. Rothermel
Please note that the optic nerve and muscles are connected to the eyeball
anyway, therefore the additional wire does not disturb as much as one might
suppose at first glance. Maybe in 2020 patients will have the choice between a
device in the form of a pair of glasses, on the one hand, and an alternative device
that requires a transmitter coil magnetically attached behind the ear, and nothing
else.
In general, the subretinal approach has the limitation that the complete subretinal
artifact has to have a certain form factor in order not to disturb the delicate retinal
tissue with its supporting blood vessels too much. Epiretinal implants generally
have a greater freedom of choice with respect to the shapes of wiring and
electrodes. It is possible that the ability of the subretinal approach to cope with
these form factor restrictions will finally define the winning approach.
In terms of improvements of electronics and electrode materials, we can expect
the following developments. The most critical parameter in all retinal implants
is the spatial resolution. Therefore, the teams following the epiretinal approach
have continuously increased the number of electrodes (see [7] for example). As the
stimulating chip and the electrodes have to be connected by wires, the wiring inside
the eyeball turns out to be a critical issue. Argus II, which is the state-of-th- art
prototype used by Second Sight Inc., uses 60 electrodes. Prototypes with higher
numbers of electrodes have been developed also by Liu’s group [23, 24]. Another
idea to increase the number of electrodes is the use of multiplexers close to the
electrodes. However, the technology of connections remains complicated.
The subretinal approach, however, has a different problem. The transferable
charge per unit area has to be maximized to get the highest possible spatial
resolution. On the other hand, it should be noted that the subretinal device does
not have to be designed for a certain pixel density. We shall try to explain this
surprising statement in the following.
The electrodes of a subretinal approach form an electrically active stimulating
layer, which is organized like being cut into pieces. Assume the pieces were to be
cut again, making the electrodes smaller. If the “fill factor” of electrode material
remains constant, basically no restrictions due to the smaller individual electrodes
arise. What happens is comparable to increasing the density of the two-dimensional
spatial sampling pattern of the electrodes. The result of this consideration is that the
electrodes themselves can be arbitrarily small without compromising the stimula-
tion strength. Of course, not one electrode alone is able to elicit visual sensations
in that case; a large enough light spot is required to generate enough charge to
result in a visible object. Still, such a device would have the advantage that every
patient would get the optimal resolution his retina is able to deliver to him.
Already today, subretinal implants use 1,500 electrodes, which allows a spatial
resolution above the resolution of the electrode–tissue charge-transfer system we
see today. It can be expected that the subretinal designs in 2020 will easily have
electrode numbers in the range of 10,000, although not necessarily delivering the
equivalent spatial resolution to all patients.
What should be solved by 2020 in addition is the increase of the “window size”
of retinal implants open for the blind patients. Here, maybe the subretinal approach
has the biggest surgical issues, as it requires a device that must be flexible in two
dimensions to be placed beneath the retina and safely away from the optic nerve.
Epiretinal stimulators may have easier access to the ganglion cells in the periphery,
however, the wiring problem worsens.
Today, it is an open question whether stimulation in the periphery will be
beneficial for retinitis pigmentosa patients. We know that the retinal signal pro-
cessing in the area of sharpest vision and in the periphery differs significantly,
therefore the spike trains of peripheral ganglion cells may be interpreted quite
differently by the brain. Only more human experiments will answer those questions
and help in designing the future implant devices.
Acknowledgments We acknowledge inspiration and continuous support by Prof. E. Zrenner

from the University Eye Hospital Tuebingen and his team. We also acknowledge support by
Dr. Wrobel, Dr. Harscher, and Mr. Kibbel from Retina Implant AG in Reutlingen, and Dr. Stett
and Dr. Gerhardt from the Natural and Medical Sciences Institute at the University of Tuebingen.
References
1. Sharma, R.K., Ehinger, B.: Management of hereditary retinal degenerations: present status and
future directions. Surv. Ophthalmol. 43, 377 (1999)
2. Berson, E.L.: Retinitis pigmentosa: the Friedenwald lecture. Invest. Ophthalmol. Vis. Sci. 34,
1659 (1993)
3. Pollack, J.: Displays of a different stripe. IEEE Spectrum 43(8), 40 (2006)
4. Gerhardt, M., Stett, A.: Subretinal stimulation with hyperpolarising and depolarising voltage
steps. In: Proceedings of the 6th International Meeting on Substrate-Integrated Micro Elec-
trode Arrays, Reutlingen, Germany, pp. 144–147, July 2008
5. Stett, A., Mai, A., Herrmann, T.: Retinal charge sensitivity and spatial discrimination obtain-
able by subretinal implants: key lessons learned from isolated chicken retina. J. Neural Eng. 4,
7 (2007)
6. Sivaprakasam, M., Liu, W., Humayun, M., Weiland, J.: A variable range bi-phasic current
stimulus driver circuitry for an implantable retinal prosthetic device. IEEE J. Solid-St. Circ.
40, 763 (2005)
7. Ortmanns, M., Rocke, A., Gehrke, M., Tiedtke, H.J.: A 232-channel epiretinal stimulator
ASIC. IEEE J. Solid-St. Circ. 42, 2946 (2007)
8. Zrenner, E., et al.: Blind retinitis pigmentosa patients can read letters and recognize the
direction of fine stripe patterns with subretinal electronic implants. 2009 Association for
Research in Vision and Ophthalmology (ARVO), Annu. Mtg., Fort Lauderdale, FL, D729,
p. 4581, May 2009
9. Zrenner, E., et al.: Subretinal microelectrode arrays allow blind retinitis pigmentosa patients to
recognize letters and combine them to words. In: 2nd International Conference on Biomedical
Engineering and Informatics, 2009, BMEI ’09. doi: 10.1109/BMEI.2009.5305315
10. Theogarajan, L., et al.: Minimally invasive retinal prothesis. In: IEEE International Solid-State
Circuits Conference 2006, ISSCC 2006 Digest Technical Papers, San Francisco pp. 99–108,
Feb. 2006
11. Rothermel, A., Liu, L., Pour Aryan, N., Fischer, M., Wuenschmann, J., Kibbel, S., Harscher,
A.: A CMOS chip with active pixel array and specific test features for subretinal implantation.
IEEE J. Solid-St. Circ. 44, 290 (2009)
12. http://www.bionicvision.org.au/ Accessed April 2011
382 A. Rothermel
13. Fujikado, T., et al.: Evaluation of phosphenes elicited by extraocular stimulation in normals
and by suprachoroidal-transretinal stimulation in patients with retinitis pigmentosa. Graefe’s
Arch. Clin. Exp. Ophtalmol. 245, 1411 (2007)
14. Zrenner, E.: Will retinal implants restore vision? Science 295, 1022 (2002)
15. Schubert, M., Stelzle, M., Graf, M., Stert, A., Nisch, W., Graf, H., Hammerle, H., Gabel, V.,
Hofflinger, B., Zrenner, E.: Subretinal implants for the recovery of vision. In: 1999 IEEE
International Conference on Systems, Man, and Cybernetics, IEEE SMC ’99 Conference
Proceedings, Vol. 4, pp. 376–381 (1999)
16. Horsager, A., Greenberg, R.J., Fine, I.: Spatiotemporal interactions in retinal prosthesis
subjects. Invest. Ophthalmol. Vis. Sci. 51, 1223 (2010)
17. Dollberg, A., Graf, H.G., H€offlinger, B., Nisch, W., Schulze Spuentrup, J.D., Schumacher, K.:
A fully testable retina implant. In: Hamza, M.H. (ed.) Proceedings of the IASTED Interna-
tional Conference on Biomedical Engineering, BioMED 2003, pp. 255–260. ACTA Press,
Calgary, (2003)
18. Rothermel, A., Wieczorek, V., Liu, L., Stett, A., Gerhardt, M., Harscher, A., Kibbel, S.:
A 1600-pixel subretinal chip with DC-free terminals and 2 V supply optimized for long
lifetime and high stimulation efficiency. In: IEEE International Solid-State Circuits Confer-
ence 2008, ISSCC 2008. Digest of Technical Papers, San Francisco, pp. 144–602 (2008)
19. Liu, L., Wunschmann, J., Aryan, N., Zohny, A., Fischer, M., Kibbel, S., Rothermel, A.: An
ambient light adaptive subretinal stimulator. In: Proceedings of 35th European Solid-State
Circuits Conference, ESSCIRC ’09. pp. 420–423. doi: 10.1109/ESSCIRC.2009.5325980
20. Fromherz, P.: Joining microelectronics and microionics: Nerve cells and brain tissue on
semiconductor chips. In: 33 rd European Solid-State Circuits Conference, ESSCIRC 2007.
pp. 36–45. doi: 10.1109/ESSCIRC.2007.4430245
21. Sooksood, K., Stieglitz, T., Ortmanns, M.: Recent advances in charge balancing for
functional electrical stimulation. In: 2009 Annual International Conference of the IEEE
Engineering in Medicine and Biology Society, EMBC 2009, pp. 5518–5521. doi: 10.1109/
IEMBS.2009.5333181
22. Wilke, R., Zrenner.E.: Clinical results, thresholds and visual sensations elicited by subretinal
implants in 8 blind patients. In: The Eye and The Chip World Congress 2008, Detroit, June
2008. http://www.eyeson.org/documents/uploads/EyeandTheChipProgramBook_2008.pdf
23. Chen, K.F., Liu, W.T.: Highly programmable digital controller for high-density epi-retinal
prosthesis. In: 2009 Annual International Conference of the IEEE Engineering in Medicine
and Biology Society, EMBC 2009, pp. 1592–1595. doi: 10.1109/IEMBS.2009.5334120
24. Weiland, J.D., Fink, W., Humayun, M.S., Liu, W.T., Li, W., Sivaprakasam, M., Tai, Y.-C.,
Tarbell, M.A.: Systems design of a high resolution retinal prosthesis. In: IEEE International
Electron Devices Meeting, 2008, IEDM 2008. doi: 10.1109/IEDM.2008.4796682
Chapter 18
Silicon Brains
Bernd Hoefflinger
Abstract Beyond the digital neural networks of Chap. 16, the more radical
mapping of brain-like structures and processes into VLSI substrates has been
pioneered by Carver Mead more than 30 years ago [1]. The basic idea was to
exploit the massive parallelism of such circuits and to create low-power and fault-
tolerant information-processing systems.
Neuromorphic engineering has recently seen a revival with the availability of
deep-submicron CMOS technology, which allows for the construction of very-
large-scale mixed-signal systems combining local analog processing in neuronal
cells with binary signalling via action potentials. Modern implementations are
able to reach the complexity-scale of large functional units of the human brain,
and they feature the ability to learn by plasticity mechanisms found in neuroscience.
Combined with high-performance programmable logic and elaborate software
tools, such systems are currently evolving into user-configurable non-von-Neumann
computing systems, which can be used to implement and test novel computational
paradigms. The chapter introduces basic properties of biological brains with
up to 200 Billion neurons and their 1014 synapses, where action on a synapse
takes 10 ms and involves an energy of 10 fJ. We outline 10x programs on
neuromorphic electronic systems in Europe and the USA, which are intended to
integrate 108 neurons and 1012 synapses, the level of a cat’s brain, in a volume of
1 L and with a power dissipation <1 kW. For a balanced view on intelligence, we
references Hawkins’ view to first perceive the task and then design an intelligent
technical response.
B. Hoefflinger (*)

384 B. Hoefflinger
18.1 Some Features of the Human Brain
The human brain consists of about 200 billion neurons, each connected via
thousands of synapses to other neurons, forming an incredibly powerful sensing,
information, and action system that has
– Unparalleled intelligence, i.e., learning, comprehension, knowledge, reasoning,
planning, and
– The unique capability of cognition, i.e., the power to reflect upon itself, to have
compassion for others, and to develop a universal ethic code.
We settle for the intelligence aspect in this chapter, and we identify the following
tasks as a collective motivation to understand and to use brain-like structures and
operations:
– All kinds of feature- and pattern-recognition,
– Understanding many-body scenes, objects, people, traffic,
– Language understanding and translation,
– Building a relationship with care-bots.
We can imagine that the first three tasks can be handled by neural networks as
described in [1] or in Chap. 16. Building a relationship with a care-bot requires a
complexity of learning and adaptation plus the incorporation of feeling and com-
passion such that we have to look for the uniqueness of the brain to solve this task.
Figure 18.1 shows a micrograph of a part of a mouse brain with the central
neuron green-enhanced. Each neuron consists of a nucleus (soma), dendrites, axons,
and synapse barriers (Fig. 18.2; for an introduction see [4]).
Fig. 18.1 A section of a

mouse brain under the
microscope. The scale bar is
100 mm. The green neuron in
the center was made visible
by a green-fluorescent
protein [2]
18 Silicon Brains 385
Dendrite
Synapse
Axons
Synapse
Fig. 18.2 Components of a neuron [3]
Nucleus: Receives and adds-up inputs from dendrites and sends outputs through
axons. Has a nonlinear response curve with thresholds, above which a spike is
generated.
Axons: Receive signal from nucleus and conduct ionic signal to synapse barrier.
Dendrites: Receive ionic signals from synapses.
Synapses: Consist of a presynaptic ending and a postsynaptic ending, separated by a
gap of a few nanometers. Neurotransmitters are released across the gap depending
on previous history and input signal, for which a simple model is the multiplica-
tion of the signal by a characteristic weight (see Fig. 16.1 for a comparison).
Contrary to the one-way feed-forward perceptrons in Chap. 16, two-way signal
transport can occur in natural neurons, and new synapses with axons and dendrites
can grow in a biological brain. Neurons are also specialized for certain tasks [5]:
– Sensor neurons sense inputs. The most prominent example is the retina
(Chap. 17).
– Inter-neurons are the large majority operating inside the brain.
386 B. Hoefflinger
– Memory neurons are a more recently defined specialty of inter-neurons, which

has strong reinforcement and long-term-potentiation (LTP) features.
– Motor neurons perform control and actuator functions.
The recent explosion of knowledge about the brain’s functions is the result of
the advancement of microelectronics and micro-sensing augmented by image
processing and micro-physical chemistry. Out of the empire of acquired knowledge,
we select a few topics, which are essential in developing biomorphic Si brains:
– The key information lies in the action potential, both in its height and in its
frequency of occurrence. The action potential is a voltage spike 0–70 mV high
and ~2 ms wide. The repetition rate can be 0.1–500 Hz.
– The resting potential across the gap of a synapse is 70 mV. When an ionic charge
arrives, a specific action potential is generated, reducing the voltage difference.
– The signal-transmission strength is given by the spike density, which resembles
a pulse-density-modulation code. The rate is <1 Hz to 500 Hz with an average of
10 Hz.
– The synapse weight determines to what degree the incoming signal at a synapse
is transferred to the postsynaptic output onto a dendrite.
Table 18.1 is a performance guide to the human brain. The basis of this charge
model is a charge density of 2 107 A s cm2 per neuron firing, fairly common to
biological brains [6] and a geometric mean for axon cross sections, which vary from
0.1 mm2 to 100 mm2.
We close this section with a brief summary on the circuit functions of the neural
elements.
– Synapses: They receive as input signal a pulse-density-coded signal with a
dynamic range of 0.1–500 Hz, equivalent to a 10-bit linear code, which could
be log-encoded for efficiency and physical meaning. They transmit an output to
the dendrite, which is weighted by the history of this connection, which means
training, memory, recapitulation, and reinforcement, certainly the most signifi-
cant and complex operation. Implementations require a multi-level, non-
volatile, up-datable memory for the weight-function, which, in simple terms,
Table 18.1 Charge and energy model of the human brain

Number of neurons 2 1011
Synapses/neuron 103
Ionic charges per neuron firing 6 1011 A s
Mean cross section synaptic gap 30 mm2
Charge/synaptic gap 6 1014 A s ¼ 4 105 ions
Action potential 70 mV
Energy per synapse operation 4 fJ ¼ 2.5 104 eV
Energy per neuron firing 4 pJ
Average frequency of neuron firing 10 Hz
Average brain power (2 1011) (4 1012) 10 ¼ 8 W
then has to be multiplied with the input signals. Thus, the electronic synapse is a
complex, most-likely mixed-signal circuit with an intelligent memory.
– Dendrites: They perform the summation of their associated synapse-outputs, a
fairly convenient electronic analog-circuit operation, but also common with
digital accumulators.
– Nucleus: Its function is the firing decision, a nonlinear characteristic with
hysteresis and historic adaptation (see the digital solutions in Chap. 16).
Beyond the dedicated technical implementations in Chap. 16, we introduce two
generalist programs to build Si brains in the following.
18.2 3D Analog Wafer-Scale Integration of a Si Brain
In 2005, a consortium of 14 university and research centers in Europe started the

FACETS project (Fast Analog Computing with Emergent Transient States) [7].
Within its broad scope of advancing neuromorphic computing, the hardware part is
a very-large-scale, mixed-signal implementation of a highly connected, adaptive
network of analog neurons. The basic element is a wafer-scale network of 50 million
synapses, real-time configurable with
– 150 k neurons and 256 synapses each, to
– 2.3 k neurons and 16 k synapses each.
One key target is a 105-fold speed-up of the natural neuron-firing rate of
10 Hz. Since the communication of firing events is digital with a minimum of
16 bits/event, communication bandwidth is a major challenge for this system. The
wafer is organized into 288 chips with 128 k synapses and 16 k inputs each,
resulting in a total of 4.6 M inputs 105 10 Hz ¼ 4.6 Tera events/s ¼ 74 Tb/s
per wafer. This traffic is organized by a wafer-level, crossbar metal-interconnect,
fabricated as a custom back-end-of-line structure. The manufacturing of this archi-
tecture would be a perfect example of future high-speed, maskless electron-beam
lithography (Chap. 8). It is also a perfect example of at least three other challenges
for the advancement of chip technology.
One is the power efficiency of chip/wafer-level interconnects. With the optimum
efficiency of 1 mW/(Gb/s) in 2010 (Chap. 5), the resulting 74 W could be handled.
The more likely level in the project at a less advanced technology node would be
1 kW. Scaling this up to mega-neurons clearly shows that power efficiency is the
number 1 concern for these complex systems.
A second challenge is the implementation of a writable and non-volatile synapse
weight. The FACETS hardware designers have opted for a multi-level (MLC)
floating-gate memory, a nice example of utilizing the tremendous development
activity of the industry in this direction (MLC Flash, Chap. 11). Providing this
specific chip technology in congruence with the underlying synapse would make a
natural application for 3D integration.
388 B. Hoefflinger
Fig. 18.3 View of the FACETS wafer-level mixed-signal brain-element with 150 k neurons and
50 M synapses [8]
This 3D integration is the third challenge for any Si brain. The memory Si layer
on top of the mixed-signal Si layer can be achieved with a through-silicon-via (TSV)
technology (Chaps. 3.7, 7, 12), and it is only one level in the task of building the
whole system, because the global programmable digital control could be added on
top. In the FACETS architecture, the digital control is implemented with FPGA
(field-programmable logic array) chips on a printed-circuit board with the wafer in
the center. An exploded view of the wafer-scale brain-element is shown in Fig. 18.3.
The chip area/synapse in the FACETS design is ~400 mm2, comprising mostly
analog circuits with about 100 devices (transistors, capacitors, resistors) operating
at 1.8 V. The design and technology rules are given by the European research wafer
service at the 130 nm level. As of 2011, Facets has been followed by BrainScaleS
[8]. There is clearly room for scaling, and it will be interesting to follow the digital-
versus-analog strategy, considering the alternative of digital 8 8 multipliers with
1 fJ and 100 mm2 per multiplication (see Sect. 3.6).
The FACETS project has advanced high-speed vision tasks as a benchmark.
Another application and a matter of cross-fertilization with the leaders in super-
computer-based simulations of the human brain is the use of the FACETS system
as a hardware accelerator for specific tasks. If a Blue Brain [9] simulation takes
100 times longer than a natural task, and the hardware could be 105 times faster than
nature, this cooperation would accelerate the acquisition of know-how about some
workings of the human brain quite significantly. This is also one of the aims of
another large-scale project.
18.3 SyNAPSE: Systems of Neuromorphic Adaptive Plastic

Scalable Electronics
This US program was launched by DARPA in 2009 [10], and it says in its
description: “As compared to biological systems. . ., today’s programmable
machines are less efficient by a factor of 1 million to 1 billion in complex,
real-world environments”. And it continues: “The vision . . . is the enabling of
electronic neuromorphic machine technology that is scalable to biological levels. . ..
Biological neural systems (e.g. brains) autonomously process information in com-
plex environments by automatically learning relevant features and associations.”
Table 18.2 gives an overview of the capacity of biological brains and a compari-
son with two major electronics programs.
The SyNAPSE program is structured into four phases of ~2 years each:
1. The entry-level specification for synapse performance is:
Density scalable to >1010 cm2, <100 nm2.
Energy per synaptic operation <1 pJ, <100 nature.
Operating speed >10 Hz (equivalent to nature).
Dynamic range of synaptic conductance >10.
2. (~2010/2012) Specify a chip fabrication process for >106 neurons/cm2,
1010synapses/cm2.
Specify an electronics implementation of the neuromorphic design methodology
supporting >1010 neurons and >1014 synapses, mammalian connectivity,
<1 kW, <2 Liter (the final program goal).
3. (~2012/2014) Demonstrate chip fabrication >106 neurons/cm2, 1010 synapses/cm2.
Demonstrate a simulated neural system of ~106 neurons performing at “mouse”
level in the virtual environment.
Table 18.2 Brain characteristicsa

Brain Neurons Synapses Frequency [Hz] Power Volume Refs.
operations weight
Human 2 1011 2 1014 0.1–500 8W
Cat 108 1012 10
Mouse 106 1010 10
8
SyNAPSE Goal 10 Fab. 1012 Fab. [10]
10
10 Des. 1014 Des. 1 kW 1 Liter
Sim.2009 108 1012 500 s / 5 s real 3 MW 227 to. [9]
9
FACETS Goal 10 1012 100 kHz [8]
a
Fab.: Fabricate a multi-chip neural system. Des.: Describe a high-level, conceptual electronics
implementation
390 B. Hoefflinger
Table 18.3 Comparison of FACETS and SyNAPSE

FACETS SyNAPSE
Area/synapse 400 mm2 <100 nm2
Frequency of events 1 MHz >10 Hz
Area/frequency 400 nm2/Hz <10 nm2/Hz
4. (~2014/2016) Fabricate a single-chip neural system of ~106 neurons and pack-

age into a fully functional system.
Design and simulate a neural system of ~108 neurons and ~1012 synapses
performing at “cat”-level environment.
5. (~2016/2018) Fabricate a multi-chip neural system of ~108 neurons and instan-
tiate into a robotic platform performing at “cat” level (hunting a “mouse”).
SyNAPSE is a program with explicit specifications. It requires an existing
system-simulation background (like Blue Brain in Lausanne [9] or C2 in Livermore)
to assess the likelihood of realizing the milestones. One striking difference between
FACETS and SyNAPSE is the density-speed figure-of-merit (FOM). See Table 18.3.
SyNAPSE is an industry-driven, much larger program and will generate new
hardware and software technologies. FACETS is based on existing chip technology,
and it is an experiment on the large-scale, international cooperation of research
institutes.
18.4 Outlook
The fascination associated with an electronic replication of the human brain has
grown with the persistent exponential progress of chip technology. The present
decade 2010–2020 has also made the electronic implementation feasible, because
electronic circuits now perform synaptic operations such as multiplication and
signal communication at energy levels of 10 fJ, comparable to or better than
biological synapses. Nevertheless, an all-out assembly of 1014 synapses will remain
a matter of a few exploratory systems for the next two decades because of several
constraints:
Now that we are close to building an equivalent electronic core, we have moved
the fundamental challenge to the periphery, namely the innumerable sensors and
actuators (motors). In this outside-of-electronics domain, that of MEMS, remark-
able progress has been made and continues with high growth (Chaps. 14 and 17).
Furthermore, the critical conversion of analog sensor signals to digital signals is
advancing rapidly (Chap. 4). And yet it is very early in the evolution of generic
neuromorphic peripheries, so that all practical systems will be focused, application-
specific solutions, certainly with growing intelligence, but with confined features of
neural networks.
The other constraint is adaptation and control of brain-like architectures. Fuzzy
and neuro-control have been practically available for 25 years, and yet their
application is still limited to applications that are just feature-oriented and not
critical for the technical performance of a product or service. It will take a major,
globally consolidated effort involving many disciplines from reliability and ethics
to social science to achieve a broad acceptance. The advancement of care-bots in
various world regions will be the test-ground for this evolution of neuromorphic
systems.
The third fundamental consideration is looking at the field of artificial intelli-
gence (AI), where the Si brain is a major constituent, from the opposite end, i.e., not
from the brain-side dealing with an environment, but from the environment, how to
perceive it and how to deal with it. An outstanding leader with this viewpoint, to
make predictions about the world by sensing patterns and then to respond intelli-
gently, is Jeff Hawkins (the inventor of the first personal assistant, the PalmPilot)
[11]. We refer to his work here in order to strike a balance with the hardware
perspective on the Si brain offered in this chapter.
References
1. Mead, C., Ismail M.: Analog VLSI Implementation of Neural Systems, ISBN 978-0-7923-
9040-4, Springer (1989)
2. www.wikipedia.org/Brain
3. Chudler, E.H.: Neuroscience for kids, http://faculty.washington.edu/chudler/synapse.html
(2009)
4. Stufflebeam, R.: Neurons, synapses, action potentials, and neurotransmission. The Mind
Project, www.mind.ilstu.edu/curriculum/neurons_intro (2008)
5. Martini, F.H., Nath, J.L.: Fundamentals of Anatomy and Physiology, Chapter 12: Neural
Tissue, Prentice-Hall (2008)
6. Sengupta, B., Stemmler, M., Laughlin, S.B., Niven, J.E.: Action potential energy efficiency
varies among neuron types in vertebrates and invertebrates. PLoS Computat. Biol (2010). DOI:
10.1371/journal.pcbi.1000840
7. www.facets-project.org
8. http://facets.kip.uni-heidelberg.de
http://brainscales.kip.uni-heidelberg.de
9. www.bluebrain.epfl.ch
10. http://www.darpa.mil/Our_Work/DSO/Programs/Systems_of_Neuromorphic_Adaptive_Plastic_
Scalable_Electronics_%28SYNAPSE%29.aspx
11. Hawkins, J., Blakeslee, S.: On Intelligence. Times Books, Henry Holt, New York (2005)
Chapter 19
Energy Harvesting and Chip Autonomy
Yiannos Manoli, Thorsten Hehn, Daniel Hoffmann, Matthias Kuhl,

