(Esther Villegas Low Power and Low Volt

IET Circuits, Devices and Systems Series 20
Low Power and Low

Voltage Circuit Design with
the FGMOS Transistor
Esther Rodriguez-Villegas
IET CIRCUITS, DEVICES AND SYSTEMS SERIES 20
Low Power and Low Voltage

Circuit Design with the
FGMOS Transistor
Other volumes in the Circuits, Devices and Systems series:
Volume 1 GaAs technology and its impact on circuits and systems D. G. Haigh and
J. Everard (Editors)
Volume 2 Analogue IC design: the current-mode approach C. Toumazou, F. J. Lidgey
and D. G. Haigh (Editors)
Volume 3 Analogue-digital ASICs R. S. Soin, F. Maloberti and J. Franca (Editors)
Volume 4 Algorithmic and knowledge-based CAD for VLSI G. E. Taylor and G. Russell
(Editors)
Volume 5 Switched-currents: an analogue technique for digital technology
C. Toumazou, J. B. Hughes and N. C. Battersby (Editors)
Volume 6 High frequency circuits F. Nibier and co-authors
Volume 7 MMIC design I. D. Robertson (Editor)
Volume 8 Low-power HF microelectronics G. A. S. Machado (Editor)
Volume 9 VLSI testing: digital and mixed analogue/digital techniques S. L. Hurst
Volume 10 Distributed feedback semiconductor lasers J. Carroll, J. Whiteaway and
R. Plumb
Volume 11 Selected topics in advanced solid state and fibre optic sensors
S. M. Vaezi-Nejad (Editor)
Volume 12 Strained silicon heterostructures: materials and devices C. K. Maiti,
N. B. Chakrabarti and S. K. Ray (Editors)
Volume 13 RFIC and MMIC design and technology I. D. Robertson and S. Lucyszyn
(Editors)
Volume 14 Design of high frequency Integrated analogue filters Y. Sun (Editor)
Volume 16 Wireless communication circuits and systems Y. Sun (Editor)
Volume 17 The switching function: analysis of power electronic circuits C. Marouchos
Volume 18 System-on-chip: next generation electronics B. M. Al-Hashimi (Editor)
Low Power and Low Voltage
Circuit Design with the
FGMOS Transistor
Dr Esther Rodriguez-Villegas
The Institution of Engineering and Technology

Published by The Institution of Engineering and Technology, London,
United Kingdom
© 2006 The Institution of Engineering and Technology
First Published 2006
This publication is copyright under the Berne Convention and the Universal
Copyright Convention. All rights reserved. Apart from any fair dealing for the
purposes of research or private study, or criticism or review, as permitted under the
Copyright, Designs and Patents Act, 1988, this publication may be reproduced,
stored or transmitted, in any forms or by any means, only with the prior permission
in writing of the publishers, or in the case of reprographic reproduction in
accordance with the terms of licences issued by the Copyright Licensing Agency.
Inquiries concerning reproduction outside those terms should be sent to the
publishers at the undermentioned address:
The Institution of Engineering and Technology,
Michael Faraday House,
Six Hills Way, Stevenage,
Herts, SG1 2AY, United Kingdom
www.theiet.org
While the author and the publishers believe that the information and
guidance given in this work are correct, all parties must rely upon their own
skill and judgment when making use of them. Neither the author nor the
publishers assume any liability to anyone for any loss or damage caused by
any error or omission in the work, whether such error or omission is the result
of negligence or any other cause. Any and all such liability is disclaimed.
The moral rights of the author to be identified as author of this work have been
asserted by him in accordance with the Copyright, Designs and Patents Act 1988.
British Library Cataloguing in Publication Data

Rodriguez-Villegas, Esther
Low power and low voltage circuit design with the FGMOS
transistor. — (IEE circuits, devices and systems series)
1. Power transistors 2. Low voltage integrated circuits
I. Title II. Institution of Engineering and Technology
621.3’81528
ISBN (10 digit) 0 86341 617 9

ISBN (13 digit) 978-086341-617-0
Typeset in India by Newgen Imaging Systems (P) Ltd, Chennai, India

Printed in the UK by MPG Books Ltd, Bodmin, Cornwall
For my parents:
Jesus and Maria Josefa
Contents
Preface xi
Acknowledgments xiii
1 Introduction 1
1.