Niklas Lotze, Dominic Maurath, Christian Moranz, Daniel Rossbach,
and Dirk Spreemann
Abstract Energy harvesting micro-generators provide alternative sources of

energy for many technical and personal applications. Since the power delivered
by such miniaturized devices is limited they need to be optimized and adapted to the
application. The associated electronics not only has to operate at very low voltages
and use little power it also needs to be adaptive to the fluctuating harvesting
conditions. A joint development and optimization of transducer and electronics is
essential for improved efficiency.
19.1 Introduction
The progress in microelectronic components and battery development has provided

us with portable equipment that is powerful and useful for our daily routine.
Wireless communication makes information available at any time and at any
place and has thus transformed the use of these devices. It is creating a whole
new business where sensor systems can be employed in unforeseen ways. They can
be used in buildings or other structures for improving maintenance, in remote areas
for environmental observation, or in cars for better safety and comfort. In many
cases this improved functionality also leads to savings in energy, e.g., fuel effi-
ciency in cars or intelligent monitoring of buildings.
However, we are observing what can be seen as an energy gap [1]. Although
microelectronic devices are becoming more power efficient, the demand for more
Y. Manoli (*) • T. Hehn • M. Kuhl • D. Maurath • C. Moranz

IMTEK Microelectronics, University of Freiburg, Georges-Koehler-Allee 102, 79110 Freiburg,
Germany
e-mail: manoli@imtek.de; maurath@imtek.de
D. Hoffmann • N. Lotze • D. Rossbach, • D. Spreemann
HSG-IMIT – Institute of Micromachining and Information Technology, Wilhelm-Schickard-
Strasse 10, 78052 Villingen-Schwenningen, Germany

394 Y. Manoli et al.
performance is outpacing the progress in power density in batteries. Batteries have

also the disadvantages of limited lifetime and the cost of replacement, maintenance,
and waste generation. The use of harvesting concepts also in combination with
rechargeable batteries can be a way to alleviate the power dilemma.
A great research effort has evolved in the last decade with the aim of improving
power efficiency and reducing the size of harvesters as well as improving the
power management electronics. We are extending the ideas of renewable resources
to the miniaturized world, employing micro-engineering techniques to convert
different forms of energy into electrical power. These energy sources can be
thermal or mechanical or come in the form of light or as chemical or biological
matter.1 A number of industrially viable solutions are already present on the
market [2, 3].
The next sections give an overview of available energy sources and show
how dedicated energy management circuits can improve the system power efficiency.
19.2 Conversion Mechanisms
A number of different ambient energy sources, such as light, temperature gradients,

motion and vibration, are available. This chapter does not attempt a comparison of
these methods but provides a function and performance overview. What makes
energy harvesting interesting for research but frustrating for the market development
is the fact that it is always very application specific. It is very rarely the case that
a task can be solved by employing a variety of these energy supplies. Environmental
conditions will determine the choice of light, heat, or motion. Further conditions
such as the allowed weight and size of the device will also constrain the options.
Finally, the required system functionality will determine the power requirements
and the selection of a solution. Should a choice of two mechanisms be actually
possible, such as heat and vibration in machines and cars or vibration and light in
industrial areas, then hybrid solutions can be considered. Since some of the sources
might be sporadic and intermittent such a solution might be the only way out.
Although it might be questionable to include radio frequency (RF) scavenging in
this discussion, we do deal with some of the electronic aspects since RF identifica-
tion (RFID) concepts can be adapted to other areas. The same is true in the case of
fuel cells, since most need energy storage. Emerging devices do not require this.
One can perhaps distinguish between generators such as thermal and solar or fuel
cells that provide a DC voltage and those that deliver an AC voltage, as is the case
for RF and vibration harvesters. The latter are often based on resonant systems and
limited to narrowband applications. Thus, great interest exists in developing active
tunable devices [4].
1
This research has been funded in part by the German Research Foundation (DFG), the Federal
Ministry of Education and Research (BMBF), the State of Baden-Wuerttemberg and industrial
partners.
19 Energy Harvesting and Chip Autonomy 395
19.2.1 Motion and Vibration Transducers
When electrical energy is to be harvested from kinetic motion, in particular mechan-

ical vibrations, the choice of a specific transducer mechanism (electromagnetic,
piezoelectric, or electrostatic) is strongly dependent on the operating conditions
(amplitude and frequency spectrum of the excitation source) and the available space
given by the application environment.
The selection process may be further rationalized by considering the scaling
behavior of electromagnetic and electrostatic forces [6]. The electromagnetic
coupling coefficient scales at a different rate to the electrostatic coupling coefficient
[7]. Therefore, when the transducer system decreases by a factor of 100 in size, the
electrostatic coupling coefficient decreases by factor of 100, whereas the electro-
magnetic coupling coefficient decreases by a factor of 1000 [7]. Thus it follows that
the electrostatic conversion mechanism is more efficient for transducer devices with
a size of a typical MEMS (micro-electro-mechanical) device (<100 mm3). On the
other hand, an electrostatic transducer device may be outperformed by an electro-
magnetic transducer for larger device sizes (>1 cm3).
Moreover, based on the fact that the technology for capacitive sensors (e.g.,
accelerometers, gyroscopes) is well established, it is beneficial to utilize the same
standard MEMS process for energy harvesting devices. Thus, if miniaturization of
a kinetic energy transducer is an important issue owing to little space available and
if a large production scale is required, electrostatic MEMS transducers are the
preferred choice. However, it must be considered that the output power of any
inertial energy transducer is proportional to the oscillating mass. Consequently,
miniaturized energy transducers with a very small proof mass will provide only low
power levels in the range of microwatts. Therefore, electrostatic MEMS transducers
require applications that are based on ultra-low-power devices.
In order to convert kinetic vibration energy into usable electrical energy by
means of energy harvesting, two functional mechanisms are required:
• A transduction mechanism to convert mechanical energy into electrical energy
• A mechanical means of coupling ambient vibrations to the transduction
mechanism.
The transduction mechanism enables the transformation of kinetic energy into
the electrical domain. In this respect an electrical current or voltage is generated,
which can be used to power electronic components. For the transduction mecha-
nism to work effectively, a relative motion between two components must be
available. A widespread principle for realizing inertial vibration generators is
based on a mechanical resonator: A proof mass is supported by a suspension,
which is attached to a frame. The oscillations of the proof mass can be damped
by a suitable transducer mechanism and thus kinetic energy is converted into
electrical power.
The mechanical oscillator should be adapted to the most energetic vibration
frequency. However, there is great interest in increasing the bandwidth by using
a piezoelectric harvester based on wideband vibrations [8] or to realize active

tunable devices by tuning the harvester stiffness through piezoelectric actuators
[9, 10]. The task in tunable devices is to be able to invest the power for the control
system without limiting the converted energy too much.
19.2.2 Application-Oriented Design of Vibration Transducers
Electromagnetic resonant vibration transducers are already commercially available

[3, 11, 12]. Commercial vibration transducers are typically add-on solutions,
greater than 50 cm3 and they fulfill standard industrial specifications or operation
conditions (such as temperature range or shock limit).
Nevertheless, requests from the industry show that there is a great demand for
application-specific solutions. This is because for each application the required
output power, the available vibration level, and the overall mass and volume are
significantly different. Moreover, it is often necessary to integrate the system into
an existing subassembly. Especially for applications where the volume of the
transducer is a critical parameter, these facts show that the required power can
only be optimally extracted by application-specific customized developments.
Thus, it can be guaranteed that the transducer is not over-dimensioned with respect
to size, mass, and output performance.
In application-specific customized developments the underlying vibration
source and the required output performance must be taken into account. Because
real vibration sources are often stochastic in nature, the first task arising from this is
to find the most energetic vibration frequency.
This is shown in the example of a real vibration source (car engine compartment)
that has already been considered for energy harvesting applications (Fig. 19.1) [5].
An underlying typical acceleration profile and the corresponding frequency
Fig. 19.1 Vibration measurements in the engine compartment of a four cylinder in-line diesel
engine [5]
Fig. 19.2 (a) Acceleration

profile during country driving
route. (b) Frequency content
of the acceleration profile and
(c) expected output power for
an oscillation mass of 10 g
dependent on the resonance
frequency
spectrum are shown in Fig. 19.2a, b. First of all, some predominant frequencies are
visible that depend on the number of cylinders and the revolutions per minute.
Based on the well-known analytic treatment of vibration transducers it can be
shown that the output power depends on the square of the acceleration amplitude
but not on the vibration frequency. Hence the 2nd and 4th harmonic conversion will
result in the same output power for this example vibration profile (Fig. 19.2c).
Beside the amplitude, the bandwidth of the vibration needs to be taken into account.
In other words, a narrow peak with high amplitude may be disadvantageous
compared to a “plateau” with slightly smaller amplitudes. The mechanical oscilla-
tor should be adapted to the most energetic vibration.
19.2.2.1 Electromagnetic Generators
Since the earliest electromagnetic vibration transducers were proposed in the mid
1990s [13], a multiplicity of prototype transducers differing in size, electromagnetic
coupling, excitation conditions, and output performance have been developed by
numerous research facilities.
The data of published electromagnetic vibration transducer prototypes are wide-
spread from 0.1 to 100 cm3 and 1 mW to 100 mW. Voltage levels well below 1 V are
in general barely applicable because of the dramatic drop in efficiency after rectifi-
cation [10]. Microfabricated devices often have an output voltage well below
100 mV.
To date, no testing standard for vibration energy harvesting devices exists.
Hence it is rather difficult to compare the transducers presented in the literature.
The devices have been tested under different conditions and important parameters
for comparison are often omitted. Nevertheless, an extensive review of electromag-
netic vibration transducers has been presented [4] (Fig. 19.3), where the theoretical
scaling law is observable even though there are orders of magnitudes between the
normalized power densities (power density related to the excitation amplitude in
100
Previously reviewed
10-01
Further prototype transducers
10-02 Theoretical scaling
Norm. Power Density (W.g2 / cm3)
10-03
10-04
10-05
∝ L2
10-06
10-07
Fig. 19.3 Normalized power 10-08 ∝ L4

density of existing vibration 10-09
transducer prototypes. The
dashed lines indicate the 10-10
theoretical scaling in the 0.01 0.1 1 10 100 1000
earlier review [4] Volume (cm3)
Fig. 19.4 Coupling Ro =Rmag

architecture used for the
design of the vibration
transducer Ri
hcoil
Coil
zmax
Magnet
hmag
N
Construction volume
g ¼ 9.81 m/s2) of the published prototypes. At very small sizes L (linear dimension)
the electromagnetic vibration transducers scale with L4, and for large L, with L2.
A possible reason why the published prototype transducers have significantly
different output performances is that many different electromagnetic coupling
architectures have been applied. A comparison of the maximum output perfor-
mance of the coupling architectures (one shown in Fig. 19.4) is presented in [14],
where the dimensions of the components have been optimized based on applica-
tion-oriented boundary conditions. Such optimization calculations are necessary in
order to achieve the highest conversion efficiencies. A further enhancement of the
transducers would result with new magnetic materials. The development potential
of (BH)max for permanent magnets is given in [1].
19.2.2.2 Piezoelectric Vibration Conversion
Piezoelectric materials are widely used in energy harvesters because of their ability
to convert strain energy directly into electric charge and because of the ease with
which they can be integrated into a harvesting system. The energy conversion
process occurs because piezoelectric materials show a local charge separation,
known as an electric dipole. Strain energy that is applied through external vibration
results in a deformation of the dipole and the generation of charge, which can be
transferred into a buffer capacitor, powering electronic devices.
Several piezoelectric materials appropriate for energy harvesting devices have
been developed. Most common is a ceramic made of lead zirconate titanate (PZT).
Although it is widely used in power harvesting applications, its brittle nature causes
limitations in the strain that it can absorb without being damaged. Cyclic loading
might cause the material to develop fatigue cracks, resulting in the device being
irreversibly destroyed [13]. Thus, more flexible materials such as poly(vinylidene
Metal layer
Piezoelectric layer
a b c
+ + +
–
–
–
Fig. 19.5 Three types of piezoelectric sensors: (a) a series triple layer (bimorph), (b) a parallel
triple layer (bimorph), and (c) a unimorph
fluoride) (PVDF), a piezoelectric polymer, have gained importance over the last
few years. Thus, larger strains can be applied, increasing the amount of energy
harvested from the piezoelectric device [15, 16].
Combining several piezoelectric layers is another strategy for increasing the
amount of harvested energy. Instead of using a single piezoelectric layer in bending
mode (unimorph), bimorphs consist of two piezoelectric layers connected electri-
cally in series or in parallel. The different configurations are depicted in Fig. 19.5.
Experiments have shown that for medium and high excitation frequencies, the
output power of bimorph harvesters outperforms that of unimorphs [17].
19.2.2.3 Electrostatic Energy Conversion
The idea of electrostatic energy conversion goes back to 1976 with a rotary non-
resonant electrostatic conversion system [18]. About 20 years later, the electrostatic
and other conversion mechanisms were combined with a mechanical resonator in
order to harvest electrical energy from kinetic motions [13]. Since then a number of
research groups have been working on different electrostatic schemes for kinetic
energy harvesting applications [7, 19–22].
The energy conversion mechanism of an electrostatic transducer is based on the
physical coupling of the electrical and mechanical domains by an electrostatic
force. The electrostatic force is induced between opposite charges stored on two
opposing electrodes. In order to convert mechanical energy into electrical energy
by means of the electrostatic transduction mechanism a variation of capacitance
over time must occur [20, 23]. Despite the progress made in electrostatic energy
harvesting devices, an energy-autonomous sensor system using this technology has
not been demonstrated so far.
Figure 19.6 shows a microscope close-up view of a fabricated electrostatic
transducer device, including one of the four mechanical suspension units and one
of the ten electrode units. Each electrode unit comprises two comb electrodes,
which consist of a number of interdigitated electrode elements.
Fig. 19.6 Microscopy image

showing a detail of the
electrostatic transducer
structure, including one of the
mechanical suspensions and
the electrode units. The inset
(upper right) shows a detailed
view of the interdigitated
electrodes, which implement
an area-overlap characteristic
19.2.3 Thermal Generators
Miniaturized devices with typically thin-film thermoelectric layers such as bismuth

telluride-related compounds on silicon substrates in common vertical architecture
are being successfully commercialized [24]. The performance data of these devices
such as DT and heat flux densities are related to the thermoelectric performance of
the materials, which are characterized by the dimensionless figure of merit ZT.
Here, the figure of merit Z for thermoelectric devices is Z ¼ sS2/k, where s is the
electrical conductivity, S the Seebeck coefficient, and k the thermal conductivity.
Although there is no theoretical limit for ZT, thermoelectric bulk materials have not
significantly exceeded ZT 1 for almost 50 years. Typical characteristics of
miniaturized devices at DT of 40–70 K are a few milliwatts [25].
19.2.4 Fuel Cells
At first glance fuel cells (FCs) do not seem to be energy harvesting devices in the
classical sense, as they do not interact with the ambient medium to convert its
energy into an electrical current. This might be true for high-power FCs utilizing
fuel reservoirs, pumps, pressure reducers, and cooling devices with a power output
of up to 300 W [26], but, concentrating on the micropower scale, several systems
have been developed in recent years that are worth mentioning as energy harvesting
devices – or at least in combination with them. Two designs of micro-FCs will be
presented in detail in this section and some future prospects will be discussed.
19.2.4.1 Polymer Electrolyte Membrane Fuel Cells
Conventional PEM FCs consist of a polymer electrolyte membrane (PEM), two gas
diffusion electrodes, two diffusion layers, and two flow fields. The reactants, e.g.,
hydrogen and oxygen, are supplied to the gas diffusion electrodes over feed pipes
out of external tanks. The amount of supplied fuel is often controlled by active
system periphery, such as pressure reducers and valves. As most of the energy
harvesting devices are aiming towards autonomous microsystems, a new setup
principle of PEM FCs was developed [27], enabling the miniaturization of its
components. This work concentrates on the possibility of building FCs integrated
into a well-established standard CMOS process. These chip-integrated fuel cells
(CFCs) are made up of palladium-based hydrogen storage and an air-breathing
cathode, separated by a PEM. The cross section of a CFC is depicted in Fig. 19.7.
Advantages of the new approach are the omission of active devices for fuel supply
and the reduction of system components such as flow fields and diffusion layers.
Due to the simple assembly process, the fuel cells can be produced by thin-film or
thick-film technologies, which fulfill all requirements for a post-CMOS fabrication.
While the CFC itself has no conductive path to the CMOS chip, both the anode and
the cathode can be connected to the CMOS circuit using wires and vias. Further-
more, the single fuel cells can be cascaded to generate multiples of the nominal
750 mV open circuit voltage. So-called integrated fuel cell cascades have been
presented with an open circuit voltage of 6 V and a maximum power density of
450 mW/cm2 [28] (Fig. 19.7).
19.2.4.2 Biofuel Cells
Other approaches to micro-FCs can be found in medical research. Biofuel cells

oxidize biofuels, either with enzymatic, microbial, or non-biological catalysts, such
as metal alloys and chelates, or with activated carbon. In the early literature, the
term “biofuel cell” is used for fuel cells operated with biofuels such as glucose,
whereas in more recent publications the term is used for biocatalyst-based fuel
cells. Direct biofuel cells directly electro-oxidize the primary fuel at the anode,
whereas in indirect fuel cells the fuel is first converted into a more reactive
intermediate, which is then electro-oxidized at the anode.
While the reactants inside of the previously presented CFC are clearly separated
from each other by the PEM, there are scenarios without such a division for biofuel
cells. Implantable fuel cells supplied by the surrounding body fluids (e.g., blood) for
example have to pick the right components for the anode and cathode out of
the same fluid by reactant selective electrodes. In [29] an implantable direct
glucose fuel cell (GFC) is presented, in which the PEM is replaced by a combina-
tion of a selectively catalyzing oxygen reduction cathode and a porous membrane
(Fig. 19.8).
Fig. 19.7 Chip-integrated micro fuel cell (CFC). (a) CFC after post-CMOS processing. (b)
Measured voltage and power density of a fuel cell cascade
This allows a setup with only one interface to the body fluids and enables
thin-film GFCs to be attached to the front of an implant (e.g., a pacemaker).
Such biofuel cells based on non-biological catalysts combine the advantages of
biological safety, long-term stability, and biocompatibility. Under physiological
conditions, the power output of these cells is around 2.5 mW/cm2, sufficient
to operate current pacemakers and many low-power MEMS implants.
Fig. 19.8 Setup and operational concept of the direct glucose fuel cell (GFC)
19.2.4.3 Rechargeable Micro Fuel Cells
In combination with energy harvesting devices, the recharging of fuel cells will
build a main focus in many laboratories. While the glucose cells are energy
harvesting devices by themselves, the CFCs have to be supplied with hydrogen
once the palladium reservoir is empty. As recharging can theoretically occur by
electrolysis, the combination with vibration or solar harvesting devices will enable
the CFCs to be used as alternatives to gold or super-capacitors. The discontinuously
delivered power produced by the energy harvesters due to the changeable avail-
ability of environmental power can be stored within the CFCs. The CFC’s capabi-
lity as a long-term energy buffer will therefore promote its application as an
uninterrupted power supply unit for autonomous microsystems. Examples of such
systems may be found within cold-chain control or autonomous medical implants.
19.3 Power Management
In contrast to supplying power by battery or wire, using an energy transducer needs

further consideration in order to achieve high efficiency. Besides rectification
efficiency, voltage converter efficiency, and storage efficiency, the harvesting
efficiency has also to be considered. There is a tremendous dependence on appro-
priate interfacing of the transducer, which is mainly achieved by input-impedance
matching or adaptive voltage converters. Here, impedance matching not only
includes the electrical load parameters but also considers the amplitude damping
by the electromechanical feedback. That means the transducer’s amplitude
decreases as the applied load increases [30–32].
As depicted in Fig. 19.9, a transducer is typically connected to a particular
low voltage rectifier (Sect. 19.3.2.2). Simply connecting a smoothing capacitor
to a rectifier is not sufficient. A load adaptation interface to the transducer
Zg Zin, eq
Cstg
Vg, oc tune unit
detector
μEH generator Rectifier Load matching Energy

(G) converter (LMC) storage
Fig. 19.9 Adaptive interface for optimal equivalent input impedance
becomes necessary, which can adapt its input impedance while converting the
actual output voltage of the transducer into a (mostly higher) voltage level for the
buffer capacitor or rechargeable battery (Sect. 19.3.3). Hence, the harvesting
efficiency remains high over a wide range of excitation conditions and storage
voltage levels [31, 33–36].
The design challenges are to create an interface that works at very low voltages,
needs little power, and sets the load point (its equivalent input impedance) so as to
obtain maximum output power. It should also convert the actual transducer output
voltage to the voltage level of an energy storage device, e.g., a capacitor or battery.
The main limitation and difficulties of designing such an interface arise from the
requirement of reducing the power consumption to a few microwatts while being
able to accept fluctuant supply voltages – often in the sub-1-V regime. This restricts
the applicable hardware components, e.g., processors, data converters, parameter
monitoring, and communication devices. These restrictions include also the possi-
ble size, limiting the number of off-chip components [30, 34].
19.3.1 Power Multiplexing
A further step towards a more reliable power supply based on energy harvesting
devices is power multiplexing, as illustrated in Fig. 19.10. That means several
generators of complementary physical principles deliver power in a redundant
manner to a central energy buffer. Thus, if one transducer is weakly excited other
harvesters may be able to deliver power in order to keep the average power input
above a minimum level [37, 38].
Similarly, power multiplexing is also reasonable at the output of the interface,
e.g., for appropriate voltage scaling issues. Instead of one energy buffer, several
buffers are present. Each buffer gets charged up to a certain voltage. Hence,
different supply requirements of application components are accomplished more
efficiently due to lower voltage conversion losses [36, 39, 40]. For example, sensors
are often driven at a 3.3 V level while a radio transceiver needs only 1.8 V. Sensors
need mostly small or moderate average currents compared to transceivers.
Fig. 19.10 Energy

harvesting and storage chain LMC
G1
High-V
sensors
buffer
with transducer interface 1
actuators
LMC
G2
2
controller
Low-V
buffer
LMC transceiver
G3
3
μEH Energy-aware power

Application
generators management unit
Fig. 19.11 RFID system comprising field generator and tag
Otherwise, if only one buffer is used, internal line regulators (low-dropout, LDO)
typically provide the required voltages and efficiency decreases [39–41].
19.3.2 RFID and Signal Rectification
Strictly speaking, powering devices using radio frequencies such as in identification

tags (RFID) cannot be regarded as energy harvesting, since the energy necessary for
the operation of the system is not extracted from freely available environmental
energy but from energy that was especially provided for this purpose. Anyhow, this
discipline should be mentioned here, as there are several analogies concerning the
circuit principles.
As described in [42], in passive RFID technology, a field generator creates an
alternating high-frequency magnetic or electromagnetic field (Fig. 19.11), out of
which battery-less independent sub-systems (tags) extract their energy. Such
a subsystem essentially consists of an antenna, i.e., a receiver resonant circuit,
and an RF frontend (energy conditioner). To convert the alternating voltage induced
in the antenna into a constant supply voltage, a rectifying circuit is needed.
Conventional CMOS integrated full-bridge rectifiers consist of four PMOS
transistors connected as diodes. Since a forward-biased PMOS diode does not

conduct any current until the threshold voltage is reached, the full-wave rectifier
reduces the available voltage by two threshold voltages. In harvesting applications
this leads to lower efficiency.
19.3.2.1 Floating Gate Diode
To minimize this voltage and the accompanying power loss, principles such as the
rectifier using floating gate diodes presented in [43] can be used.
By applying a voltage (slightly larger than the usual operating voltage) between
a MOSFET’s control terminal and the channel, electrons can be transferred through
the insulating oxide layer (Fowler–Nordheim tunneling, Fig. 19.12). If the circuit
node receiving the charge carriers is isolated from surrounding circuit elements,
i.e., is floating, a defined amount of charge can be deposited on it. This corresponds
to a static bias voltage, which effectively reduces the voltage required to turn on the
transistor. This technique needs an additional programming step after actual
production.
19.3.2.2 Negative Voltage Converter
Another way to reduce the voltage loss is the use of the negative voltage converter
followed by an active diode [10]. As illustrated in 19.13, the negative voltage
converter also comprises only four transistors, but in contrast to the conventional
full wave rectifier they are not connected as MOS diodes but as switches.
The switch creates a low-resistive connection between the alternating input
voltage and the storage capacitor. Since a switch, unlike a diode, does not prevent
a c
Fig. 19.12 Different types of

diodes used in rectifiers: (a)
PN diode; (b) PMOS diode;
(c) floating gate diode. (d)
Cross section of a single poly
floating gate transistor
Fig. 19.13 Operation

principle and transmission
characteristic of the negative
voltage converter with
subsequent active diode
a possible backflow of charge that was already accumulated on the storage capaci-
tor, a diode has to be attached to the negative voltage converter’s output. In order to
avoid reintroducing the typical voltage drop over the PMOS diode, one makes use
of an active diode. As shown in Fig. 19.13, a voltage comparator evaluates
the potentials at the input and output of the active diode and thus determines the
direction of a potential current flow. It controls the PMOS switch so that only
charge flowing onto the capacitor is permitted.
In order to achieve the optimum efficiency of the rectifier, the voltage compara-
tor needs to be designed for very low power consumption. As a consequence its
switching speed is limited and the rectifier is only suitable for systems with a
frequency of a few-hundred kilohertz. To avoid the need of an active diode circuit
and thus to make the rectifier suitable for higher frequency operation, [44] proposes
that only two of the four diodes used in the conventional full-wave rectifier be
replaced by switches. This topology requires a drop of only one threshold voltage
between its input and output, while inherently preventing a current back-flow.
19.3.3 Input-Impedance Adaptive Interfaces
Instead of impedance matching, the term “load adaptation” between the generator
and its electronics (Fig. 19.9) is sometimes preferred since typical conjugate
impedance matching as normally understood is actually not sufficient. Therefore,
the intention is to adjust the equivalent input impedance of the interface so as to
maintain an optimal transducer output voltage to approximate the maximum power
condition [33, 34, 36].
The following examples demonstrate this approach. One method continuously
adapts to the actual open circuit voltage of the electromagnetic transducer, while
two flying capacitor arrays are switched out of phase in order to best match the input
voltage (Fig. 19.14). By switching the arrays, voltage conversion is achieved
simultaneously [34]. Thus, non-harmonic and fluctuant transducer output voltages
are also covered. As introduced in Sect. 19.2.1, rather low-frequency excitations are
a
Vrect Vstg
<Zin>
Zg
ftoggle kV
rectifier
Vg, oc Vg
Cin Cstg
Electromagnetic Low-voltage
Adaptive interface
transducer rectifier
b
IP
Piezo VP VCC
Fig. 19.14 (a) An input-load matching adaptive charge pump for efficient interfacing of electro-
magnetic transducers. (b) A SSHI energy harvesting circuit for efficient harvesting with piezo-
electric transducers
converted into electrical power with highly inductive transducers. Thereby, the high
resistive transducer coils appear as a quasi-resistive source impedance [10, 30, 45].
In a first step, interface circuits for piezoelectric harvesters, as for all AC
generators, need to efficiently rectify the input voltage. Second, the voltage level,
which is often very high, has to be down-converted to usable levels. One method
keeps the piezoelectric harvester unloaded until the piezo voltage reaches a maxi-
mum, followed by a transfer phase extracting the total energy stored in the piezo-
electric harvester temporarily into a coil and finally transferring it into a buffer
capacitor. This principle is described as synchronous electric charge extraction
(SECE) [46]. When tested experimentally against a linear impedance-based con-
verter design, the SECE principle increases power transfer by over 400%. SECE
circuits implemented in CMOS technology have been presented in [47] and [48].
In similar studies [35, 49–51] another nonlinear technique, called synchronous
switch harvesting on inductor (SSHI), has been developed (Fig. 19.14). Unlike
SECE, the coil–switch combination is placed between the piezoelectric harvester
and the rectifier. A control circuit monitors the deflection of the piezoelectric
harvester and triggers the switch at a deflection maximum, flipping the piezo
voltage instantaneously. Thus, the rectifier is conducting most of the time, avoiding
phases where the piezo voltage is below the voltage level of the buffer capacitor.
Tests with discrete electronics show that the SSHI circuit is capable of extracting
400% more power than the standard rectifier circuit.
Since electrostatic transducers are capacitive, just like the piezoelectric ones,
similar methods can be used. The initial approach for implementing electrostatic
energy conversion was the implementation of a variable capacitor in a switched
system [20]. Through the operation of switches a reconfiguration of the variable
capacitor at different parts of the conversion cycle was induced, enabling the
process of energy harvesting. A new approach based on charge transportation
between two variable capacitors continuously connects the capacitors to the cir-
cuitry and therefore no switches are required [22]. To electrically bias the variable
capacitors, precharged capacitors or polarized dielectric materials (electrets) can be
used. Power outputs are predicted in the range of 1–40 mW based on theoretical
analysis.
All these examples use very minimalistic, power-saving control and drive
circuits. Compared to early harvesting and interfacing solutions, the tendency is
clearly towards processorless and rather analog solutions [30, 52]. Light-load
efficiency and high integration density in sub-milliwatt power ranges mark the
direction of development [37, 41].
19.3.4 Circuitry for Fuel Cells
Both solar and fuel cells deliver very low output voltages, which need to be boosted
for application using DC–DC converters. Furthermore, both the presented micro-
FCs – the chip integrated fuel cells (CFCs) and the direct glucose fuel cell (GFC) –
need to be monitored and controlled for efficient and safe operation. In [28]
a stabilized power supply based on the CFCs with monitoring-and-control circuitry
was realized within an extended CMOS process (Fig. 19.15). The electronic control
circuitry consists of a low-dropout voltage regulator (LDO) [53], an on-chip
oscillator, a programmable timing network, and a monitoring-and-bypass circuit
for each CFC. All of these components are driven by the integrated CFCs. For the
benefit of the system’s lifetime, the core system is restricted to fundamental control
elements such as an oscillator and an asynchronous timing network to avoid current
peaks. It is powered by a redundant two-CFC-stack generating 1–1.5 V, depending
on the charging level of the fuel cells.
For high reliability of the proposed power supply system all fuel cells within
each cascade have to be checked for functionality before they are used to drive the
LDO. Each cell’s anode is compared to its slightly loaded cathode. In the case of
defective or empty fuel cells, a small load current results in a significant voltage
drop across the cell, enabling a dynamic comparator to classify this cell as “unus-
able”. These defective or empty fuel cells are shorted by CMOS transmission gates
V7 LDO
Bypass
Bypass –
+
Bypass
Bypass
6
Bypass
Bypass Vout
Bypass
Bypass
LDO active
FC Ready
Bypass
Bypass
FC Rst
Vcore 6
Bypass
Bypass
fosc
Osc Timer
Bypass
Bypass
Vcore
Fig. 19.15 Schematic of the CFC monitoring-and-control circuitry for a stabilized power supply
with chip-integrated micro-FCs
to prevent loss of a complete cascade or to keep its input resistance as low as

possible.
By stacking the fuel cells to cascades with an output voltage higher than 3.3 V
for load currents up to 7 mA, a classical LDO was used to stabilize the output voltage
to 3.3 V. As presented, six fuel cell cascades are integrated on chip. To increase the
system’s driving capability, these cascades are connected in parallel to the output
by separate PMOS transistors of the LDO, used as pass elements, to avoid
equalizing currents between the different cascades. Thus, the required load current
is distributed to the cascades with respect to their driving capability. The complete
circuitry is implemented without external components.
19.4 Ultra-Low-Voltage Circuit Techniques
The harvesting applications being focused on are sensor and data-acquisition

devices and micro-actuators with wireless communication. Considering the previ-
ously discussed systems and architectures, the necessary building blocks are power
interfaces for the transducers, voltage converters, and power management control
units. All these modules include amplifiers, comparators, references, clock
generators, and digital blocks. Owing to the limited energy resources, these circuits
have to be designed compatible to the low voltages and power available rather than
for performance. Feasibility, functionality, and reliability are seen as more critical
and essential than absolute accuracy, bandwidth, and speed.
To handle the high voltages that might occur at the output of harvesting
transducers and owing to the voltage requirements for driving certain sensors
commonly used, bulk CMOS processes for energy harvesting processing modules
remain rather constant at 0.18–0.35 mm technologies [54–58]. These processes,
which usually include high voltage options, have relatively high threshold voltages,
making the design of ultra-low-voltage circuits a challenging task.
19.4.1 Analog Design and Scaling Considerations
To enable analog integrated circuits to operate at low supply voltages, different

strategies are available: (1) weak/moderate inversion operation, (2) bulk-driven
stages [34, 59–61], (3) floating gate programming [62, 63], (4) input-level shifting,
(5) field-programmable analog arrays (FPAAs), such as reconfigurable unity-gain
cells [59, 61], and (6) digitally assisted architectures and sub-threshold designs [35,
64, 65].
Besides low supply voltages, the scaling of bulk CMOS technologies also has to
be considered. As down-scaling reduces threshold voltages and minimum supply
voltages [66], several degenerative effects arise. Most challenging are mismatch
and reduced output resistance (gain), as well as lower dynamic range and signal-to-
noise ratio [59, 67]. Often, speed has to be traded for noise margins. As an
additional scaling result, the body effect will decrease, which reduces the potential
of bulk-driven stages.
In conjunction with increased mismatch, unity-gain amplifiers with the option of
reconfiguration become an interesting way to deal with these obstacles [68, 69].
Further, the technique of using digitally assisted architectures [65] helps to save
power, while being useful for mismatch correcting and improving linearity.
19.4.1.1 Amplifiers and Comparators
For high amplification and current efficient operation the strong inversion regime
might be omitted. The use of particular input-stages, e.g., bulk-driven MOS
transistors or specially biased gate-driven stages, allows wide input common
mode ranges while linearity remains high [59–61]. Fig. 19.16 gives further circuit
examples.
Controlling the channel of MOS transistors by both the gate and the bulk
terminals allows (exchangeable) use of either terminal as an input or for biasing.
Fig. 19.16 Input-stage 0.5V

with two stacked transistors
only [61] Vin+ Vin–
+ – +
+ – Vout–
– + Vout+
Avg.
Table 19.1 Low-voltage amplifiers with different design techniques: gate-driven (GD), bulk-
driven (BD), multi-path (MP), multi-stage (MS), switched-capacitor (SC)
GD[59] BD[59] BD[61] MP[70] MS[71] SC[72]
VDD [V] 0.5 0.5 0.5 2.0 1.0 1.0
PV[mW] 75 100 1.7 400 10,000 1.8
GBW [MHz] 10 2.4 0.1 4.5 200 0.07
DC gain [dB]
pffiffiffiffiffiffi 62 52 93 100 84 90
Noise ½nV/ Hz 220 280 500 – 27 60
FOM [(MHz pF)/mW] 2.7 0.48 1.18 1.35 0.02 1.16
Load CL [pF] 20 20 20 120 1 pF, 25kO 30
Technology [mm] 0.18 0.18 0.35 0.8 0.065 0.065
These bulk-driven stages make wide input ranges possible while biasing at the gate
terminals guarantees high linearity [59]. However, low body transconductance
adversely affects gain and noise margins, and an increased input capacitance limits
the bandwidth [59, 61]. Prevention of latch-up is necessary and nMOS devices
require triple-well processes [61].
The concept of switched biasing allows strong inversion operation [60]. This,
however, requires a clock and continuous switching of the level-shifting input
capacitors.
Floating gate programmed amplifiers are an interesting opportunity, as described
in the rectifier section (Sect. 19.3.2). The threshold voltage and transconductance
can be adjusted granularly, and device mismatch can be compensated [62]. But this
technique is a research topic rather than an industrially applicable one, since each
single device has to be carefully programmed.
Another emerging technique is using unity-gain cells and implementing them
with reconfigurable options. Thus, mismatches could also be compensated and
certain bandwidth and gain characteristics could be configured. This technique
may especially help to compete with the challenging characteristic of short-channel
transistors at process technologies with digital emphasis [62, 63]. For increasing
accuracy or linearity, digital correction mechanisms are recommended [65, 73].
(Table 19.1)
For comparators, in addition to very efficient designs with regenerative or

clocked latches, the use of asynchronous comparators is a reasonable choice
[34, 65]. That is justified by the shallow clock edges and low slew rates if operation
in strong inversion is not possible [35, 64], or if clocking is totally avoided, as
in asynchronous digital circuits, as explained later [74]. In conjunction with
active diodes, common-gate input stages allow a wide range of input and supply
voltage levels. Even at low biasing currents adequate response and transition delay
is maintained.
19.4.2 Ultra-Low-Voltage Digital CMOS Circuits
The requirements for digital blocks used in interface circuits for energy harvest-
ing devices are typically very strict: Energy budgets are tight and the minimum
supply voltage should be reduced as far as possible to allow for an early start-up of
the device. Even though circuits operating at ultra-low supply voltage are typically
also highly energy efficient, optimization for minimum energy consumption and
minimum supply voltage are not equivalent targets.
In this section, we focus our discussion on supply voltage reduction for mini-
mum power in digital circuits rather than energy-per-operation minimization as
discussed, e.g., in [75], as the ultimate limit for energy harvesting devices typically
is the startup supply voltage required by the operated electronics.
An early discussion about a theoretical lower supply voltage limit [76] states
a minimum of kT/q based on noise considerations, more specifically for CMOS
circuits a limit of 2(ln 2)kT/q for inverters [77] and 2(ln 5)kT/q for practical circuits
[78], derived from the requirement that a logic gate needs more than unity gain.
Regarding practical circuit implementations, a very low supply voltage CMOS
design was presented as early as the 1990s operating down to 125 mV [79] at room
temperature. The operation occurs in saturation with transistor threshold voltages
tuned to near-zero values though. As the convenience of such optimized process
technology is not available for most applications, most ultra-low-voltage circuits
operate the transistors in weak inversion. Examples are a hearing aid filter operating
at 300 mV [80], a fast Fourier transform (FFT) processor operational at 180 mV
[81], and a microcontroller optimized for the sub-threshold regime working at
supplies as low as 168 mV [82]. Also memories with supply voltages down to
160 mV have been demonstrated [83]. These designs are optimized mainly for
minimum energy consumption by supply voltage reduction. It is possible to achieve
considerably lower supply voltages when optimizing specifically for this target
though, as reported for a memory operational down to 135 mV [84], a FIR filter
operational to 85 mV [85], and multiplier circuits remaining fully functional down
to 62 mV [86].
Figure 19.17a gives an overview of a selection of publications regarding ultra-
low-voltage digital circuits, relating the technology node used to the minimum
supply voltage. It becomes clear that there is an obvious improvement in minimum
a 1000
750
Vmin [mV] trend line
500
Logic
Memory
250 Logic+Memory
Ring oscillator
0
50 500 5000
Lmin [nm]
b 500
400
Vmin [mV]
300
200 Logic
Memory
100 Logic+Memory
Ring oscillator
0
65 90 130 180 250 350
Lmin [nm]
Fig. 19.17 Minimum supply voltage of circuits versus technology node: (a) a long-term interpo-
lation, and (b) a zoomed view of recent technologies
supply coming from old technologies, but the supply voltage values level off or
even increase for current ones. To understand this trend, a closer analysis of the
reasons that limit the scaling of the supply voltage is necessary.
The ultimate limit for supply voltage reduction is imposed by a degradation of
static noise margins (SNMs) to zero. More illustratively, the output levels of CMOS
gates degrade due to the decrease in the “on” to “off” current ratio with decreasing
supply voltage until the output level of some digital gate is falsely evaluated by the
succeeding logic. Three main factors impact output level degradation: The circuit
structure, process technology factors and process variabilities.
Regarding circuit structure, gate sizing has an important impact on minimum
supply, as discussed in [87] and [81], requiring symmetric devices for minimum
supply [88]. Furthermore, numerous circuit structures should be avoided, i.e., long
transistor stacks (leading to an inherent worst-case asymmetry in CMOS), ratioed
circuits [89] and dynamic circuits. All of these requirements though do not neces-
sarily optimize energy consumption, as discussed for sizing in [87] and are there-
fore applied within limits in most circuits.
The major process technology factors affecting SNM are the sub-threshold slope
factor and drain-induced barrier lowering (DIBL). DIBL corrupts SNM as it causes
additional mismatch between “on” and “off” devices [90]. The sub-threshold slope
defines the “on” to “off” current ratio of a transistor and therefore directly impacts
SNM. It improves with shrinking oxide thickness, but then also degrades with
decreasing gate length owing to short-channel effects. The fact that oxide thickness
is scaled less than minimum gate length to limit gate leakage results in an overall
sub-threshold slope factor deterioration in modern technologies [91].
Process variability impairs SNM as the imbalance of devices is increased,
causing critical output level deviations at higher supply voltages. Threshold voltage
variations are the major source of variability in sub-threshold circuits owing to their
exponential impact. Many designs compensate global variabilities by the use of
adaptive body biasing [82, 85, 92, 93], whereas local variability is dominated by
random dopant fluctuations [94] and can therefore only be compensated by increas-
ing device sizes following Pelgrom et al.’s model [95]. This approach is primarily
used in designs optimized for ultra-low-voltage operation though (e.g., [86]) as it
also increases the energy consumed per operation.
The increase in minimum supply voltage with shrinking technology in
Fig. 19.17b is therefore easily explained by the fact that most of these circuits
have a stronger optimization for minimum energy than for minimum supply and
therefore suffer from increased variability, DIBL and sub-threshold slope factor. To
profit from technology scaling, it is however possible to use larger than minimum-
length devices [90, 91] to alleviate the DIBL and variability impacts and effectively
profit from increased channel control due to the reduced oxide thickness.
If possible improvements in minimum supply voltage due to technology
improvements are to be estimated, it needs to be noted that the only major process
parameter relevant to minimum supply that can improve with scaling is sub-
threshold slope. This factor is limited to approximately 60 mV/decade for physical
reasons and has a value of 70–80 mV/decade in current processes. Using a first-
order linear approximation, an ideal process might therefore improve supply volt-
age in a circuit highly optimized for minimum supply voltage operation from 62
[86] to 53 mV and in a design optimized for minimum energy per operation from
168 [82] to 144 mV. It is therefore obvious that for supply voltage reduction, circuit
design innovations are more relevant than technology scaling. To achieve major
technology-driven improvements for minimum supply voltage reduction, new
device technologies are necessary, e.g., multi-gate FETs, improving channel con-
trol and removing random dopant fluctuations; band-to-band tunneling transistors
[96] or carbon nanotubes [97], allowing considerably reduced sub-threshold slopes;
or even devices such as the BiSFET [98], for which nominal supply voltages in the
range of 50 mV are predicted.
19.5 Conclusion
The latest research on high-performance energy transducers and adaptive circuits has
been presented. The challenge remains to close the gap between the energy genera-
tion capabilities of the harvesters and the energy consumption of the adaptive
electronics used to optimize the output power of the harvesting devices. Whereas
previous research focused mainly on characterizing the harvesting devices, in the
future complete energy harvesting systems, including harvester, interface circuits,
storage, and application electronics, need to be developed and optimized as a system.
References
1. Delagi, G.: International Solid-State Circuits Conference Digest of Technical Papers,

pp. 18–24. IEEE (2010). Reprinted as chap. 13 in the present book
2. EnOcean. http://www.enocean.com/
3. Perpetuum. http://www.perpetuum.com/
4. Arnold, D.: Trans. Magn. 43(11), 3940–3951 (2007)
5. Spreemann, D., Willmann, A., Folkmer, B., Manoli, Y.: In: Proceedings of PowerMEMS,
pp. 261–264 (2008)
6. Trimmer, W.: IEEE J. Solid-State Circuits 19, 267 (1989)
7. Hoffmann, D., Folkmer, B., Manoli, Y.: J. Micromech. Microeng. 19(094001), 2939,
11 pp (2009)
8. Ferrari, M., Ferrari, C., Guizzetti, M., And, B., Baglio, S., Trigona, C.: Proc. Chem 1, 1203
(2009)
9. Eichhorn, C., Goldschmidtboeing, F., Woias, P.: J. Micromech. Microeng. 19(094006) (2009)
10. Peters, C., Spreemann, D., Ortmanns, M., Manoli, Y.: J. Micromech. Microeng. 18(10
104005), 1203 (2008)
11. Ferrosi. http://www.ferrosi.com/
12. KFC-Tech. http://www.kfctech.com/
13. Williams, C., Yates, R.: In: International Conference on Solid_State Sensors and Actuators,
and Eurosensors IX, pp. 369–372 (1995)
14. Spreemann, D., Folkmer, B., Manoli, Y.: In: Proceedings of Power_MEMS (2008),
pp. 257–260
15. Lee, C., Joo, J., Han, S., Lee, J.H., Koh, S.K.: Appl. Phys. Lett. 85, 1841 (2004)
16. Lee, C., Joo, J., Han, S., Lee, J.H., Koh, S.K.: In: Proceedings of International Conference on
Science and Technology of Synthetic Metal, pp. 49–52 (2005)
17. Ng, T.H., Liao, W.H.: J. Intell. Mater. Syst. Struct. 16, 785797 (2005)
18. Breaux, O.P.: Electrostatic energy conversion system. 1976 U.S. Patent 04127804, 28 Nov
1978
19. Basset, P., Galayko, D., Paracha, A., Marty, F., Dudka, A., Bourouina, T.: J. Micromech.
Microeng. 19(115025), 12 pp (2009)
20. Meninger, S., Mur-Miranda, J., Amirtharajah, R., Chandrakasan, A., Lang, J.: IEEE Trans.
VLSI Syst. 9(1), 64 (2001)
21. Naruse, Y., Matsubara, N., Mabuchi, K., Izumi, M., Suzuki, S.: J. Micromech. Microeng. 19
(094002), 5 pp (2009)
22. Sterken, T., Baert, K., Puers, R., Borghs, G., Mertens, R.: In: Proceedings of Pan Pacific
Microelectronics Symposium, pp. 27–34 (2001)
23. Williams, C.B., Yates, R.B.: In: Proceedings of Transducers/Eurosensors, pp. 369–372 (1995)
24. Bttner, H., et al.: IEEE J. Microelectromech. Syst. 13(3), 414 (2004)
25. Micropelt. http://www.micropelt.com

26. Hopf, M., Schneider, K.: Fraunhofer ISE – Annual Report 2008. Fraunhofer ISE, Freiburg
(2009)
27. Erdler, G., Frank, M., Lehmann, M., Reinecke, H., Mueller, C.: Sens. Actuators A 132(1),
331336 (2006)
28. Frank, M., Kuhl, M., Erdler, G., Freund, I., Manoli, Y., Mueller, C., Reinecke, H.: IEEE J.
Solid State Circuits 45, 205 (2010)
29. Kerzenmacher, S., Ducre, J., Zengerle, R., von Stetten, F.: J. Power Sources 182, 1 (2008)
30. Mitcheson, P., et al.: Proc. IEEE 96, 1457 (2008)
31. Mitcheson, P.D., Green, T.C., Yeatman, E.M.: Proc. Microsyst. Technol. 13(11–12), 1629
(2007)
32. Spreemann, D., Willmann, A., Manoli, B.F.Y.: In: Proceedings of PowerMEMS, pp. 372–375
(2009)
33. Lefeuvre, E., Badel, A., Richard, C., Petit, L., Guyomar, D.: J. Sens. Actuators, A Phys. 126,
405416 (2006)
34. Maurath, D., Manoli, Y.: In: Proceedings of IEEE European Solid State Circuits Conference
(ESSCIRC), pp. 284–287. IEEE (2009)
35. Ramadass, Y.K., Chandrakasan, A.: In: IEEE International Solid-State Circuits Conference
Digest of Technical Papers, pp. 296–297. IEEE (2009)
36. Yi, J., et al.: In: Proceedings of IEEE International Symposium on Circuits and Systems
(ISCAS), pp. 2570–2573 (2008)
37. Lee, S., Lee, H., Kiani, M., Jow, U., Ghovanloo, M.: In: IEEE International Solid-State
Circuits Conference Digest of Technical Papers, pp. 12–14 (2010)
38. Li, Y., Yu, H., Su, B., Shang, Y.: IEEE Sens. J., 678–681 (2008)
39. Flatscher, M., Dielacher, M., Herndl, T., Lentsch, T., Matischek, R., Prainsack, J., Pribyl, W.,
Theuss, H., Weber, W.: In: IEEE International Solid-State Circuits Conference Digest of
Technical Papers, pp. 286–28 (2009)
40. Guilar, N., Amirtharajah, R., Hurst, P., Lewis, S.: In: IEEE International Solid-State Circuits
Conference Digest of Technical Papers, pp. 298–300 (2009)
41. Pop, V., et al.: In: Proceedings of PowerMEMS, pp. 141–144 (2008)
42. Finkenzeller, K.: RFID Handbook. Wiley, England (2003)
43. Peters, C., Henrici, F., Ortmanns, M., Manoli, Y.: In: IEEE International Symposium on
Circuits and Systems (ISCAS), pp. 2598–2601 (2008)
44. Ghovanloo, M., Najafi, K.: IEEE J. Solid State Circuits 39, 1976 (2004)
45. Kulah, H., Najafi, K.: In: Seventh IEEE International Conference on MEMS, pp. 237–240.
IEEE (2004)
46. Lefeuvre, E., Badel, A., Richard, C., Guyomar, D.: J. Intell. Mater. Syst. Struct. 16, 865876
(2005)
47. Hehn, T., Manoli, Y.: In: Proceedings of PowerMEMS, pp. 431–434 (2009)
48. Xu, S., Ngo, K.D.T., Nishida, T., Chung, G.B., Sharma, A.: IEEE Trans. Power Electr. 22, 63
(2007)
49. Badel, D., Guyomar, E., Lefeuvre, C., Richard, J.: Intell. Mater. Syst. Struct. 16, 889901
(2005)
50. Guyomar, D., Badel, A., Lefeuvre, E., Richard, C.: IEEE Trans. Ultrason. Ferroelectr. Freq.
Control. 52(4), 584594 (2005)
51. Lefeuvre, E., Badel, A., Richard, C., Guyomar, D.: In: Proceedings of Smart Structures and
Materials Conference (2004), pp. 379–387
52. Ottman, G., Hofmann, H., Lesieutre, G.: IEEE Trans. Power Electron. 18, 696 (2003)
53. Rincon-Mora, G., Allen, P.: IEEE J. Solid State Circuits 33, 36 (1998)
54. Amirtharajah, R., Chandrakasan, A.: IEEE J. Solid-State Circuits 33, 687 (1998)
55. Chao, L., Tsui, C.Y., Ki, W.H.: In: IEEE/ASM International Symposium on Low Power
Electronics and Devices (ISLPED), pp. 316–319. IEEE (2007)
56. Man, T., Mok, P., Chan, M.: IEEE J. of Solid-State Circuits 44(9), 2036 (2008)
57. Mandal, S., Turicchia, L., Sarpeshkar, R.: IEEE Pervasive Comput. 9(1), 71 (2010)
58. Torres, E., Rincon-Mora, G.: IEEE Trans. Circ. Syst. I (TCAS I) 56(9), 1938 (2009)
59. Chatterjee, S., Tsividis, Y., Kinget, P.: IEEE J. Solid State Circuits 40(12), 2373 (2005)
60. Chiu, Y., Gray, P., Nikolic, B.: IEEE J. Solid-State Circuits 39, 21392151 (2004)
61. Maurath, D., Michel, F., Ortmanns, M., Manoli, Y.: In: Proceedings of IEEE MWSCAS (2008)
62. Henrici, F., Becker, J., Trendelenburg, S., DeDorigo, D., Ortmanns, M., Manoli, Y.: In:
Proceedings of IEEE International Symposium on Circuits and Systems, pp. 265–268. IEEE
(2009)
63. Naess, O., Berg, Y.: In: IEEE International Symposium on Circuits and Systems (ISCAS), pp.
663–666. IEEE (2002)
64. Lotze, N., Ortmanns, M., Manoli, Y.: In: Proceedings of IEEE International Symposium on
Low-Power Electronic Devices, pp. 221–224 (2008)
65. Murmann, B.: In: Proceedings of IEEE Custom Integrated Circuits Conference (CICC), pp.
105–112 (2008)
66. ITRS. The international technology roadmap for semiconductors. http://public.itrs.net (2009)
67. Vittoz, E., Fellrath, J.: IEEE J. Solid State Circuits SC-12, 224 (1977)
68. Becker, J., Henrici, F., Trendelenburg, S., Ortmanns, M., Manoli, Y.: IEEE Trans. Solid-State
Circuits 43(12), 2759 (2008)
69. Knig, A.: In: International Conference on Instrumentation, Communication and Information
Technology (ICICI) (2005)
70. Lee, H., Mok, P.: IEEE J. Solid-State Circuits 38(3), 511 (2003)
71. Sancarlo, I.D., Giotta, D., Baschirotto, A., Gaggl, R.: In: ESSCIRC 2008 – 34th European
Solid State Circuits Conference, pp. 314–317. IEEE (2008)
72. Fan, Q., Sebastiano, F., Huijsing, H., Makinwa, K.: In: Proceedings of IEEE European Solid
State Circuits Conference (ESSCIRC), pp. 170–173. IEEE (2009)
73. Witte, P., Ortmanns, M.: IEEE Trans. Cir. Syst. I: Reg. Papers 57(7), 1500 (2010)
74. Lotze, N., Ortmanns, M., Manoli, Y.: In: Proceedings of International Conference on Com-
puter Design (ICCD), pp. 533–540 (2007)
75. Calhoun, B., Wang, A., Chandrakasan, A.: IEEE J. Solid-State Circuits 40(9), 1778 (2005)
76. Keyes, R., Watson, T.: In: Proceedings of IRE (Correspondence), pp. 24–85 (1962)
77. Swanson, R., Meindl, J.: IEEE J. Solid-State Circuits 7(2), 146 (1972)
78. Schrom, G., Pichler, C., Simlinger, T., Selberherr, S.: Solid-State Electron. 39, 425 (1996)
79. Burr, J.: In: Proceedings of IEEE Symposium on Low Power Electronics, pp. 82–83 (1995)
80. Kim, C.I., Soeleman, H., Roy, K.: IEEE Trans. Very Large Scale Integr. (VLSI) Syst. 11(6),
1058 (2003)
81. Wang, A., Chandrakasan, A.: IEEE J. Solid-State Circuits 40(1), 310 (2005)
82. Hanson, S., Zhai, B., Seok, M., Cline, B., Zhou, K., Singhal, M., Minuth, M., Olson, J.,
Nazhandali, L., Austin, T., Sylvester, D., Blaauw, D.: In: Proceedings of IEEE Symposium on
VLSI Circuits, pp. 152–153 (2007)
83. Chang, I.J., Kim, J., Park, S., Roy, K.: IEEE J. Solid-State Circuits 44(2), 650 (2009)
84. Hwang, M.E., Roy, K.: In: CICC 2008 – Custom Integrated Circuits Conference, 2008, pp.
419–422. IEEE (2008)
85. Hwang, M.E., Raychowdhury, A., Kim, K., Roy, K.: In: IEEE Symposium on VLSI Circuits,
154–155. doi:10.1109/VLSIC.2007.4342695 (2007)
86. Lotze, N., Manoli, Y.: In: ISSCC 2011 – IEEE International Solid-State Circuits Conference
Digest of Technical Papers (to be published) (2011)
87. Kwong, J., Chandrakasan, A.: In: Proceedings of International Symposium on Low Power
Electronics and Design, pp. 8–13 (2006)
88. Calhoun, B., Chandrakasan, A.: IEEE J. Solid-State Circuits 39(9), 1504 (2004)
89. Calhoun, B.H., Chandrakasan, A.: IEEE J. Solid-State Circuits 41(1), 238 (2006)
90. Bol, D., Ambroise, R., Flandre, D., Legat, J.D.: IEEE Trans. Very Large Scale Integr. Syst. 17
(10), 1508 (2009)
91. Hanson, S., Seok, M., Sylvester, D., Blaauw, D.: IEEE Trans. Electron Dev. 55(1), 175 (2008)
92. Jayapal, S., Manoli, Y.: In: Proceedings of 13th IEEE International Conference on Electronics,
Circuits and Systems ICECS’06, pp. 1300–1303 (2006)
93. Jayapal, S., Manoli, Y.: In: Proceedings of International Symposium on VLSI Design, Auto-
mation and Test VLSI-DAT 2007, pp. 1–4 (2007)
94. Roy, G., Brown, A., Adamu-Lema, F., Roy, S., Asenov, A.: IEEE Trans. Electron Dev. 53(12),
3063 (2006)
95. Pelgrom, M., Duinmaijer, A., Welbers, A.: IEEE J. Solid-State Circuits 24(5), 1433 (1989)
96. Koswatta, S., Lundstrom, M., Nikonov, D.: IEEE Trans. Electron Dev. 56(3), 456 (2009)
97. Appenzeller, J., Lin, Y.M., Knoch, J., Avouris, P.: Phys. Rev. Lett. 93(19), 196805 (2004)
98. Banerjee, S., Register, L., Tutuc, E., Reddy, D., MacDonald, A.: IEEE Electron. Device Lett.
30(2), 158 (2009)
Chapter 20
The Energy Crisis
Bernd Hoefflinger
Abstract Chip-based electronics in 2010 consumed about 10% of the world’s total
electric power of 2 TW. We have seen throughout the book that all segments,
processing, memory and communication, are expected to increase their performance
or bandwidth by three orders of magnitude in the decade until 2020. If this progress
would be realized, the world semiconductor revenue could grow by 50–100%, and the
ICT industry by 43–66% in this decade (Fig. 6.1). Progress sustained at these levels
certainly depends on investments and qualified manpower, but energy has become
another roadblock almost overnight. In this chapter, we touch upon the life-cycle
energy of chips by assessing the energy of Si wafer manufacturing, needed to bring
the chips to life, and the power efficiencies in their respective operations. An out-
standing segment of power-hungry chip operations is that of operating data centers,
often called server farms. Their total operating power was 36 GW in 2010, and we
look at their evolution under the prospect of a 1,000 growth in performance by
2020. One feasible scenario is that we succeed in improving the power efficiency of
• Processing 1,000,
• Memory 1,000,
• Communication 100,
within a decade.
In this case, the total required power for the world’s data centers would still
increase 4 to 144 GW by 2020, equivalent to 40% of the total electrical power
available in all of Europe.
The power prospects for mobile/wireless as well as long-line cable/radio/satellite
are equally serious. Any progression by less than the factors listed above will lead to
economic growth smaller than the projections given above. This demands clearly
that sustainable nanoelectronics must be minimum-energy (femtojoule) electronics.
B. Hoefflinger (*)

422 B. Hoefflinger
20.1 Life-Cycle Energy
The tiny micro- and nano-chips are so ubiquitous today that they have become an
issue on our overall environmental balance sheet. As a result, the science of a life-
cycle analysis (LCA) of the chips from their conception to their disposal has been
established. We focus here on two major aspects of their life cycle:
– The manufacturing of chips,
– The energy of data centers, the most energy-hungry operation of chips.
We simplify these highly complex topics by studying selected characteristics as
they appear in 2010, and we project them to 2020 in two scenarios, a pessimistic
one and an optimistic one.
The characteristics of manufacturing plants (fabs) are listed in Table 20.1. The
projections in the table take into account that 3D chip-stacking will be introduced
on a large scale, in order to achieve the expected advancements of performance
per chip-footprint. We assume that the weighted average of stacked chip- or
wafer-layers will be two layers. That is reflected both in the number of wafers
and in the energy per wafer-stack and in the production energy per chip-stack. The
required electrical power for the global chip production rises significantly to
levels equivalent to 1/2–2/3 of the total electrical power used in Italy. The
inflation in the total area of crystalline Si is a concern, too. However, thin wafer
films have to be used anyway, which are the natural results of ELTRAN
(Fig. 3.19) or CHIPFILM (Fig. 3.40), putting the usage of crystalline Si on a
new, more favorable metric.
In any event, the energy is a major cost factor in the production of chips. At
the price of 0.05 $/kWh, the cost breakdown in 2010 for a finished 30 cm wafer
would be:
Energy/wafer: $5,000
Plant/equipment $1,700
Labor $300
Total cost/wafer $7,000
Table 20.1 Si wafer manufacturing 2010 and 2020

2010 2020 conserv. 2020 optim. Source
Revenue [$billion] 300 450 600 Fig. 5.1
Chip units [109] 150 225 300
ASP [$] 2.00 2.00 2.00
Wafers [30 cm diam. equiv., million] 60 180 240
Total Si area [km2] 4 12 16
Plant and equipment [$billion] 80 120 160
Elect. power [GW] 6 18 24
Energy/wafer [MWh] 1 2 2
Energy/chip/cm2 [kWh] 1.7 3.4 3.4
Cost/chip/cm2 [$] 10 18 18
20 The Energy Crisis 423
Critical as this breakdown may be, the future trend of cost/wafer is unfavorable
due to the rising cost of smaller feature sizes. We have seen in Chap. 8 that the cost
of nanolithography makes up an increasing part not only of the total equipment cost
but also of the total energy cost. The rule also holds that the smaller the minimum
dimensions on the product, the larger the overhead on plant volume and plant
power for environmental control (temperature, air, media, vibration, etc.). It is
clear that this scenario is yet another inflection point for scaling down much beyond
16 nm (the 2015 node).
However, even with these limitations on the manufacturing roadmap, it is
possible to foresee that, from 2010 to 2020, the much quoted and demanded
1,000-fold improvement for high-performance processors is possible.
20.2 Operating Performance and Energy
Here, we focus on the data centers (or server farms) because, not only do they
operate with high-performance chips, hot chips [1], but also, as the digital brains of
the global knowledge world and as the backbones of the internet, they are respon-
sible for the largest part of the energy bill of the information and communication
technology (ICT) industry. In 2010, worldwide 36 million servers operated and
consumed ~36 GW (roughly equivalent to the total electrical power of all of Italy).
This power level is also six times the power consumption of all chip-manufacturing
plants. And it is expected that the world’s data centers improve their performance
1,000-fold in the present decade from 2010 to 2020.
We will try to analyze this 10-year perspective, starting from the present relation
that processors, memory and interconnects make up about 1/3 each of the total
energy bill of the data centers. We now look at these three entities and their
potential based on our analyses in previous chapters. We start with processor
performance in Table 20.2.
The 1,000-fold improvement in operations/s appears in the roadmap for
supercomputers, Fig.6.4, and we have taken that as representative for the
expectations on servers. In the two scenarios for 2020, the improvement of the
power-efficiency by a factor of 100 is assumed in the mainstream of the industry,
and it would increase the necessary electrical power by an order of magnitude in
this decade. However, we have seen in Sect. 3.6 in the examples on digital
multipliers that improvements in power-efficiency by more than an order-of-
magnitude are possible with respect to the mainstream (Fig. 3.32). There are
Table 20.2 The advancement of processor performance and power consumption

Processors 2010 2020 pessim. 2020 optim. Source
Rel. MOP/s 1 1,000 1,000 Fig. 6.4
Rel. power-efficiency 1 0.01 0.001 Fig. 3.32
Rel. power 1 10 1
424 B. Hoefflinger
certainly many bottlenecks in processors (memory will be treated separately in the

following) other than multipliers. However, it is also a strong point in Chap. 13 that
the trend line of the significantly more power-efficient special processors for
wireless applications (Fig. 13.8) is a sign for improvement by another factor of
10. Therefore, in the optimum scenario, a 1,000 improvement in operations/s is
feasible at constant energy. We will find that this amount of progress is harder to
predict for the other two constituencies of computers, namely memory and
interconnects.
What is needed with processors is high-speed instruction and data memory, so that
we now have to look at the progress of SRAM and DRAM with regard to their power-
efficiency. We find in Fig. 11.6 that SRAM dynamic power-efficiency can advance to
as low as 10 nW/Gb/s, while DRAM has reached its limits at 10 mW/Gb/s, but could
gain another factor of 10 in a “differential DRAM”, offering close to 1 mW/Gb/s. On the
pessimistic side, we have to settle for a 100 advancement of memory power-
efficiency, and an optimistic mix of SRAM and differential DRAM may achieve
1,000 (Table 20.3).
The third critical component is interconnects. In Chap. 5, we came to
the conclusion that the power-efficiency will advance at best 100 in the decade
2010–2020 to 10 mW/(Gb/s) so that we obtain Table 20.4.
The development of the total electrical power of data centers is shown in
Fig. 20.1. The pessimistic forecast would require 10 more power for the data
centers, i.e., 360 GW needed in 2020 for data centers, equivalent to the present total
electrical power of all of Europe. Since this is not realistic, the performance
improvement of data centers would slow down to 100/decade, resulting in a
serious setback for the expansion of computer and internet services. Even in the
optimum scenario, the electrical power of data centers would increase 4 to 144 GW,
which is hard to imagine from an investment and environment point-of-view. It was
predicted in 2008 [2] that energy costs could eat up to 40% of IT budgets. Data
centers are at the forefront of this escalation, and major developments are under
Table 20.3 The advancement of memory performance and power consumption

Memory 2010 2020 pessim. 2020 optim. Source
Rel. Gb/s 1 1,000 1,000
Rel. power-efficiency 1 0.01 0.001 Fig. 11.6
Rel. power 1 10 1
Table 20.4 The Interconnects 2010 2020 Source

advancement of interconnect
performance and power Rel. Gb/s 1 1,000
consumption Rel. power-efficiency 1 0.01 Fig. 5.2
Rel. power 1 10
360 GW
144 GW
120 Interconnect
36 GW
Interconnect 12
Memory 12 12 Memory
Processors 12 12 Processors
Worst Best
Case Case
2010 2020 2020
Year
Fig. 20.1 The electrical power consumption of data centers in 2010 and 2020
way to alleviate this situation. Already, more than half of the energy bill of data
centers is spent on cooling, so that recovering a major part of this by efficient water-
cooling of the servers is an ongoing development. If the water is used for remote
heating or industrial thermal processing, half of the original electrical energy could
be recovered for other applications [3]. Sustainable growth will depend more and
more on a serious commitment to ultra-low-energy architectures and circuits as
well as integral optimization of the energy issue. Societies, regions, and future
strategies will play an important role in view of the size of the energy issue. This is
true as well for the two following domains.
20.3 Mobile and Wireless Energy Scenarios
The wireless and mobile interconnect/communication infrastructure has very

specific energy constraints, which were covered in Chaps. 13 and 5. We recapitu-
late here that the quality of the mobile-communication world for billions of users
426 B. Hoefflinger
is anything but certain, unless 1,000-fold improvements in processor power-

efficiency can be realized. Among the four constituents, processors, memory,
displays, and transceivers, we have good prospects for processors and memories
to meet the challenge (Sect. 3.6, Chaps. 11 and 13). Displays have their own
specific roadmap for power-efficiency, measured in mW/(Mpix/s), where a 10
larger resolution and a 10 reduction in drive-power/pixel would save the status
quo in energy. However, the always-on, always-connected goal at constant or
reduced battery energy will require highly intelligent snooze modes and on-board
energy harvesting (Chap. 19). The expected video and cloud-computing bandwidth
causes an insatiable appetite for transceiver energy at the terminals as well as in the
cellular grid. The prospects for transceiver power-efficiency in Fig. 13.6 and
according to Table 5.5, and for bandwidth in Fig. 5.4 put serious limits on a high-
rate sustainable expansion. Breakthroughs in bit-rate compression and in the short-to-
wide-range handshake will be needed. This makes us look at the long-distance
evolution.
20.4 Long-Distance Fiber- and Satellite-Communication

Energy
For decades, the installed long-line bandwidth, both cable and satellite, has been an
invitation to invent ever more sophisticated web-based services. However, recently,
the worldwide capacity has been filled up rapidly with the overwhelming expansion
of high-resolution, high-speed video traffic. For its further potential expansion in
the current decade 2010–2020, estimates could be:
– Low: (bit-rate 10) (users 100) ¼ 1,000
– High: (bit rate 100) (users 1,000) ¼ 105
Capital investment, energy, and environment will dictate how much of this
potential can be sustained, and business creativity will be the surprise factor
pushing this expansion.
The other major new-use component is cloud-computing and web-supported
large-scale education (Chap. 22). More and more regions worldwide provide web-
based educational materials and library and laboratory access to their schools –
free-of-charge. Estimates of growth for these services are difficult because they
depend on regional and global policies as well as new organizational structures for
these services.
In any event, the challenges on providing the electrical energy needed for the
hardware in the internet, as well as at the user sites, have dimensions similar to
those of the data centers, that is hundreds of gigawatts in 2020. These communica-
tion nodes and user facilities are all candidates for renewable energy.
20.5 Conclusion
We have seen in the technology and product chapters that we have the innovation
potential to increase significantly all chip functions, processors, memories, com-
munication, which would lead to >1/3 of the total world electric power being used,
a level that would not be economical or sustainable. In any event, to support
reasonable growth, large-scale development, driven by serious requirements for
systems-on-chip with minimum energy-per-function is needed, starting from
proven prototypes and meeting significant milestones in short periods of time.
References
1. www.hotchips.org
2. McCue, A.: Energy could eat up to 40% of IT budgets. www.businessweek.com (2006). 7 Nov
2006
3. Taft, T.K.: IBM building zero-emission data centers. www.eweek.com (2009). 6 Nov 2009
Chapter 21
The Extreme-Technology Industry
Bernd Hoefflinger
Abstract The persistent annual R&D quota of >15% of revenue in the semicon-
ductor industry has been and continues to be more than twice as high as the OECD
definition for High-Technology Industry. At the frontiers of miniaturization, the
Cost-of-Ownership (COO) continues to rise upwards to beyond 10 billion $ for
a Gigafactory. Only leaders in the world market for selected processors and
memories or for foundry services can afford this. Others can succeed with high-
value custom products equipped with high-performance application-specific
standard products acquired from the leaders in their specific fields or as fabless
original-device manufacturers buying wafers from top foundries and packaging/
testing from contract manufacturers, thus eliminating the fixed cost for a factory.
An overview is offered on the leaders in these different business models. In view of
the coming highly diversified and heterogeneous world of nanoelectronic-systems
competence, the point is made for global networks of manufacturing and services
with the highest standards for product quality and liability.
21.1 Extreme Technology
The term high technology arose in the 1950s to denote industries with a significant
commitment to research and development (R&D). Since the 1970s, the OECD
has classified industries as high-tech, if their R&D budget exceeds 7% of their
revenues. A ranking of such high-tech industries is shown in Table 21.1.
As another reference, electrical machinery and apparatus with 3.6% R&D is
classified as medium-high-technology. A comparison with the R&D percentage of
the semiconductor industry, shown in Fig. 21.1, demonstrates why the classification
extreme-technology industry is justified. The R&D percentage has been above
B. Hoefflinger (*)

430 B. Hoefflinger
Table 21.1 R&D quota in Pharmaceuticals 10.5%

high-tech industries (% of
Air and space 10.3%
revenues)
Medical, precision, optical 9.7%
Radio, TV, communication 7.5%
Office and computing machines 7.2%
Fig. 21.1 Integrated-circuits industry R&D expenditure [1]. # IC Insights
15% for the IC industry since the mid-1990s, twice as high as the closest electronic-
equipment industry, TV, communication and computing. The leaders in the indus-
try have spent and are spending closer to 20% or even higher.
This is only the R&D effort. In a way more extreme is the investment in facilities
and equipment, which has consistently exceeded the R&D. The semiconductor
industry has had and continues to have unique dynamics in its evolution and
continuing expansion. The global picture is well represented by the installed
wafer capacity (Fig. 21.2).
In 2010, investment in facilities and equipment ran high at ~$60 billion, arriving
late after a cautious 2009 and being effective only in 2011. Therefore, wafer
capacity in 2010 did not meet demand, causing healthy prices and great revenues
for 2010. However, the next unfavorable ratio of demand/capacity can be predicted
for 2012. These extreme cycles and the sheer magnitude of required capital led to
three distinct business models:
– The original-device-manufacturer model
– The foundry model
– The fabless-manufacturer model
Together, the top-10 representatives of these models serve well over 50% of the
world market. However, we look at the remaining players as well, because they are
21 The Extreme-Technology Industry 431
Americas Europe Japan Korea Taiwan China ROW
4.08
4.50
4.00
3.10
3.00
3.50
2.78
2.76
Millions of Wafers
2.71
2.66
3.00
2.31
2.10
2.50
1.89
1.88
1.81
1.76
1.63
2.00
1.37
1.25
1.21
1.12
1.04
1.03
1.00
1.50
1.00
0.50
0.00
July 2010 July 2011F July 2015F
12.36M 13.37M 16.77M
Fig. 21.2 The installed capacity in wafers/month (200 mm equivalents), July 2010, July 2011, and
July 2015 [2]
often leaders in their field with application-specific integrated chips (ASICs) or

chip-systems. This type of customization is also being served by the PCB (printed-
circuit-board) industry, which merits attention for that reason.
21.2 The Original-Device-Manufacturer Model
In the early years of microelectronics, the 1960s, chips were produced by semicon-
ductor divisions within vertically integrated electronics enterprises, with some
notable exceptions such as Intel (Chap. 2). Intel has been the world leader since
the early 1990s and continues to hold that position, as shown in Fig. 21.3.
Besides Intel, five others have a history throughout most of the 50 years of
microelectronics. The Koreans and Taiwanese started in the 1980s. There are many
dramatic stories behind these leaders. The striking observation about the 5-year data
is that the memory heavyweights grew much faster than the group 5-year average of
35%, and that the world’s largest foundry also did much better than this average.
21.3 The Foundry Model
By the early 1980s, the business cycles and the necessary critical-size investments
led several leading chip manufacturers to reduce their risk by securing some wafer
capacity through shares in a contract manufacturer for wafers, which later received
the label silicon foundry. With his formidable track record at Texas Instruments,
432 B. Hoefflinger
2010 Revenue (est.) : Bio.$ 5-year growth
1 INTEL USA 39.2 +10 %

2 SAMSUNG KO 31.7 +72 %
3 TI USA 12.8 +11 %
4 TSMC TW 12.8 +51 %
5 TOSHIBA JP 12.3 +38 %
6 RENESAS JP 11.2 +35 %
7 HYNIX KO 11.0 +89 %
8 ST Micro FR / IT 9.8 +8 %
9 MICRON USA 9.1 +76 %
10 ELPIDA JP 7.3 +94 %
Avg. Top Ten +35 %
Total Top 10 : 157.2 Bio.$ 50 % World 2010
Fig. 21.3 The world top-10 chip manufacturers in 2010 and their recent 5-year growth
Si Foundries 2008 Revenue (Bio.$)
1 TSMC TW 10.5
2 UMC TW 3.1
3 GlobalFoundries 1) USA 2.6
4 SMIC CN 1.4
5 Dongbu KO 0.5
6 Vanguard TW 0.5
7 IBM 2) USA 0.4
8 Samsung 2) KO 0.4
9 Grace CN 0.3
10 He Jian CN 0.3
1) 2009
2) Foundry Business
Fig. 21.4 The leading Si foundries with their 2008 revenue
Morris Chang became the most famous leader, returning to his home country
Taiwan and founding TSMC (Taiwan Semiconductor Manufacturing Corporation).
Offering top chip technology totally to outside customers became in a way the
silicon open source, like open-source code in software, launching a tsunami of chip
innovations starting in the 1980s, continuing through today and becoming even
more important in the future, now that the super-foundries are at the forefront in Si
nanotechnology.
The top-10 Si foundries are listed in Fig. 21.4 with their 2008 revenue, with the
exception of Global Foundries, which is listed with 2009 data after Chartered Semi-
conductor (Singapore), the global number 3 until 2008, joined in the second year of
this new enterprise, which also includes now the foundry business of IBM. IBM and
Samsung’s foundry business are examples of world leaders with the mixed portfolio of
original, proprietary products, and outside services (even for potential competitors),
a seemingly touchy constellation, which IBM has practiced since the mid-1990s.
The dynamics in the foundry scene show that the tremendous Si technology
challenges are mastered with adaptive business models, and, in some prominent
cases, with a radical decision from the start or early-on. This leads us to the group of
fabless chip suppliers.
21.4 The Fabless Model
Companies with a fables business model contract Si-wafer manufacturing, packag-

ing, and testing to outside sources, namely to foundries and contract manufacturers.
Their 2009 ranking is shown in Fig. 21.5.
It is interesting to note that many of these are leaders in their fields, Qualcomm
and Broadcomm in very-high-speed transceivers, Xilinx and Altera in field-
programmable arrays, and that they require from their foundries advanced
technologies with special features. They demonstrate that high-performance chips
Fabless IC Suppliers Revenue 2009 (Bio.$)
1 Qualcomm USA 6.5
2 AMD USA 5.3

3 Broadcom USA 4.2
4 Media Tek TW 3.5
5 NVIDIA USA 3.1
6 Marwell USA 2.7
7 Xilinx USA 1.7

8 LSI Corp. USA 1.5
9 Altera USA 1.2
10 Avago USA 0.9
Fig. 21.5 Fabless IC suppliers in 2009

434 B. Hoefflinger
can be brought to the market without an in-house wafer-manufacturing line, and

that the ecosystem for the production of high-performance chips has a lot to offer.
Competition for the next node on the roadmap and global alliances have enabled
a large-scale sharing of progress for all three models. What may future strategies
bring at the end of the roadmap?
21.5 Manufacturing Strategies Towards the End

of the Roadmap
In the past, a definitive competitive strategy for the leaders in chip manufacturing
has been to reduce the minimum features on a wafer faster than the competitor. The
consensus is growing rapidly that this linear strategy will come to an end between
16 nm (2015 node) and 10 nm (2018 node)
– Economically because of diminishing returns and
– Physically because of atomic variance and vanishing amplification.
At the same time,
– Power-efficiency of processing, memory and communication can be improved
100–1,000 (femtoelectronics), and
– The use of power and, really, energy is becoming the no.1 global force for
progress – and economic growth.
Therefore, the future performance questions for the chip industry will be:
– Which femtojoule (104 eV) and attojoule (10 eV) functions/bit do we offer in our
portfolio?
– By how much do we reduce the energy/year necessary for the production of
a chip function from sand to product?
This will amount to a life-cycle assessment of the chips and ultimately of the
systems into which they are integrated, and of their operating life. With this
perspective, it is worthwhile looking at the whole network of facilities involved
in the creation and operation of micro- and nanochip systems.
21.6 Networks for the Creation of Micro-

and Nanochip Systems
Our attention to the giga- and megafactories in the previous section and, regarding
their energy and investment characteristics, in Chaps. 8 and 20 is well justified
because of their challenging dimensions. We now consider these factories as
backbones in an elaborate network of facilities, which are all essential in the design
and production of optimum-performance and high-quality chip-based products for

highly diversified customers, as illustrated in Fig. 21.6.
The world’s leading and aspiring regions in nanoelectronics have to be
concerned about the total quality of this network and its facilities, regionally and
globally. We highlight the diversity of facilities in this network in Table 21.2 by
looking at the global characteristics of three types of manufacturers, namely those
for
– Standard chips
– Custom chips and microchip modules (MCMs)
– Printed-circuit boards (PCBs)
Fig. 21.6 Networks and value chains for the design and production of chip systems
Table 21.2 Scenario in 2000 of chip and PCB manufacturers

Standard Custom chips and PCBs
chips micromodules
World market [$ billion] 160 40 60
No. of companies 200 400 4,000
No. of companies for 90% of the market 40 200 1,500
Top-10 revenue share 50% 5% 3%
Avg. revenue of top 10 [$ billion] 10 0.2 0.2
No. of design starts/year 30 k 200 k 2 mio.
Avg. volume [units/product] 5 mio. 10 k 5k
436 B. Hoefflinger
We include the PCB industry because their role is changing fast with the level of
integration on chips and systems-on-chip (SOCs).
Contrary to the importance and to the dominance of the top-10 suppliers of
standard chips in their segment, the supply of custom chips and chip systems,
including PCBs, is much more distributed in market shares and, at the same time,
characterized by millions of design starts/year, without which standard chips would
not reach their large volumes.
This constituency of manufacturers sees a continuing and often disruptive shift
both in their product complexity as well as in their technology complexity:
– The progress of integrating more functionality on a chip continues to be driven
by cost, performance, and energy.
– SOCs and 3D chip stacks mean more customization, more specific services, and
fewer units/product.
– The PCB industry has to follow a roadmap, synchronizing itself with the
advancing interconnect hierarchy (Fig. 3.39 and Chap. 12).
With the increasing importance of 3D integration of chip stacks, where the chips,
processors, memory, transceivers, MEMS, etc. come from different vendors, the
diversity and specialization in services will increase, and entirely new levels of total
quality management (TQM) have to be established. The new SOCs and SIPs
(systems-in-package) with chips from many sources and complex value chains
require
– New standards,
– New test strategies,
– New compliance rules,
– New warranty and liability rules,
– Life-cycle assessment.
At the end of the scaling roadmap and in the era of the energy roadmap,
a diminishing part of the total investment will go into the gigafactories, and
an increasing part will go into the most effective network to produce the best 3D
multi-chip systems.
21.7 Conclusion
Chips have become the basis of 50% of the world’s gross product, and this share
will rise. The ingenuity of the concept has triggered and feeds an unprecedented
research, development, and factory investment of over 30% of annual revenue, and
that with exponential growth for half a century. The R&D expenditure in the chip
industry was more than $400 billion over the past 10 years, and the accumulated
know-how in this industry is equivalent to over three million man-years, not
counting the millions from the concurrent and essential software industry. It will
take a lot of manpower and many years to make any measurable impact on the
continuing silicon innovation wave with alternative technologies such as molecular,

bio-, or quantum electronics.
The move of the mega and gigafactories to Asia continues. Soon, more than 80%
of the world’s chips will come from the Far East (Fig. 21.2).
In spite of Europe’s investment in megafabs between the late 1980s and the early
2000s, Europe’s share of the world production of chips never exceeded 10%.
Therefore, Europe has been a net importer of finished chips and of custom-
processed wafers for a long time. The only surviving wafer fabs are high-value-
added fabs for proprietary smart chips with integrated sensing and actuating.
Substantial employment can result only from creative, early entry into sizable
markets such as the internet, mobility, and health care in the high-value-added
early phase before the rapid worldwide commoditization. This requires a highly
qualified work force, highly innovative companies with global alliances for mar-
keting and standardization, and super-critical investment to reach the market in
time.
The US share of global chip production has steadily declined from 50% in the
1980s to 15% in 2010. However, the US has been and still is the single largest
provider of manufacturing know-how and the largest customer for the Far-East
silicon foundries, preferably in Taiwan and Singapore. The US has by far the largest
integrated network power to open and serve the world’s dominant markets such as
the internet and personal computing. Even so, the implementation of their
strategies, such as interactive and internet TV, has been delayed due to a lack of
manpower for product development rather than by a lack of chip technology. In
order to be a part of the global competition for qualified manpower, US
corporations and universities have established a vast cooperative network with
the many rising universities in Taiwan, South Korea, India, and China, which
have grown since the 1980s at a rate much faster than the established academia in
the USA or Europe and Japan. Both the return of US-educated compatriots and the
establishment of affiliated campuses in those countries will secure a super-critical
force for innovation. However, the concern over qualified manpower and over the
rate of innovation persists in the USA and everywhere else where nanoelectronics is
considered to be essential, as we will assess in the following chapter.
References
1. The McClean report . www.icinsights.com (2010)

2. IC Insights, after D. McGraw: Taiwan to beat Japan wafer capacity. http://www.eetasia.com/
ART_8800626005_480200.HTM. Nov. 16, 2010
Chapter 22
Education and Research for the Age
of Nanoelectronics
Bernd Hoefflinger
Abstract Nanoelectronics has great potential for further, sustainable growth, and
this growth is needed worldwide, because new chips provide the technology
foundation for all those products and services that shape our lives. However, the
concern is justified that this truth is not the perception of the public in the first
decades of the new millennium. How can we work towards a broad, sustained
commitment to an innovation ecosystem involving education, research, business,
and public policy?
Reminding ourselves of the 10x programs invoked in Chap. 2 to describe major
milestones in advancing microelectronics towards today’s nanoelectronics, we
notice that all of them demanded requirements-driven, top-down research with
ambitious, often disruptive targets for new products or services. Coming closer to
the end of the nanometer focus, the new task of global proportion should be a femto-
Joule focus on minimum-energy nanoelectronic systems research.
22.1 Chips in High School
The children in our schools are the critical long-term resource for a creative, healthy
society. However, in spite of ICT (information and communication technology)
being the largest contributor to global economic growth in the past 40 years, the
interest of the public, and with it within schools, in engineering and microelectronic
engineering, in particular, has declined since the early 1990s. For an excellent
review of important studies on this subject, see John Cohn’s lecture “Kids today –
Engineers Tomorrow?” [1]. His statement “What we need is a new Sputnik to
galvanize the next generation of students. . .” (and we might add the public), makes
us reflect on the waves in electronics innovation, which we covered in Chap. 2.
B. Hoefflinger (*)

440 B. Hoefflinger
It was indeed Sputnik and Apollo that launched the rise of integrated circuits in the
1960s, prominently in the US. This investment waivered in the 1970s in the US for
well-known reasons, letting the Japanese take over the lead, and it needed the
renaissance of 1981 to generate a worldwide engagement in microelectronics as an
industrial – and, very significantly – as a public policy, a phenomenon unparalleled
since then. It was in a 1984 meeting at MCNC (Microelectronics Center of North
Carolina) that their communication expert showed me the Lego transistor model,
which she used in her classes with elementary-school children. I have used Lego
since that day, and you will find examples in Figs. 2.9 and 3.5, no longer as a toy,
but now, due to their buttons, as a visualization of individual boron and arsenic
atoms in a nanotransistor.
Certainly, the internet and educational-robotics programs have risen to a formi-
dable offering for children in the meantime, and a listing of proven programs is
contained in [1]. At a more advanced level and in the serious concern to carry
microelectronics to high-school juniors in time for their professional inclinations,
we launched a 2-week intensive lab course on the making of microchips in 1992,
which has since been offered annually (Table 22.1 and Fig. 22.1).
The course program, although challenging, with its subject matter reaching from
freshman- to junior-year material, has been very well received by the high-school
students. Over the past 18 years, more than 600 students from some 60 high schools
participated, among them 30% girls. Returning to their schools, the students
reported in their classes about their experience, thereby multiplying the effect of
this program. The experience with that material was used to establish a vocational
training program in collaboration with regional SMEs (small and medium
enterprises). This program and its certificate have been accepted by the regional
association of electronics enterprises as an additional professional qualification for
electronics technicians.
22.2 University Programs in Micro- and Nanoelectronics
The teaching of microelectronics has been promoted throughout its history for
60 years by monographs and textbooks written by authors based in the US (and
Canada). It is important to note that the originators of new material in the early
years were the world leaders in industrial research:
– William Shockley: Electrons and Holes in Semiconductors, 1950 (Bell
Laboratories)
– Simon Sze: Physics of Semiconductor Devices, 1970 (Bell Laboratories),
– The Series on Bipolar ICs, ~1967 (Texas Instruments and Motorola) and MOS
ICs, 1970 (Texas Instruments)
The leading universities, especially Berkeley and MIT, were quick to structure
and present these new topics in textbooks, including important sections on
problems (and solutions), and the competitive US university system established
22 Education and Research for the Age of Nanoelectronics 441
Table 22.1 SMC (Sch€ uler und Sch€

ulerinnen machen Chips) – High-School Juniors Make Chips,
a 10-day laboratory course
Day 1
C1 Microchips for health, mobility, and entertainment
C2 The electron, force, energy, potential
C3 Resistors, capacitors, voltage and current sources
L1 Tour of the institute
C4 Ideal and real amplifiers
C5 Semiconductors – the shockley garage model
Day 2
C6 PN-junction and diode
C7 MOS transistor
C8 Analog and digital signals, codes
C9 Logic gates and standard cells
C10 CMOS logic – the inverter, its gain and noise margin
M1 Meeting with engineering students
Day 3
C11 CMOS gates and counters
C12 Field-programmable arrays
L1 Design of a digital clock (group B)
Day 4
C13 Logic synthesis with VHDL I
C14 Logic synthesis with VHDL II
C15 Logic synthesis with VHDL III
L1 Design of a digital clock (group A)
Day 5
C16 Digital circuits testing
C17 CMOS process technology front-end
C18 CMOS process technology back-end, parameter test
L2 Design of a digital counter (group B)
M2 Reunion with previous classes
Day 6
L2 Design of a digital counter (group A)
L3 Chip manufacturing line (group A)
L4 Counter in VHDL (group B)
Day 7
L5 Programming of the FPGA (group B)
L3 Chip manufacturing line (group B)
L4 Counter in VHDL (group A)
Day 8
C19 ASICs – from specification to test
C20 Test program generation
C21 Quality and reliability
L5 Programming of the FPGA (group A)
Day 9
L6 Clock printed-circuit board assembly and test (group B)
L6 Clock PCB assembly and test
Day 10
M3 Professional development – an industry representative
C22 High-dynamic-range CMOS video cameras
M4 Certificates and course review/feedback
442 B. Hoefflinger
Fig. 22.1 High-school juniors make chips, class of 2000 (www.ims-chips.de)
Fig. 22.2 US engineering degrees/year [1]. # 2009 IEEE
the expansion within the electrical-engineering curricula, supported by a national

accreditation program, executed by the AAEE (American Association for Engi-
neering Education), whose teams consist of peer professors as well as industry
leaders. This system, fed by the world’s best students from both home and abroad,
enabled the annual number of US engineering degrees to rise from 40,000 in 1960
to a peak of almost 100,000 in 1985 (Fig. 22.2) and led to the sustained world
leadership of US universities.
A worldwide ranking of engineering programs generated by Shanghai Technical
University [2], which is quoted widely, shows among the top 50 engineering
programs:
34 from the US (among them 15 at the top)

4 from the UK
4 from Japan
3 from Canada
1 from Switzerland
1 from Singapore
1 from Hong Kong
1 from Israel
1 from Taiwan
Table 22.2 shows more details of this global ranking:
In spite of this sustained leadership, Fig. 22.2 shows that the US system has
stagnated below 80,000 engineering degrees/year since 1990. At the same time,
China has strongly increased its engineering output so that by 2008 their number of
young engineering professionals became three times that in the USA (Fig. 22.3).
The concern about US leadership is quoted here because, in the past, it has been
the largest force to revive and to re-focus the innovation system, not only in the US
but also globally. The concern has been voiced in the US since 2001 with increasing
intensity, and it has been addressed as a perfect storm: “There is a quiet crisis in US
science and technology. . .This could challenge our pre-eminence and capacity to
innovate. . .” (Shirley Ann Jackson, President of Rensselaer Polytechnic Institute,
2004, quoted from [4]). The global financial and economic crisis of 2008/2009 has
made it even clearer that the exit strategy has to be focused innovation for global
markets and services, an Apollo program as the answer to Sputnik (where the
present crisis is much larger than Sputnik). Apollo was the largest requirements-
driven program ever, and there will certainly not be a simple analogue today.
However, we will look at the present situation in nanoelectronics research and
technology transfer with special attention to the requirements that have been set
before the scientists and which they have promised to meet.
Table 22.2 Top 10 university engineering programs USA/ROW 2007 [2]

USA Rest of world (ROW)
1. MIT 16 Cambridge U., UK
2. Stanford U. 17 Tohoku U., Japan
3. U. of Illinois, Urbana-Champaign 19 U. Toronto, Canada
4. U. of Michigan, Ann Arbor 25 Kyoto U., Japan
5. U. of California, Berkeley 27 Imperial College, UK
6. Penn State U. 28 EPF Lausanne, Switzerland
7. Georgia Tech 28 Tokyo Institute of Technology, Japan
8. U. of Texas, Austin 32 Singapore U., Singapore
9. U. of California, San Diego 37 Hong Kong U. of Science and Technology,
Hong Kong
10. Purdue U., IN 38 Technion, Haifa, Israel
444 B. Hoefflinger
Fig. 22.3 The number of young professionals in the US, China, and India [3]
22.3 Large-Scale Nanoelectronics Research
The microelectronics age has had its two prominent international scientific
meetings for well over half a century, and they continue to be the most representa-
tive yardsticks for the global development of micro- and now nanoelectronics:
– The International Electron-Devices Meeting (IEDM), held annually in Decem-
ber, as the name says with a focus on new devices, their physics, technology and
characterization.
– The International Solid-State-Circuits Conference (ISSCC), held annually in
February, the no.1 event for the presentation of new chips, and most often
quoted in this book.
To get a view of the global participation in 2010, we show the origins of the
scientific papers by country in the categories universities and research institutes
and industry in Fig. 22.4. As a foretaste, we note that, among the more than 450
papers, 85 are the result of industry–university/institute cooperations, and of these
46 are international/global ones, a remarkable new development since ~2000. We
have counted these cooperations in both categories to weigh in their significance.
The US lead in papers remains significant, and it has gained a visible cooperative
note with well over 20 cooperations. Japan is holding a solid second place. Signifi-
cantly, Taiwan and Korea have become key contributors to global innovation.
Chinese universities appear on the map as well. The single most remarkable force
is IMEC in Belgium (the Inter-University Microelectronics Center), which has
become the world’s largest nanoelectronics research center. One other message is
the success of the university–industry network of Grenoble, France, and northern
Italy, around the major locations of ST Microelectronics. This has given France/
Italy a significant scientific output in nanoelectronics.
Fig. 22.4 Origins of the Universities and Research Institutes

scientific papers presented at
the 2010 ISSCC and IEDM 1 USA 110
2 Japan 42
3 Belgium 32
4 Italy 25
5 Taiwan 22
6 NL 17
7 France 15
8 China 14
9 Korea 14
10 Switzerland 12
11 Germany 11
12 UK 10
13 Canada 4
14 Spain 3
Industry
1 USA 118
2 Japan 53
3 Korea 21
3 Taiwan 21
5 France 17
6 Italy 15
7 NL 10
8 Switzerland 6
9 Austria 3
9 Canada 3
9 Germany 3
The ranking in Table 22.2 reflects very closely the 2010 presence of US
universities at ISSCC and IEDM. However, the ROW scenario of top universities
regarding nanochip innovations is marked by great dynamics, which is supported
by the world’s leading chip manufacturers (Chap. 21). Therefore, on a global map
of nano-chip research regions, we have to look at the following:
USA
– NANO NY, Albany NY with IBM, AMD, and Global Foundries.
– The universities in Table 22.2, related institutes with the US companies
among the top 10.
Japan
– Tokyo-Kanagawa Pref. with the Japanese companies among the top 10.
446 B. Hoefflinger
Taiwan
– Taipeh-Hsinshu with TSMC, UMC, and MediaTek.
Korea
– Seoul-Taejon with Samsung and Hyundai.
Europe
– IMEC, Leuven, Belgium, cooperating with most of the world top 20 companies.
– Grenoble, France, and Lombardy, Italy, with ST Micro, Micron, and
Numonix.
These top regions and their governments have highly diverse cultures in their
ultimate task: Transfer of nanoelectronics innovations.
22.4 Requirements-Driven Research and Innovation

Transfer in Nanoelectronics
The concern in this section is the timely question of how to energize the innovation
constituency of academia, public research institutes, and industrial research
divisions with a sustained drive out of the 2008/2009 world crisis and consecutive
critical conditions, accepting that in many parts of the world, notably the West, that
is North America and Europe, in the words of the US Commerce Secretary [5], “The
Innovation Ecosystem is broken.” In his speech on January 15, 2010, before the
PCAST, the President’s Council of Advisors on Science and Technology, he said:
“Just a decade or two ago, the US had the unquestioned lead in the design and
production of things like semiconductors, batteries, robotics and consumer elec-
tronics. No more. Our balance of trade in advanced technology products turned
negative in 2002, and has shown little signs of abating.”
The European balance of trade in semiconductors has always been negative,
with the notable exception of automotive semiconductors, and the situation of
Germany and Great Britain, as illustrated by Fig. 22.4, shows that the US concern
should be a concern in Europe as well. The research community, together with the
research policy makers in governments, has to commit itself to initiate a sustained
recovery with tangible, real-world solutions.
As far as nanoelectronics is concerned, approaching the end of the roadmap (an
effective, simple requirement in the past) and after two decades of too much
scientific optimism about quantum-nano solving all problems, it is time to ask
which top-down requirements nanoelectronics should meet with clear milestones
and with a tight time-frame.
We have seen throughout this book that energy efficiency in processing, memory
and communication is the single most important task and the most effective one
generating new performance and added value. And let us remember again: power
efficiency tells us the atomistic value of how much energy we need for a single
operation, but not how much energy we need to solve a task such as compressing
one HDTV frame to a minimum of bits for communication.
It is an interesting test of relevance to search recent major research programs for
power efficiency or, better, energy-efficiency goals. We covered the Silicon Brains
program SyNAPSE [6] under the leadership of DARPA, and, as a case in point, it is
demanded there that the electronic synapses perform an operation with an energy
<10 pJ. Another DARPA program, Ubiquitous High-Performance Computing [7],
also sets clear goals: Build a PetaFLOP-class system in a 57 kW rack [7]. In a
simple model, like the one we used in Chap. 20, where we assumed that processing,
memory, and interconnects each take about 1/3 of the total power, this would leave
19 kW for 1015 floating-point operations/s or 19 pJ/FLOP, which is mostly a
54 54-bit multiply. This is one tenth that of the best comparable graphics
processor in 2010 (200 pJ) and a significant challenge if we consider the data on
short-word-length multipliers in Fig. 3.32.
By contrast, if we read through the program statements of ENIAC, the European
Nano Initiative [8], there is nothing other than nanometer nodes of the ITRS and
“more-than-Moore” commonplaces. The 2010 update of the European 7th
Programme for ICT for the tasks 2011–2012 is the most generic catalog of hit-
words without any target other than at least asking for “radically improved energy
efficiency”[9]. While energy efficiency is the right performance goal, why not
demand from the research community a reduction in the energy for recording,
compression, communication, decompression, and delivery of a single HD frame to
a display by at least a factor of 10,000. This would make a serious 10 program for
the 2010s, considering that video-bandwidth and -energy is the no.1 roadblock for
high-quality video, not only for entertainment but, more seriously, in professional
applications such as health care and public safety.
While system energy is a requirement forcing a top-down approach (researching
functions, components, and physics at the bottom), the higher levels of system
applications and benefits for the public should be focused on. The ones with the
biggest public interest and with broad government obligations are
– Health and care
– Education
– Safe mobility
– Communication
Here again, test questions have to be asked regularly:
– What is the dedication in funded research on these subjects,
– Their requirements on nanoelectronics,
– Their benefits from nanoelectronics and, most important and most often
forgotten,
– Their sustained program of implementation?
448 B. Hoefflinger
22.4.1 Nanoelectronics in Health and Care
The potential of intelligent, portable, potentially energy-autonomous diagnostic and

therapeutic nanochips is boundless so that we can only name a few representative
ones: The development of a superficial, intelligent heart-monitoring coin for all
patients with potential heart problems would benefit hundreds of millions of
patients as well as their health-care and health-insurance systems. Supporting the
introduction of this device through public health insurance would accelerate its
development and its approval, and it would push low-energy chip technology in the
semiconductor industry with spin-offs into many other new products.
22.4.2 Nanochips for Safe Mobility
Saving lives and preventing injuries and life-long handicaps in health care and
traffic are tasks with a tremendous social significance, certainly as great as
providing energy compatible with the environment. In the latter domain, we see
billions of dollars being spent in subsidies for renewable power generation such as
solar and wind. Avoiding fatal accidents on public roads caused by tired or
distracted truck and bus drivers has been proven by hundreds of thousands of
kilometers of test drives by the automotive industry, and as yet no legal requirement
for these safe-distance and automatic-braking systems has been issued by the
governments and supported with some public cost sharing for their rapid
introduction.
The rapidly advancing urbanization of our world presents formidable challenges
and opportunities for nanoelectronics-enabled ICT to support public and personal
mobility. The intelligent-transportation-systems (ITS) programs in Europe, Japan,
and the USA merit a review and a new, sustained effort in view of the new era of
mobility driven by environmental and life-cycle energy-efficiency issues.
22.4.3 Education with and for Nanoelectronics
The Texas Instruments Speak & Spell system introduced in the late 1970s was a
pioneering technical achievement with its linear-predictive-coding (LPC) voice
encoder, and it was the most effective toy ever to advance language skills. Educa-
tional electronics today could be incredibly more powerful than Speak & Spell, and
yet it is a domain with very patchy success given the base of billions of potential
customers from kindergarten to old age. In Chap. 6, on requirements and markets,
we covered broad issues such as “Open Libraries for Practically all Children” and
carebots as personal tutors and assistants in rehabilitation. These examples raise the
extremely challenging issue of producing and supporting content for the application
of nanoelectronic systems – and for the training of knowledge in nanoelectronics.
Education and continuing-education (CE) programs in microelectronics were
a vital part in corporate CE programs and nationally as well as internationally
(like COMETT in Europe) funded R&D programs. In spite of the boundless
opportunities with the internet, there are no critical-size public programs today.
Several world-class universities have made internet course offerings into a business
model, and a truly interactive mode should receive sustained support. This requires
standards and certification, as were set up in the USA by the AAEE and the NTU
(National Technological University) for satellite-broadcast programs.
22.5 Conclusion
Overall, it is worthwhile reconsidering the public research strategy. The advent of

the transistor and of the integrated circuit has not been the result of financing large
national research labs for the advancement of physics. It was rather driven by
ambitious system requirements, in system language: not bottom-up, but top-
down. First came the radar requirement and then the massive-computing require-
ment of the military establishment in the 1940s and 1950s. Soon after, in the 1960s,
the requirement by the Japanese industry of a single-chip desk-top calculator
changed the world, again a system requirement forcing a creative solution with
the best, albeit “available” technology, not a search for new physical phenomena.
ARPANET, the predecessor of the internet, is another example of the significance
of an ambitious government requirement. Similarly, the European success of the
digital cellular phone based on the GSM standard became possible with the long-
term strategic planning of the still functional public telecom authorities requiring
and pushing a joint standard, while the USA was occupied with taking its telecom
system, majestically guided by ATT, apart.
The lesson for our public and governments in their efforts to support innovation
for growing markets and employment should be to assign more resources to realize
ambitious systems and services requirements with top-down R&D, noting that there
is a lot in place and sitting at the bottom.
References
1. Cohn, J.: Kids today! Engineers tomorrow? In: IEEE ISSCC (International Solid-State Circuits
Conference) 2009, Digest of Technical Papers, pp. 29–35, Feb 2009, San Francisco, USA
2. http://www.arwu.org/fieldENG
3. BusinessWeek, Special Report: China 6 India, 22 August 2005
4. Friedman, T.L.: The World Is Flat, Chaps. 7 and 8. Farrar, Straus and Giroux, New York (2005)
5. US Department of Commerce News, Feb 2010
450 B. Hoefflinger
6. http://www.darpa.mil/Our_Work/DSO/Programs/Systems_of_Neuromorphic_Adaptive_Plastic_
Scalable_Electronics_%28SYNAPSE%29.aspx
7. http://www.darpa.mil/Our_Work/I2O/Programs/Ubiquitous_High_Performance_Computing_
(UHPC).aspx
8. www.eniac.eu
9. www.cordis.europa.eu/fp7/ICT/Objective ICT-2011.9.8, p.106.
Chapter 23
2020 World with Chips
Bernd Hoefflinger
Abstract Although we are well advised to look at the future 1 day at a time, we
have seen in the chapters of this book, and they necessarily could cover only a
selection on the features and applications of those tiny chips, that their potential
continues to grow at the exceptional rates of the past. However, the new commit-
ment has to be towards Sustainable Nanoelectronics, guided by creating sensing,
computing, memory, and communication functions, which move just a few elec-
trons per operation, each operation consuming energy less than one or a few
femtojoule, less than any of the 1014 synapses in our brains.
At these energy levels, chips can serve everywhere, making them ubiquitous,
pervasive, certainly wireless, and often energy-autonomous.
The expected six Billion users of these chips in 2020, through their mobile,
intelligent companions, will benefit from global and largely equal access to infor-
mation, education, knowledge, skills, and care.
23.1 Users of 2020 Chips
The six billion users in 2020 will use, on average, about 40 new chips per year, on
top of the hundreds they will have had in use statistically, giving everyone the
power of a supercomputer. Most of these chips will go unnoticed, from the terabit
memory in the MP3 ear plug to the heart monitor in your shirt. The always-on
mobile companion will be ready to perform trillions of multiplies per second to
provide its owner with high-quality video and animated information.
The access to information, education, knowledge, and skills will be global and
largely equal. The number of users in Asia will be almost ten times from the number
in North America and Europe.
B. Hoefflinger (*)

452 B. Hoefflinger
The most visible products powered by minimum-energy chips will be paper-thin

e-books and tablets for information and education, and, at least in Japan, Korea,
Taiwan, and China, personal robots for children and for the elderly and
handicapped. These shocking new companions will change human lives more
than the phone, the car, the TV, and the internet. Health monitoring, diagnostics,
therapy, and telemedicine will be most significantly advanced by 2020, again
enabled by high-resolution and large-bandwidth imaging and video. Personal and
public, safe and effective mobility can be greatly enhanced by intelligent
nanoelectronic systems. Both health and mobility are a public concern – and
obligation.
23.2 Regional and Global Policies for Sustainable

Nanoelectronics
Given the potential and power of chips in 2020, a valid question is whether this
progression happens just because of market forces such as competition and new-
product appeal or as a result of long-term strategies such as those pursued in Asia or
by DARPA in the USA.
R&D strategies should focus on system requirements in health, education, and
mobility, and they should be accompanied by certification policies to assure the
beneficial introduction of these new, intelligent nanosystems such as silicon brains.
23.3 The Makers of 2020 Chips
The re-focusing from nanometers to femtojoules is another one of the more

fundamental turning points in the history of microelectronics, such as the transitions
from bipolar to NMOS and NMOS to CMOS, however, this time, not from one
technology to another, but from technology to functional energy.
Certainly, the broad introduction of a 10-nm Si process technology is a sufficiently
large challenge for a decade. But to stop pushing the nanometer competition further
and to invest instead in energy, and to get this done overnight and in parallel with the
remaining nanometer work, is very tough. The time scale is short indeed, if we recall
the energy crisis exemplified for data centers in Chap. 20. However, the focus on
nanoelectronics as a sustainable, minimum-energy, even self-sustained technology
also has the potential to interest the public and the young, in particular [1].
We identified the interconnect- and communication-energy problem in Chap. 20,
and, if we turn our attention all the way back to Fig. 1.4, we see the magnitude of the
problem. Correspondingly, since 2005, the 3D integration of thinned chips with
TSVs has appeared as one measure to shorten interconnects. This heterogeneous
23 2020 World with Chips 453
process with heterogeneous, multiple chip suppliers is reshaping the industry and
one sure sign of change.
With the large-scale wafer manufacturing – >80% of the world total – in Asia,
the West and the rest of the world (ROW) have to realize their high-value-added
share on finished chip-systems in highly innovative ways. By 2020, this will have
evolved largely on the basis of available skilled, creative manpower.
The two energy worlds of processors, general-purpose as opposed to application-
specific, minimum-energy, well described in Chap. 13, will have hit each other
decisively by 2020. Even more fundamentally, the professionals have to come to
grips with the more radical, adaptive, and self-learning biomorphic chip systems
regarding their certification and fields of applicability. Here, the chip industry has to
work closely with the public and with policy makers.
23.4 Beyond 2020
Will a 10-nm chip industry have enough potential to grow at > 4%/a beyond 2020?
The answer is yes, because all of the developments described in this book and all
those based on the same large-scale Si technology but not described here will just be
achieved by a few demonstrators and by selected leaders in their field, and the
2020s will be marked, in the first place, by the widespread insertion of these
achievements in a very large available global market.
Of course, there are longer-term research issues, and to keep these focused, we
identify some fundamentals for future large-scale electronics in a natural environ-
ment with operating temperatures >70 C, advancing, by at least two orders of
magnitude, the 2020 performance of Si nanoelectronics projected in this book:
For programming, writing, reading, and communication within the natural
environment, electrons, currents and voltages are a must. In 10nm transistors,
MOS or bipolar, currents are generated by single electrons, and the 2020 perfor-
mance levels are as follows:
Digital logic and high-speed memory:
– Differential, full-swing regenerator
– Footprint < 60 60 nm2
– Dynamic energy <7 eV
– DC energy <7 eV.
Tenfold improvements on footprint and energy are a tough challenge and/or
limited by the electron thermal energy of 25 mV at room temperature.
In any more radical innovation of high-speed, low-energy nanometer-logic, the
fundamental component still has to be a differential, amplifying regenerator such as
a cross-coupled inverter, with a variance per chip compatible with the power
supply. In a carbon-based technology, this means complementary n- and p-type
carbon nanotube (CNT) or graphene transistors. A 3D vertical, doped CNT struc-
ture with a topography as in Fig. 3.42 may be better than a thin-film graphene
454 B. Hoefflinger
implementation. In any event, a harmonious mix of, most likely short-range,

vertical and (long-range) horizontal integration should be achieved. Self-assembly
by selective, doped crystal growth should receive sustained development because of
its density, quality, and efficiency.
A different scenario exists for non-volatile (NV), field-programmable memory,
because here the selection of the token for a memory bit has a high priority. Charge-
and resistance-based memories such as multi-level per cell (MLC) Flash, phase-
change memory (PCM), and resistive RAM (RRAM) are ruling for the time being.
They have intrinsic energies per cell per bit > 100 eV and require large program-
ming energies.
A most fascinating token for a bit is the spin of the electron so that spintronics
for NV memories continues to be an important research area. Programming
(writing) and reading require electronics resulting in unfavorable energy and area
requirements (see Table 11.4). The write and read energy remain the fundamental
issue of NV memories. It is here that the single-electron nanodevice, the SE device-
chain with Coulomb confinement, built possibly with direct, focused-ion-beam
doping, has fundamental potential, if, as a thought experiment, we revive the
charge-coupled-device (CCD) serial-memory concept [2] and imagine a single
electron (or several electrons, for that matter) per bucket in a CCD, either holding
the electrons in the retention state without power dissipation, or reading them out
serially, coupled with a re-write. In such a multi-phase, clocked architecture, write
and read energies have to be assessed carefully.
The fantastic progress in understanding ionic and organic information
processing in nature will continue with exciting insights and discoveries, many of
them achieved with ever more powerful electronic sensing, processing, and pattern-
recognition functions. This is of significant merit for diagnostics, therapy, learning,
and for the interaction of electronic systems with functions and components of the
human body via surface contacts, implants, and prosthetics.
However, R&D with the aim of replacing the age of micro- and nanoelectronics
by a new age of molecular or organic computing is ill conceived, as summarized in
the final section.
23.5 Conclusions
– The electronic circuits being built in the present decade 2010–2020 perform the
operation of a neuron’s synapse with less energy and one million times faster
than nature.
– Electronics is a fundamental technical invention.
– Electronic circuits are for information and communication what wheels and
motors are for mobility.
– We should improve these with all our creativity and energy, but not try to fly like
birds!
23 2020 World with Chips 455
References
1. Cohn, J.: Kids today! Engineers tomorrow? IEEE ISSCC (International Solid-State Circuits
Conference) Digest Technical Papers, pp. 29–35 (2009)
2. Boyle, W.S., Smith, G.E.: Charge Coupled Semiconductor Devices, Bell Sys. Tech. J. Vol. 49,
pp. 587–593, Apr. 1970.
Index
A Ambient energy sources, 394

ACC. See Autonomous Cruise and AMD, 200. See Age-related macular
Anti-collision Control (ACC) degeneration (AMD)
Acceleration, 294, 296, 299–308, 312, 332 AMI, 19
profile, 396, 397 Amplification, 3
Accelerator(s), 67, 75, 223–224, 388 Amplifier, 13, 16
Accelerometers, 294, 299 Analog–digital converters, 6–7, 288
Acceptor, 45 Analog–digital interfaces, 6
doping, 51 Analog front end, 300, 305
Access Analog-Mostly Network ACE, 336–339
time(s), 230, 233, 234, 239, 240 Analog neural network, 26, 27
to information, 451 Analog neurons, 387
Action potential, 368, 369, 386 Analog sensor, 390
Active area (blue), 45, 46 Android robot, 153, 154
Active diode, 407, 408, 414 Angular rate sensing, 302–304
Active-pixel sensors (APS), 322 Anneal, 68
Activity levels, 197 Anode, 13
Adaptive connectivity, 274, 276, 282–283 Anti-aliasing, 116, 118
Adaptive power management, 287–288 API. See Application programming
ADCs. See Analog–digital converters (ADCs) interface (API)
A/D conversion, 96 Apollo, 440, 443
Adders, 75 Applicability, 453
Advanced porous silicon membrane (APSM), Application development, 215
297 Application processor, 280
technology, 308–310 Application programming interface (API), 223
Advanced RISC Machines (ARM), 147 Application-specific integrated chips (ASICs),
Age-related macular degeneration (AMD), 367 431
Aging, 220 Application-specific integrated processors
Airbag, 294, 299 (ASIPs), 331
Air bridges, 135 APSM. See Advanced porous silicon
Air insulation, 135 membrane (APSM)
Airy disk, 319, 320 APS. See Active-pixel sensors (APS)
All-ion-implanted, 63 Architectural level, 205
Always-aware, 274 Area of crystalline Si, 422
Always-on, 274 ArF water-immersion DPT, 178
Amazon, 201 Arithmetic units, 75
Ambient-brightness adaptation, 375 ARM. See Advanced RISC Machines (ARM)

DOI 10.1007/978-3-642-23096-7, # Springer-Verlag Berlin Heidelberg 2012
458 Index
ARPANET, 449 Behavior, 192

Artificial intelligence (AI), 391 Behavioral models, 192
ASA, 318, 319 BEOL. See Back-end-of-line (BEOL)
ASA 12.800, 318 Bicentennial, 15
ASIC, 151 BiCMOS, 60–67, 138
ASIPs. See Application-specific integrated BigDecimal, 224
processor (ASIPs) Billions of sensors, 293
Aspect ratio, 22 Binning of n pixels, 319
Assembly, 265–267 Biofuel cell, 402–404
Associative processing, 148 Biological brains, 385, 386, 389
Asynchronous comparators, 414 Biological sensors, 285
Asynchronous digital circuits, 414 Biological systems, 389
Atomic-layer epitaxy (ALE), 74 Biomedicine, 154
Atomic variance, 2, 86, 434 Biomorphic, 26, 79
AT&T, 64 chip, 11
Automatic braking, 448 chip systems, 453
Automatic steering, 7, 28 vision system, 326
Automotive, 145 Bio-sensing, 275
applications, 8 Biphasic output, 372
conditions, 305 Bipolar quad, 236
environments, 304 Bipolar transistor, 41–45
Autonetics, 21 BIST. See Built-in-self-test (BIST)
Autonomous Cruise and Anti-collision Control Bit-line sense amplifier, 237
(ACC), 157 Bit-rate compression, 138
Autonomous systems, 274 Blind from birth, 367
Available global market, 453 Blindness, 367–371
Available voltage gain, 56 Blind patients, 367–381
Avalanche breakdown, 327 Blood sugar, 155
Average selling price (ASP), 142 Blue brain, 388, 390
Axon cross sections, 386 Blue Gene, 226
Axons, 384–386 Bluetooth, 138
BMI. See Brain–machine interface (BMI)
Boards, 211–212
B Body-area networks (BAN), 10, 275, 285
Backbones, 434 Booth encoding, 76
Back-end-of-line (BEOL), 173, 247 Bosch process, 297
Backside-illuminated (BI), 321 Brain, 2
Back-to-back bonded wafers, 81, 82 characteristics, 389
Balance of trade, 446 power, 386
Ballistic transport length, 133 Brain–machine interface (BMI), 10, 275, 277
BAN. See Body-area networks (BAN) Braun, K.F., 15
Bandwidth, 51, 53, 216, 223, 273, 387, 395, Breakdown field, 51
398, 412, 413 Brick wall, 195
Barrier-controlled current, 55 Broadband wireless, 138
Barrier-control mode, 53, 56, 169 Built-in-self-test (BIST), 251, 270
Base Bulk-driven amplifier, 412, 413
transit time, 42 Bulk micromachining, 297, 309
width, 42 mBumps, 79
Battery, 393, 404–406 Buried collector, 41
life, 254 Buried layers, 23
technology, 276 Buried-oxide, 45
Beam-blanker, 181 Business
Beam-lead technology, 135 analytics, 216
Index 459
cycles, 431 Checkpointing, 221

intelligence, 216–220 Chemo-mechanical polishing (CMP), 74
model, 271 Chip carrier (CC), 135
Chip–chip transceivers, 135
Chip industry, 11
C Chip-level stacking, 81
C4, 79 Chip meets timing, 192
Caballero, 344 Chip-stacking, 245
Calculator chips, 21 Choroid, 369, 371
Calibration, 310 Chua, L.O., 34
Camera Cip-integrated fuel cell, 402, 410
capsule, 154 CISC. See Complex-instruction-set computing
pill, 154 (CISC)
systems, 328–330 Cleanroom footprint, 186
volume, 316 Clocked MOS logic, 75
CAMs. See Content-addressable memories Clock-tree synthesis, 197
(CAMs) Cloud computing, 147, 426
Capacitance, 47 CMOS. See Complementary metal oxide
Capacitor, 15 semiconductor (CMOS)
Carbon-based technology, 453 CNN. See Cellular neural network (CNN)
Carbon nanotube (CNT), 133, 169, 453 CNT. See Carbon nanotube (CNT)
Cardiogram, 154 Code-division multiple access (CDMA), 148
Carebot, 141, 153, 384, 391 Cognition, 384
Care-free monitoring, 155 CO2 laser, 178
Carrier frequencies, 138 Collaboration, 274
Carry chains, 75 Collision-avoidance, 157
Cascaded, 121 Color constancy, 326
CASTLE, 340, 341, 349 Color management, 326
Cat, 383, 389, 390 Commercial workloads, 220
Catch fire, 210 Common-mode, 239
Cathode, 13 Communication, 145
CCD. See Charge-coupled device (CCD) Companions, 153, 451
CDMA. See Code-division multiple access Compiler, 224
(CDMA) Complementary bipolar nanotransistors, 88
CE. See Continuing-education (CE) Complementary metal oxide semiconductor
CECC-EN 1001–14, 151 (CMOS), 17, 23, 60–67, 194
Cell, 223 active pixel, 320, 322
broadband engine, 226 compatible, 297–298
projection, 181 driver, 133
size, 230, 236 inverter, 3
Cellular neural network (CNN), 148, 332, 333, quad, 73
335–336, 339–363 6T SRAM, 230, 232
chips, 10 6T SRAM 2020, 235
processor, 333, 343, 355–361 Complex-instruction-set computing (CISC),
Certification, 453 147
policies, 452 Complexity, 76
procedures, 155 Compliance rules, 436
programs, 12 Computational requirements, 277
Chalcogenide, 246 Computed tomography (CT), 155
Channel engineering, 59 Computer on a silicon wafer, 20–21
Charge-coupled device (CCD), 247, 454 Computing, 145
Charges, 48 Conductance, 13, 15
Charge-transfer, 380 Constant contrast resolution, 323
460 Index
Consumer, 145 shot noise, 318, 324

market, 255 Dark limit, 318
Contact-less, 264 DARPA, 146, 389
Contacts (black), 45, 46 Database, 216
Contamination, 299 Data centers, 11, 422–426
Content, 449 Dataflow graph, 359
Content-addressable memories (CAMs), 148 dBm, 137
Context awareness, 275 DC–DC converter, 288
Continuing-education (CE), 151, 449 3D chip
Contract manufacturer, 431, 433 stacking, 422
Contrast resolution, 322, 323, 326 stacks, 436
Contrast-sensitivity function (CSF), 324, 325 3D CMOS, 69–74, 240
Control electrode, 15 SRAM, 73, 74
Control gate, 241 transistor pair, 23, 69, 71
Controlled-collapse chip connect, 79 DC standby current, 231
Convergence, 147, 252–255 DCVSL, 75
Conversion energy, 96, 99 3D display, 275
Cooling, 195, 222, 425 DEC, 64
Copper (Cu), 132 Deep UV, 31
pillars, 263 Defect count, 178
resistivity, 133 Delays, 133
sulfide, 15 Delay times, 133, 134
Co-processing, 252 Delta-sigma ADC, 116
Coriolis, 303–305 DS-modulators, 305
force, 301, 304 Demand/capacity, 430
Corners, 202 Dendrites, 384–387
Correct colors, 322 Dennard, R.H., 23
Cost/layer, 187 Density, 245
Cost of nanolithography, 423 Design
Cost of patterning tools, 183 automation, 61, 267–268
Coulomb know-how, 33
blockade, 247 productivity, 63
box, 247 starts, 197–198
confinement, 86, 454 Detectable signal, 318
Critical field, 49 Device-level 3D integration, 89
Cross-bar memory, 245 DFS. See Dynamic frequency scaling (DFS)
Crossbar metal-interconnect, 387 Diabetics, 155
Cross-coupled inverter pair, 5–6, 65 3D ICs, 282
Crosstalk, 171 Dielectric constant, 48
Cryptography, 220 Die stacking, 282
CSF. See Contrast-sensitivity function (CSF) Die-to-die, 251
Cube stack, 258, 259 DiffDRAM. See Differential DRAM
Cu. See Copper (Cu) (DiffDRAM)
Current Differential DRAM (DiffDRAM), 8, 239, 240
density, 85, 87 Differential input trees, 75
gain, 13 Differential logic, 74–78
Custom chips, 435–436 Differential sense amplifier, 240
Customization, 431, 436 Differential signaling/logic, 66
Diffusion constant, 42
Diffusion current, 53
D DIGILOG, 76, 352
DAC. See Digital-analog converter (DAC) Digital–analog converter (DAC), 288
Dark current, 318, 320, 321, 324 Digital correction, 114
Index 461
Digital light processing (DLP), 294, 296 3D vision, 328

Digital logic, 5 DVS. See Dynamic voltage scaling (DVS)
Digitally assisted architecture, 412 3D wafer-level integration, 329
Digitally assisted design, 105, 108 Dynamic energy, 5, 453
Digital neural networks, 383 Dynamic frequency scaling (DFS), 199
Digital number (DN), 324 Dynamic power, 192, 194, 199
Digital pattern generator (DPG), 183 Dynamic random-access memories (DRAMs),
Digital Rosetta stone, 246 64, 229, 237–240
3D imaging, 273, 275 Dynamic range, 316, 320–324, 326, 327,
Diminishing returns, 434 373, 412
DIN, 318 Dynamic-switching power, 67
3D integration, 24, 236, 244, 387, 388, 452 Dynamic voltage and frequency scaling
of memory, 245 (DVFS), 192, 199, 201
Direct electron-beam writing, 21 Dynamic voltage scaling (DVS), 199
Direct glucose fuel cell, 402–404, 410 Dystrophy, 368
Discretionary wiring, 21
Display power, 286
Display standards, 253 E
Disruptive, 363 Early voltage, 43, 87
Divide and conquer, 83 Ear plug, 451
DLP. See Digital light processing (DLP) E-books, 147
3D medical imaging, 155 ECL. See Emitter-coupled logic (ECL)
D-MIPS (million Dhrystone instructions Economic impact, 1
per second), 253 Economic issues, 4
DMIPS/MHz, 279 Ecosystem, 2, 11–12, 434
DN. See Digital number (DN) EDA. See Electronic design automation
Domino logic, 75 (EDA); Electronic design automation
Donor, 45 (EDA) tools
Doped crystal growth, 454 Edge detection, 361
Double-patterning, liquid-immersion Education, 439–449, 451, 452
lithography (DPT ), 176 Educational electronics, 448
3D packaging, 210 EEG. See Electroencephalography (EEG)
2.5D packaging technology, 256 Effective gate voltage, 50
3D-photodiode-on-CMOS, 322 Efficiency, 393, 394, 398, 404–408, 410
DPT. See Double-patterning, liquid-immersion Einstein relation, 42
lithography (DPT ) Electrical polarization, 246
Drain, 18 Electrical power, 5
as the seed, 69, 71 Electrical power of all of Europe, 424
current, 48 Electrical wafer sort (EWS), 268
induced barrier lowering, 243 Electric field, 13, 48
DRAMs. See Dynamic random-access Electric vehicles, 157
memories (DRAMs) Electro-cardiogram (ECG ), 289
DRIE process, 297 Electrodes, 13, 14
Driver fatigue, 157 Electroencephalography (EEG), 284
3D sensors, 327–328 Electromagnetic coupling, 395, 398, 399
3D stacking, 251 Electromagnetic interference (EMI), 281
3D strategies, 256–262 Electromagnetic transducer, 395, 396, 408, 409
3D technology, 362 Electromechanical system simulation, 299
3D torus interconnect, 225 Electron-beam lithography, 180
Dual-gate, 86 Electron-beam writing, 245
transistors, 69, 167 Electronic control, 157
3D vertical, 453 Electronic design automation (EDA), 267
DVFS. See Dynamic voltage and frequency Electronic design automation (EDA) tools, 251
scaling (DVFS) Electronic Eye, 79
462 Index
Electronic implants, 154 Epiretinal, 369, 370, 378–381

Electronic stability program (ESP), 157, Epitaxial growth, 81
300–308 Epitaxial lateral overgrowth (ELO), 68
Electronic system level (ESL), 204 Epitaxial-layer transfer (ELTRAN), 68
Electron mobilities, 34 Epitaxy, 68
Electrons, 453, 454 Equilibrium, 51
Electron-spin, 246 Equivalent circuit, 14
Electrostatic actuation, 301 Equivalent oxide thickness (EOT), 3, 165, 238
Electrostatic coupling, 395 ERC. See Engineering Research Centers
Electrostatic model, 53 (ERCs)
Electrostatic transducer, 395, 400, 401, 410 Error-detection, 221
ELO, See Epitaxial lateral overgrowth (ELO) Errors, 220–221
ELTRAN. See Epitaxial-layer transfer ES2, 25
(ELTRAN) ESL, 205. See Electronic system level (ESL)
Embedded memory, 238 ESP. See Electronic stability program (ESP)
EMC, 150 ESPRIT, 24, 25
Emerging Research Devices (ERD), 167–170 Ethics, 391
EMI. See Electromagnetic interference (EMI) EUREKA, 28
Emitter, 41 European 7th Programme, 447
Emitter-coupled logic (ECL), 41, 194 EUV. See Extreme-ultraviolet (EUV)
End of the Nano-roadmap, 2–5 Evolution, 169
Energy, 47, 66 EWS. See Electrical wafer sort (EWS)
autonomous sensor system, 400 Extended temperature range, 150
bit, 230 Extreme-technology, 429–437
consumption, 414, 416, 417 Extreme-ultraviolet (EUV), 4
crisis, 11, 421–427, 452 lithography, 30, 176
efficiency, 169, 175–188, 190, 274, 288, Eye tracking, 369, 379
446–448
efficiency from plug to wafer, 4–5
figure-of-merit (FOM), 2 F
gap, 286, 393 Fabless IC suppliers, 433
harvester, 399, 404 Fabless-manufacturer model, 430
harvesting, 274 Fabless model, 433–434
per chip-stack, 422 Face recognition, 275
per function, 1 Face-to-back, 265
per neuron, 386 Face-to-face, 265
per operation, 2 FACETS, 148, 387–390
per operation minimization, 414 Face-up stacking, 84
per synaptic operation, 389 Fail-safe, 152
per wafer-stack, 422 Failure analysis, 268–270
roadmap, 436 Failure in 107 hours, 151
sources, 6 Fairchild, 20
test, 150 Fan-out wafer-level package (FO-WLP), 257
E-newspapers, 147 Faster than the competitor, 434
Engineering Fatigue-warning, 157
cleverness, 166 Fault-tolerant architecture, 152
output, 443 FD SOI. See Fully depleted silicon-on-
Engineering Research Centers (ERCs), 25 insulator (FD SOI)
ENIAC, 447 FED, 26, 31
Environmental Protection Agency (EPA), 190 Femtoelectronics, 1–2, 434
EOT, 86, 168, 240 Femtojoule, 1, 451
EPA. See Environmental Protection Agency FEOL. See Front-end of line (FEOL)
(EPA) FeRAM. See Ferroelectric RAMs (FeRAM)
Index 463
Ferroelectric RAMs (FeRAM), 246 Fuzzy logic, 152

Field-effect triode, 15
Field-programmable analog array, 412
Field-programmable arrays, 433 G
Field-programmable gate array (FPGA), 198, 3G, 148
223, 335, 341, 343 Gain, 169
Field-programmable logic array (FPGA), 388 Gaming, 251
Figure-of-merit (FOM), 43, 288, 390 Ganglion cells, 368, 369, 379, 381
Fill-factor, 320–322 Gate arrays, 62, 63
Film thermoelectric layers, 401 Gate dielectric, 167
Financial crisis, 142 General Instruments, 19
FIPOS, 68 General-purpose microprocessors, 215
Firing rate, 387 General-purpose processors, 278
Firmware, 220 Generic neuromorphic peripheries, 390
First metal ½ pitch, 164 Gesture interface, 275
FIT, 150 Gesturing, 278
Fixed-dictionary compression, 220 GFLOPS, 277
Flash memory, 81, 150 3G/4G, 273
Flash RAM, 166 1 GHz barrier, 217
Flipped chips, 79, 80 Gigafactories, 180, 187, 436, 437
Floating gate, 240, 407, 412, 413 Giga-scale, 20
memory, 387 MOS nano-ROM, 245
Floating-point operations per second (FLOPS), Global policies, 452
146, 447 Global wire pitch, 132
FLOPS. See Floating-point operations per Google, 201
second (FLOPS) GOPS, 277
Flying capacitor arrays, 408 Government, 291
Focal plane coding and processing, 327–329 GPS, 138
Focal-plane processing, 150 location, 275
FOM. See Figure-of-merit (FOM) GPU. See Graphics processing units (GPUs)
Footprint, 23, 422 Gradient optimization, 335
Form factor, 254 Graphene, 133, 453
Foundry Graphics processing units (GPUs), 223
business model, 25 Graphics processor, 193, 221, 253, 257, 362, 447
model, 430–433 “Green” movement, 190
Fourier transform, 362 Grid, 13
Fowler–Nordheim tunneling, 407 Gross product, 142, 143
FO-WLP. See Fan-out wafer-level package Ground, 199
(FO-WLP) bounce, 71
FPGA. See Field-programmable gate array GSM, 148, 449
(FPGA) 800/1900, 138
Frequency race, 222 Gummel number, 43, 87
Front-end of line (FEOL), 170, 173, 264 Gummel–Poon transistor model, 43
Fuel cell, 394, 401–404, 410–411 Gyroscope, 298, 302
cascades, 402, 411
Full-bridge rectifier, 406
Full-swing regenerator, 5 H
Full-swing signal, 240 Hardware accelerators, 289
Fully depleted silicon-on-insulator (FD SOI), Hardware description languages (HDL), 204
167, 194 Hardware/software co-design, 225
Fundamental variance, 58, 64, 67 Hardware-software collaboration, 224–225
Future, 453 Harvest energy, 9
Future electron devices, 24 HD camera, 273
464 Index
HDL. See Hardware description languages IEEE Std 1801–2009, 208

(HDL) I2L, 41
HDTV, 149 Illuminance, 317, 318, 323, 324, 326
video cameras, 329 Image processing, 150
Head-mounted displays, 272 Image resolution of lenses, 320
Health, 452 Image sensors, 265
and care, 447, 448 Imaging, 8
Heart monitor, 451 IMEC, 28
Heart-monitoring, 448 Imitate biology, 335
Heat, 212 Immunity, 66
flow, 290 Impact ionization, 327
Heating, 425 Impedance-based converter design, 409
Heterogeneous process, 452–453 Impedance matching, 404, 408
Heterogeneous technologies, 251, 270 Implantation, 376, 377
HiDRaLoN, 327 of oxygen, 68
High-density 3D design, 269 Inductive, 371, 379
High dynamic range, 8 coupling, 135, 244, 246, 264
sensors, 322–327 spiral loops, 136
High-k, 167 transceivers, 136
High-level models, 191 Industrial, 152
High-level synthesis (HLS), 192, 204, 205 Inertial sensors, 294, 301, 305
High-performance processing, 276 Infinite retention time(s), 245
High-performance processors, 6 Inflection point, 195, 198, 423
High school, 439–442 for simple scaling, 317
High-speed memory, 66 Information and communication technology
High-tech, 429, 430 (ICT), 142, 156
High technology, 429 Information on single electrons, 3
High-value-added share, 453 Injuries, 448
Hilsch, R., 15 Inkjet nozzles, 294, 296
H implantation, 68 Innovation
HIPERLOGIC, 75, 76, 78 ecosystem, 446
Hitachi, 25 potential, 423
HLS. See High-level synthesis (HLS) Input impedance, 16, 404, 405, 408–410
Hoerni, J.A., 17 Input port, 14
Holter tape, 154 Insertion, 453
Hot chips, 423 Instruction-fetch/branch (IFB), 219
Hot-electron injection, 241 Instruction set, 356–357
Hotspots, 210 Instructions-per-cycle (IPC), 222
HP, 64 Insulated gate, 16
HP Labs, 34 FETs, 16
Human brain, 384–390 Integrated circuit, 16
Human-interaction, 153 3D-integrated process, 5
Humanoid, 152 3D integration, 2
Hybrid, 157 Intel, 195, 201
Hypoglycemia, 155 Intelligence, 384, 390
Intelligent memory, 387
Intelligent nanosystems, 452
I Intelligent vision system, 79
IBM, 81, 210 Interconnection fabrics, 280
ICT. See Information and communication Interconnects, 6, 131–140, 170–173
technology (ICT) technologies, 274
IEDM. See International Electron-Devices International electron-devices meeting
Meeting (IEDM) (IEDM), 444, 445
Index 465
International solid-state-circuits conference L

(ISSCC), 444, 445 Lane-keeping, 157
International Technology Roadmap for Language
Semiconductors (ITRS), 4, 23, 28, skills, 448
29, 47, 132, 161–173, 193, 204, understanding, 384
205, 230 Large-scale integration, 20
International working groups, 162 Laser recrystallization, 68
The Internet bubble, 142 Latch-up, 413
Internet connections, 216 Latency, 223, 274
Internet of Things, 275 LCA. See Life-cycle analysis (LCA)
Inter-neurons, 385–386 Lead sulfide, 15
Inter-processor communication, 278 Lead-time, 163, 224
Intestines, 154 Leakage, 53, 237
Intrinsic frequency limit, 42 current, 195, 286
Intrinsic gain, 67 Learn by plasticity, 383
Intrinsic voltage gain, 169 Learning, 334, 335
Inverter, 18, 19, 64, 135 Lego, 19, 20
pair, 135 blocks, 45
Investment in facilities, 430 Lens, 176, 177
Ion-beam doping, 454 Level shifters, 209
Ion-beam lithography (IBL), 30 Liability, 84, 157, 271, 436
Ionic charge, 386 Libraries, 147
Ion implantation, 60, 63 Life-cycle analysis (LCA), 422
IPC. See Instructions-per-cycle (IPC) Life-cycle assessment, 434, 436
ISO 9000:2000, 151 Life-cycle energy, 422, 423, 448
ISO 90001, 151 Light energy, 290
Isolation cells, 209 Light-section method, 328
ISSCC. See International solid-state-circuits Lilienfeld, J., 15
conference (ISSCC) Linear charge-integration, 320
ITRS. See International Technology Roadmap Linearity, 412, 413
for Semiconductors (ITRS) Linear strategy, 434
Line capacitances, 134
Line delays, 134
Lithium ion technology, 255
J Lithography, 304
JESSI. See Joint European Submicron Silicon Load adaptation, 404, 408
(JESSI) Load point, 405
JFETs. See Junction FETs (JFETs) Local memories, 236
Joint European Submicron Silicon LOCMOS, 63
(JESSI), 27 Logarithmic photosensor, 373
JND. See Just noticeable difference (JND) Logarithmic response to illuminance, 323
Junction FETs (JFETs), 16 Logarithmic sensor data, 327
Just noticeable difference (JND), 318 Long-line bandwidth, 426
Low-dropout voltage regulator, 410
Low-k dielectrics, 134
Low-voltage differential signaling (LVDS),
K 75, 136, 166, 169, 239
KGD. See Known-good die (KGD) Low voltage rectifier, 404
Kilby, 17 Low-voltage standby, 66
Kinetic energy transducer, 395 LPDDR2, 258
4k 2k, 149 LTE, 274
Knee current, 44 Luminance, 322, 326, 327
Known-good die (KGD), 84, 244, 271 LVDS. See Low-voltage differential signaling
Known-good tested chips, 210 (LVDS)
466 Index
M energy harvesters, 312

Machine–machine Interface, 284 everywhere, 312–313
Magnetic flux, 247 history, 295
Magnetic RAM (MRAM), 246 market, 295–297
Magnetic-tunneling junction (MTJ), 246 microphones, 297, 312
Main computer, 279 Micro-fuel cells, 294, 312
Mainframes, 222 Microlenses (MLs), 321
Mainstream, 423 Micro-machining, 295
Makers, 452–453 Micro-mirrors, 294
Managing energy, 279 Microprocessor(s), 21, 45, 193
Manchester carry chain, 75 Microsurgery, 154
Mandatory, 157 Milestones, 50
Man–machine interfaces, 279 Military, 145
Manpower, 11, 436, 437, 453 Millennium, 142
Manufacturer model, 430, 431 MIL-STD–883, 151
Manufacturing tolerances, 58 Miniature cameras, 328, 329
Many-core CNN architecture, 343–355 Minimalistic design, 105–108, 113
MAPPER, 181 Minimally invasive, 154
Market(s), 141–158, 430, 434–437 Minimum energy, 3, 5
forces, 452 consumption, 414
segments, 141 Minimum supply voltage, 3, 65, 234, 412,
shares, 436 414–416
Maskless, 245 MIPS, 279
electron-beam lithography, 387 MITI, 21
nano-patterning, 183 Mixed-signal, 63
Masks, 176 full-custom CNN, 337
Maturity, 272 MLC. See Multi-level-cell (MLC)
Maximum available current, 50 mmWave, 274
Maximum velocity, 168 Mobile
MCM. See Multi-chip module (MCM) communication, 63
MEB. See Multiple-electron-beam (MEB) companion(s), 9, 10, 273, 451
Mechanical vibration, 290 handset, 273
Media protection, 299 medical, 272
Media-rich mobile electronics, 75 phones, 145, 251
Medical, 152 Mobility, 42, 48, 133, 141, 441, 447, 448,
electronics, 153 452, 454
monitoring, 290 Modem link, 274
Megaprojects, 25 Molecular-beam epitaxy (MBE), 74
Memories, 6 Monocrystalline silicon, 294, 298
connectivity, 282 Moore, G., 20, 21, 221
density, 231 Moore’s law, 20, 229
energy, 8 More than Moore, 210
neurons, 386 effect, 331
Memristor, 247 MOS ICs, 41
MEMS. See Micro-electro-mechanical systems MOSIS, 25
(MEMS) MOSTEK, 20, 60
Metal gate, 167 Motor neurons, 386
Metal–1 half-pitch, 165 Motorola, 64
Metallurgical channel, 51 Mouse, 389, 390
Metric, 226 brain, 384
M1 half pitch, 132 Movement extraction, 362
Micro-electro-mechanical systems (MEMS), 8, MPEG, 362
181, 272, 293–313, 436 MPEG4, 277, 278
Index 467
MP3 players, 150 Neural controller, 27, 28

MTJ. See Magnetic-tunneling junction (MTJ) Neural-network event-recorder, 154
Multi-bit per cell, 247 Neural networks, 10, 26–30, 331–363
Multi-bit storage per transistor, 229 Neural signal processing, 79
Multi-chip, 210 Neurology, 154
Multi-chip module (MCM), 79, 217 Neuromorphic, 389–391
Multi-chip package, 135 computing, 387
Multi-core application processors, 279 machine, 389
Multi-core chips, 215 networks, 150
Multi-core processors, 6, 200–203 Neuron(s), 2, 10, 27, 28, 333, 334, 368,
Multi-layer feed-forward network, 334 384–390
Multi-level-cell (MLC), 242 Neutral base, 87
flash, 387 New media applications, 333, 361–363
NAND Flash density, 243, 244 Next-generation-lithography (NGL),
Multi-level CNN, 362 30, 176, 178
Multi-level flash, 8 Next-Generation-Mobile, 273–291
Multi-level implementation, 359–360 NGL. See Next-generation-lithography (NGL)
Multimedia-centric, 286 NHTSA. See National Highway Traffic Safety
Multimedia processing, 274 Administration (NHTSA)
Multi-mode design, 202 NMOS, 61
Multi-phase clocking, 247 select-transistor, 237
Multiple-electron-beam (MEB), 180–185 transistor, 17
direct write, 4 Nm radiation, 178
lithography, 176 Noise, 3, 13, 197, 211
Multiple exposure, 322 margin (NM), 65, 75, 234
Multiple input, 279 performance, 294, 305–307
Multiplier(s), 6, 75, 76, 338, 343, 349–351 shaping, 110, 121
mW/MIPS, 285 shaping techniques, 305
transfer function, 118
Non-invasive, 155
N Non-volatile memory, 241
NANDs, 70, 72 Non-volatile synapse weight, 387
array, 240, 241 Notebooks, 145
flash, 239, 240 Noyce, R.N., 17, 41
Nanocarbon, 133 NPN
Nano CMOS, 64 nanotransistor, 86
Nanoelectronics, 30–33 transistor, 41
Nanoimprinting, 178 N-type dopants, 45
Nanolithography, 4 Nucleus (soma), 384, 385, 387
Nano-roadmap, 3 Numerical aperture (NA), 176
Nanotransistor, 56, 59 NVidia, 205
National Highway Traffic Safety NV memories, 454
Administration (NHTSA), 300
Natural bit-rate compression, 323
Natural senses, 323 O
Near-field range, 136 OECF. See Optoelectronic conversion
Near-Infrared spectrometer, 155 function (OECF)
NEC, 60 Office automation, 251
Negative voltage converter, 407–408 Offset drift, 305, 307–308
NEMS, 183 On-Chip Interconnects, 132–135
Netbooks, 46 One-Laptop-per-Child (OLPC), 147
Network of facilities, 434 One transistor per bit, 237
Neumann, 332 ON/OFF current ratios, 57, 236
468 Index
OpenCL, 215, 223 Personal robots, 152, 452

Open SystemC Initiative (OSCI), 191 PetaFLOP-class, 447
Operating performance, 423–425 Petritz, R.L., 20
Operating system, 215 Phase-Change Memory (PCM), 246
Operations per second per milliwatt, 166 Phosphenes, 368, 376
Optical coherence tomography, 376 Photodetector, 136
Optical connections, 281 Photometric continuous sensor, 326
Optical interconnect, 136 Photometric flux density, 317
Optical lithography, 176 Photoreceptor, 367–369
Optical patterning, 176 Photosensitive resists, 176
Optic nerve, 380, 381 Physical chemistry, 386
Optoelectronic conversion function (OECF), Physiology, 154, 375–376
322, 324–326, 429 Piezoelectric, 290
Organic information processing, 454 transducer, 409
ORTC. See Overall roadmap technology Pinch-off, 50
characteristics (ORTC) Pipelined CNN, 339–342
OSAT. See Outsourced assembly and test Pixel(s), 338–342, 349, 356–358, 371,
(OSAT) 373–375, 379, 380
OSCI. See Open SystemC Initiative (OSCI) area, 316, 321
Output conductance, 56 operations, 332, 355–357, 359
Output port, 14 retina contact, 375–376
Output resistance Rout, 14 size, 316, 319, 321
Outsourced assembly and test (OSAT), 268 Placement and routing, 63
Overall roadmap technology characteristics Planar manufacturing process, 17, 18
(ORTC), 163 Planar structure, 17
Overlay tolerance, 171 Platform, 198
Overload-proof, 309 power, 191
Oversampling, 110 PMOS, 61
Oxide insulator, 17 transistor, 17
Oxide isolation, 41, 67 PN junction, 16
depletion layers, 51
Pohl, R.W., 15
P Policy makers, 453
Pacemaker, 154 Poly gate (red) and source/drain (yellow),
Package design, 191 45, 46
Package-on-package (POP), 257 Polysilicon, 45
Packaging, 210–211 POP. See Package-on-package (POP)
Parallel architectures, 274 Porous electrode, 375
Parallel computation, 280 Porous Si, 84
Parallelism, 215, 222, 383 Post-exponential growth era, 215
Parallel-plate capacitor, 47 Postsynaptic output, 386
Pattern-recognition, 384 Potassium bromide, 15
PCB. See Printed-circuit board (PCB) Potential barrier, 51, 241
PDN. See Power-distribution networks (PDNs) Power-aware, 192, 206–208
Perceptron, 11, 26–27 Power budgets, 195–197
Per-column analog/digital (A/D) converters, Power compiler®, 198
320 Power consumption, 273
Perkin-Elmer, 25 Power density, 85, 87, 223, 272
Permanent embedded ROM, 245 Power-distribution networks (PDNs), 211
Permanent magnets, 399 Power does not scale, 194–197
Permanent ROM (read-only memory), Power efficiency, 1, 7, 212, 276, 285–290, 387
245–246 of multipliers, 77, 79
Personal robot, 152 per bit, 234
Index 469
Power gain, 13, 43 Quantizer, 124–127

Power-gating, 192, 199 Quantum dot, 29
Power integrity, 192 Quantum efficiency, 317, 318
Power management, 207, 261, 394, 404–412 Quantum-nano devices, 1
techniques, 192
Power multiplexing, 405, 406
Power, Performance, and Area (PPA), 191 R
Power-saving, 201 RADC, 24
Power switches, 209 Radiation, 220
Power under test, 209 Radical, 10–11
Power wall, 332 innovations, 154
PPA. See Power, Performance, and Area (PPA) Radio (World War I), 15
Pressure sensors, 297, 298, 300, 308–310, 312 Radio-frequency (RF) transceiver, 261
Printed-circuit board (PCB), 135, 211, 431, Random-access memory (RAM), 45
435, 436 Range maps, 327, 328
design, 191 Ranking of engineering programs, 442
footprint, 256 Ratio design, 74
Process variabilities, 415 RCA, 19, 20
Production ramp-up, 162, 163 R&D. See Research and development (R&D)
Professional segments, 150–151 Read-only memory (ROM), 245–246
Programmability, 332, 338 Real-time, 335, 362
features, 216 Real-world environments, 389
Programming energies, 454 Recapitulation, 386
Programming models, 215 Receivers, 6, 131–140
Programming speed, 245 Recovering, 425
Programming voltage, 241 Recovery, 220–221
Program of implementation, 447 Recrystallize, 68
Projection, 275 recrystallized Si, 69
PROMETHEUS, 156 recrystallizing Si, 23
Prostheses, 367, 376 Rectification, 398, 404, 406–407
Prosthetics, 154 Redundancy, 82
P-type doping, 45 Reflectance, 326
Public interest, 155 Reflective-mirror optics, 178
Public research policy, 155 Reflectivity, 179
Public-safety interests, 157 Refocusing, 12, 452
Pulse-density-coded signal, 386 Refresh, 237
Punch-through current, 244 Regeneration amplifier, 61
P-well, 18 Regenerator, 5, 64, 453
Register transfer level (RTL), 192, 205
Reinforcement, 386
Q Reliability, 84, 197, 220–221, 272
QCIF, 277, 278 Rent’s rule, 20
QML. See Qualified manufacturing lines Repetition rate, 386
(QML) Requirements, 141–158
QMS. See Quality-management systems Requirements-driven research, 11, 446–449
(QMSs) Research, 439–449
QS 9000, 151 strategy, 449
Quad, 65 Research and development (R&D) quota,
Qualified manufacturing lines (QMLs), 24, 151 430, 436
Quality, 84 Reset, 320, 324, 326
chains, 151 Reshaping, 453
of life, 285 Resistive RAM (RRAM), 246
Quality-management systems (QMSs), 150 Resist sensitivity, 187
470 Index
Resolution, 8, 176, 177, 316, 319, 320, 322, Sapphire, 20

323, 326–328 Satellite-communication, 426
limit, 307 Saving lives, 448
Resolution-enhancement techniques (RETs), Scalability, 169, 245
176 Scaling, 22, 56, 395, 398, 405, 412–416
Resolution-scaling factor, 176 laws, 22, 43–45
Resonant vibration, 396 trends, 223–224
transducers, 396 Scan chain(s), 209
Resting potential, 386 Scenes, 384
Retention Schematic capture, 204
registers, 208, 209 Sclera, 371, 375, 377
time(s), 243 Seamless connectivity, 283–284
Reticle Sea-of-gates, 62
flatness, 179 Searchability, 216
life, 179 Secondary-electron detection, 181
magnification, 178 Security protocols, 285
parameters, 176, 177 SE device chain, 86, 454
Retina, 367–381 Seeding program, 156
Retinal implants, 9, 154, 367–381 Selected transistor, 240
Retinitis pigmentosa (RP), 367–369, 371, Selective, 454
379, 381 epitaxial lateral overgrowth, 69
Retinomorphic, 150 growth (self-assembly), 89
RETs. See Resolution-enhancement techniques Selectively doped Si crystal growth, 89
(RETs) Selector, 70
Re-write circuitry, 247 Self-sustained technology, 452
Re-writing, 237 Self-test, 294, 304, 305
RF. See Radio-frequency (RF) transceiver SEMATECH, 27, 178
RF energy, 290 Semiconductor, 15–16
RF power, 137, 138 market, 142–145
Rich media, 216–220 Semi-global wire pitch, 132
RISC, 147, 219 Sensing axes, 302
Roadmap, 1 Sensitivity, 8
Roadmap (1995), 27 standards ASA and DIN, 318
Road traffic, 156 Sensors, 6, 8
Robot electronics, 153 cluster, 300–308
Robot–human interaction, 153 neurons, 385
Robotics market, 153 Sequoia, 226
Robust ASIC, 293 Server, 137, 201
Rockwell, 21 farms, 201, 423
Rolling shutters, 320 SET. See Single-electron transistor (SET)
Room-temperature process, 63 Share content, 216
Roughness, 179 Sharing, 273
Router, 138 Sharp, 21
RTL. See Register transfer level (RTL) Sheet-beam illumination, 328
Shockley, 15, 49
Short-channel transistor, 54
S Short string, 238
Safe-distance, 448 Shot noise, 318, 324, 326
Safe personal mobility, 156 Shrink factor, 162
Safety and availability, 305 Si carrier, 79, 80
Safety applications, 294 Side-loading, 279
Safety-critical tasks, 151 Si dioxide, 45
Sah, C.T., 17 SiGe, 168
Index 471
Sigmoid transfer functions, 334, 335 Special-purpose processors, 215

Signal evaluation, 300, 304–305 Specialty engines, 223
Signal regenerator, 67 Spectral sensitivity, 317
Signal-to-noise ratio (SNR), 305, 323, 324, Spectrum, 138, 139
326, 412 Speed-up, 387
Signal transfer function, 118 Spike density, 386
Signal-transmission, 386 Spin, 8
Silicon brains, 11, 148, 383–391 of the electron, 454
Silicon compact disk, 150 Spin-torque-transfer (STT), 246
Silicon interposer, 262, 263 Spintronics, 454
Silicon-on-insulator (SOI), 67–69 Sputnik, 11, 439, 440, 443
Silicon open source, 433 SRAM. See Static random-access memory
SIMD FPU. See Single-Instruction (SRAM)
Multiple-Data Floating-Point Units SRC, 24
(SIMD FPUs) SSDs. See Solid-state drives (SSDs)
SIMOX, 68 Stacked DRAM, 82
Single electrons, 453 Stacked NMOS, 70
Single-electron transistor (SET), 3, 29, 57, 85, Stacked wafers, 32
169, 247 Stacking, 210
Single-Instruction Multiple-Data Floating- CMOS technology, 362
Point Units (SIMD FPUs), 226 devices, 136
Single-photon avalanche diode (SPAD), 327 of wafers, 81
Single photons, 327 transistors, 69
Sinker, 71 Stack wafers, 362
SiON, 241 Standard(s), 251, 436
Si p-n junction diode, 317 chips, 435, 436
Si retina, 26 definition, 329
Si tiles, 156 deviation, 58, 318, 327
Slew-rate, 240 Standardization, 272
Smart cameras, 361 Standard logic cells, 63
Smart cut, 68 Standards, 251
Smart-mobile-companion, 275 Standby, 67
(4G) Smart phones, 147 current, 67
Smart-power, 80 lifetimes, 199
SmartReflex, 287 power, 230
Smart-sensor, 80 Start-up, 305
SMP (symmetric multi-processing), 253 Static power, 193, 199
Sn droplet, 178 Static random-access memory (SRAM), 8, 64,
SNR. See Signal-to-noise ratio (SNR) 66, 88, 230–236, 289
Social interactions, 273 Stencil masks, 181
SOCs. See Systems-on-chip (SOCs) Stereo, 327–328
Soft-error rate, 221 cameras, 327
Software, 191, 223 Stimulation, 368–370, 373–377, 381
exploitation, 227 Strained Si, 168
SOI, 19, 45 Strategies, 452
Solid-state analogue, 15 String, 238
Solid-state drives (SSDs), 8, 150, 229 of NMOS transistors, 240
Sony, 223 Strong injection, 44
SOS, 68 Strontium bismuth tantalite (SBT), 238
Source, 17 STS. See Suprachoroidal transretinal (STS)
Source-follower transistor, 319 STT. See Spin-torque-transfer (STT)
SPAD. See Single-photon avalanche diode Sturdy delta-sigma ADC, 122
(SPAD) Subchoroidal, 371
472 Index
Subretinal, 369–373, 375–380 Terabit memories, 229–248

implant, 371, 372, 376–380 Terabit/s, 281
Subscribers, 141 Tera events/s, 387
Substrate, 18 TeraFLOPS, 146, 277
bias, 200 Terahertz, 45
Subthreshold, 195 Testability, 33
current, 53 Test flow, 270
design, 412 Test-generation, 208
leakage, 238 Testing, 208–210
Super capacitor, 404 for mechanical, 299
Supercomputers, 137, 145, 146, 215–228, 451 Test patterns, 209
Supercomputing, 362 Test strategies, 251, 436
Superprocessors, 215–228 Texas Instruments, 20
Supplant CMOS, 193 Thermal analysis, 211, 212
Supply chain, 268–271 Thermal-budget, 63
Supply voltage, 101–104, 199 Thermal integrity, 255
Suprachoroidal transretinal (STS), 371 Thermal issues, 210
Surface micromachining, 297–298, 301, Thermal limits, 280–282
302, 308 Thermal noise, 101–104
Surgery, 155 Thermal oxidation, 45
Surgical procedure, 371, 375–377 Thermal processing, 425
Surround gate structures, 86 The sixth Sense, 275
Surround-gate transistors, 167 Thin-film graphene, 453
Sustainable, 422, 425, 427 Thin wafer films, 422
expansion, 426 Third dimension, 23–26
nanoelectronics, 452 III-V materials, 169
Switching activity, 209 Thresholds overlap, 244
Switching energy CVDD2, 230 Threshold voltage, 50, 61, 222, 285
Switching speed, 47 Throughput, 178
Switching time, 42 computing, 215
SyNAPSE, 148, 389–390, 447 Through-silicon vias (TSVs), 32, 78, 210, 229,
Synapse, 2, 26, 27, 248, 384–390, 454 244, 251, 282
barriers, 384 TSV-first, 173
weight, 386, 387 TSV-last, 173
Synaptic gap, 386 TSV-middle, 173
Synaptic operations, 389, 390 TI, 64
Synchronous electric charge extraction, 409 Tigers, 27
Synchronous switch harvesting, 409 Time-encoding, 125–127
Synergistic Processing Elements, 223 Time-integral of voltage, 247
System drivers, 165–167 Time-interleaving, 110–111
System-level design, 191 Time-of-flight (TOF), 328
System requirements, 452 Time-to-volume, 251
Systems-on-chip (SOCs), 6, 63, 81, 256, Tire pressure, 312
285, 436 TLM 2.0, 191
TOF. See Time–of–flight (TOF)
Token for a memory bit, 454
T Tone mapping, 326
Tablet, 257 TOP500, 146, 226
PC, 147 Top–10 chip manufacturers, 432
Technology-cycle, 162, 163 Top-down R&D, 449
Technology scaling, 101–105, 285 Topography, 46
Technology transfer, 443 Torus, 156
Temperature range, 212 Toshiba, 223
1 Templates, 178 Total quality management (TQM), 436
Index 473
TQM. See Total quality management (TQM) Value chains, 271, 435, 436
Traffic technology, 156 Vapor deposition, 45
Training, 386 Variable shaped beam, 181
Transaction level modeling, 191 Variance, 5, 58, 165, 236, 453
Transceiver energy, 426 of the dark-current, 318
Transceivers, 433, 436 VDC. See Vehicle dynamics control (VDC)
Transconductance, 44, 49, 56, 58, 168 Vector processors, 226
Transconductor efficiency, 104 Vector scan, 181
Transfer characteristics, 64, 234, 235 Vehicle dynamics control (VDC), 300, 308
Transformer coil, 136 Velocity limit, 49
Transistor, 16 Velocity saturation, 49
column, 89 Verification, 192, 206
Transit time, 48 Verilog, 204
Transmitters, 6, 131–140 Vertical isolation, 20
Transport modes, 133 Very-large-scale integration (VLSI), 20
Trench etching, 304 VHDL, 25, 204
Trend lines, 164, 166 VHSIC, 24
Trimming, 305, 310 Via-first process, 264
TRUE and COMPLEMENT, 239 Via-last process, 265
TSMC, 25, 197, 205 Via-middle process, 264
TSVs. See Through-silicon vias (TSVs) Vias, 32, 83, 85
TTL, 41 Vibration, 394–401, 404
Tube, 13 harvester, 394
Turning points, 452 Video, 8
bandwidth, 447
codec, 280
U extraction, 362
Ubiquitous high-performance computing, 146, traffic, 426
447 Virtual retinal displays (VRDs), 272
UL. See Underwriters Laboratories (UL) Vision-in-the-loop, 362
ULP ICs, 290 Visual information, 316, 330
Ultimate cell area, 245 Vitreous body, 369, 376
Ultra-high-bandwidth 4G, 251 VLSI Lab, 22
Ultra-low-energy architectures, 425 Voice codecs, 323
Ultra-low-power, 147 Voltage-controlled oscillator, 124–125
sensors, 285 Voltage-divider, 237
Ultra-low-voltage, 288 Voltage gain, 3, 19, 43, 57, 58, 61
Ultra-thin-wafer handling, 269 Voltage scaling, 197
UMTS, 148 Voltage spikes, 220
UMTS/3G, 138 Volume wiring density, 137
Understanding, 332, 335, 361 Von Neumann, 26
Underwriters Laboratories (UL), 151 VRDs. See Virtual retinal displays (VRDs)
Unit-size, 61, 62, 75
Universities, 291
University programs, 440–444 W
UPF, 207 Wafer bonding, 68
Urbanization, 448 Wafer capacity, 430, 431
Users, 451 Wafer manufacturing, 422
Wafer-scale brain, 388
Wafer-scale integration, 387–389
V Wafers/month, 431
Vacuum cavity, 308–310 Wafer thinning, 78
Vacuum tubes, 13–15, 194 Wanlass, 17
474 Index
Warranty, 436 Workstation-on-the-go, 279

Water-cooling, 425 World economy, 142, 143
Wave computing, 361 Worst-case, 64, 65
Waveguides, 136
Weak inversion, 414
Web-based services, 426 X
Weight, 333–335, 338 X-ray, 176
Weighting, 333 lithography, 25
White saturation, 322, 326
Wideband vibration, 396
Wide-IO, 254 Y
DRAM, 257–260 Yaw rate, 296, 299–308
WiFi, 138, 274 Year 2020, 194
Wintel, 147 Yield, 20
Wire bonding, 256 loss, 269
Wireless energy scenarios, 425, 426 Y2K effect, 30
Wireless multimedia, 251–272
Wireless transceivers, 137–140
Wireless video, 138 Z
WLAN, 138 Zero-leakage standby, 289–290
Titles in this Series
Quantum Mechanics and Gravity

By Mendel Sachs
Quantum-Classical Correspondence
Dynamical Quantization and the Classical Limit
By Josef Bolitschek
Knowledge and the World: Challenges Beyond the Science Wars

Ed. by M. Carrier, J. Roggenhofer, G. Küppers and P. Blanchard
Quantum-Classical Analogies
By Daniela Dragoman and Mircea Dragoman
Life - As a Matter of Fat

The Emerging Science of Lipidomics
By Ole G. Mouritsen
Quo Vadis Quantum Mechanics?

Ed. by Avshalom C. Elitzur, Shahar Dolev and Nancy Kolenda
Information and Its Role in Nature

By Juan G. Roederer
Extreme Events in Nature and Society

Ed. by Sergio Albeverio, Volker Jentsch and Holger Kantz
The Thermodynamic Machinery of Life

By Michal Kurzynski
Weak Links
The Universal Key to the Stability of Networks and Complex Systems
By Csermely Peter
The Emerging Physics of Consciousness

Ed. by Jack A. Tuszynski
475
476 Titles in this Series
Quantum Mechanics at the Crossroads

New Perspectives from History, Philosophy and Physics
Ed. by James Evans and Alan S. Thorndike
Mind, Matter and the Implicate Order

By Paavo T. I. Pylkk€anen
Particle Metaphysics
A Critical Account of Subatomic Reality
By Brigitte Falkenburg
The Physical Basis of The Direction of Time

By H. Dieter Zeh
Asymmetry: The Foundation of Information

By Scott J. Muller
Decoherence and the Quantum-To-Classical Transition

By Maximilian A. Schlosshauer
The Nonlinear Universe

Chaos, Emergence, Life
By Alwyn C. Scott
Quantum Superposition
Counterintuitive Consequences of Coherence, Entanglement, and Interference
By Mark P. Silverman
Symmetry Rules
How Science and Nature Are Founded on Symmetry
By Joseph Rosen
Mind, Matter and Quantum Mechanics

By Henry P. Stapp
Entanglement, Information, and the Interpretation of Quantum Mechanics

By Gregg Jaeger
Relativity and the Nature of Spacetime

By Vesselin Petkov
The Biological Evolution of Religious Mind and Behavior

Ed. by Eckart Voland and Wulf Schiefenh€
ovel
Homo Novus - A Human Without Illusions

Ed. by Ulrich J. Frey, Charlotte St€
ormer and Kai P. Willf€
uhr
Brain-Computer Interfaces
Revolutionizing Human-Computer Interaction
Ed. by Bernhard Graimann, Brendan Allison and Gert Pfurtscheller
Titles in this Series 477
Extreme States of Matter

on Earth and in the Cosmos
By Vladimir E. Fortov
Searching for Extraterrestrial Intelligence

SETI Past, Present, and Future
Ed. by H. Paul Shuch
Essential Building Blocks of Human Nature

Ed. by Ulrich J. Frey, Charlotte St€
ormer and Kai P. Willf€
uhr
Mindful Universe
Quantum Mechanics and the Participating Observer
By Henry P. Stapp
Principles of Evolution
From the Planck Epoch to Complex Multicellular Life
Ed. by Hildegard Meyer-Ortmanns and Stefan Thurner
The Second Law of Economics

Energy, Entropy, and the Origins of Wealth
By Reiner K€ummel
States of Consciousness
Experimental Insights into Meditation, Waking, Sleep and Dreams
Ed. by Dean Cvetkovic and Irena Cosic
Elegance and Enigma

The Quantum Interviews
Ed. by Maximilian Schlosshauer
Humans on Earth
From Origins to Possible Futures
By Filipe Duarte Santos
Evolution 2.0
Implications of Darwinism in Philosophy and the Social and Natural Sciences
Ed. by Martin Brinkworth and Friedel Weinert
Probability in Physics
Ed. by Yemima Ben-Menahem and Meir Hemmo
Chips 2020
A Guide to the Future of Nanoelectronics
Ed. by Bernd Hoefflinger
From the Web to the Grid and Beyond

Computing Paradigms Driven by High-Energy Physics
Ed. by René Brun, Federico Carminati and Giuliana Galli Carminati

Chips 2020

Uploaded by

Copyright:

Available Formats

Chips 2020

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chips 2020

Uploaded by

Copyright:

Available Formats

What is the purpose of the Frontiers Collection?

What is the purpose of the Frontiers Collection?

What topics are covered in the Frontiers Collection?

What topics are covered in the Frontiers Collection?

THE FRONTIERS COLLECTION

THE FRONTIERS COLLECTION

Series home page – springer.com

# Springer-Verlag Berlin Heidelberg 2012

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

1 Introduction: Towards Sustainable 2020 Nanoelectronics . . . . . . . . . . . . 1

2 From Microelectronics to Nanoelectronics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3 The Future of Eight Chip Technologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

5 Interconnects, Transmitters, and Receivers . . . . . . . . . . . . . . . . . . . . . . . . . . 131

6 Requirements and Markets for Nanoelectronics . . . . . . . . . . . . . . . . . . . . . 141

7 ITRS: The International Technology Roadmap

9 Power-Efficient Design Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

10 Superprocessors and Supercomputers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215

11 Towards Terabit Memories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

12 3D Integration for Wireless Multimedia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251

13 Technology for the Next-Generation-Mobile User Experience . . . . . . 273

14 MEMS (Micro-Electro-Mechanical Systems) for Automotive

15 Vision Sensors and Cameras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315

16 Digital Neural Networks for New Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331

17 Retinal Implants for Blind Patients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367

18 Silicon Brains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383

19 Energy Harvesting and Chip Autonomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393

20 The Energy Crisis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421

21 The Extreme-Technology Industry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429

22 Education and Research for the Age of Nanoelectronics . . . . . . . . . . . . 439

23 2020 World with Chips . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451

Titles in this Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479

Greg Delagi Texas Instruments, 12,500 TI Boulevard, MS 8723, Dallas, TX,

Dr. Udo-Martin Gómez Robert Bosch GmbH, Automotive Electronics (AE/

Thorsten Hehn IMTEK Microelectronics, University of Freiburg, Georges-

Prof. Bernd Hoefflinger Leonberger Strasse 5, 71063 Sindelfingen, Germany,

Dr. Daniel Hoffmann HSG-IMIT, Institute of Micromachining and Information

Dr. Matthias Keller IMTEK Microelectronics, University of Freiburg, Georges-

Matthias Kuhl IMTEK Microelectronics, University of Freiburg, Georges-

Niklas Lotze HSG-IMIT, Institute of Micromachining and Information Technology,

Suleiman Malki Department of Electrical & Information Technology,

Prof. Yiannos Manoli Fritz Huettinger Chair of Microelectronics, University

Dominic Maurath IMTEK Microelectronics, University of Freiburg, Georges-

Christian Moranz IMTEK Microelectronics, University of Freiburg, Georges-

Daniel Rossbach HSG-IMIT, Institute of Micromachining and Information

Prof. Albrecht Rothermel University of Ulm, Institute of Microelectronics,

Prof. Lambert Spaanenburg Dept. of Electrical & Information Technology, Lund

Dirk Spreemann HSG-IMIT, Institute of Micromachining and Information Tech-

Bernd Hoefflinger started his career as an assistant professor in the Department of

Matthias Keller was born in Saarlouis, Germany, in 1976. He received his

Georg Kimmich is product manager at ST-Ericsson in Grenoble, France. He is

Dr. Lin is the editor-in-chief of the Journal of Micro/nanolithography, MEMS, and

Yiannos Manoli was born in Famagusta, Cyprus. As a Fulbright scholar, he

Boris Murmann is an Associate Professor in the Department of Electrical

Barry Pangrle is a Solutions Architect for Low Power in the Engineered

architecture and VLSI design while performing research in high-level design

Albrecht Rothermel received his Dipl.-Ing. degree in electrical engineering

Lambert Spaanenburg received his master’s degree in electrical engineering

Kai Weber received his Dipl.-Ing.(BA) in information technology from the

AAEE American Association for Engineering Education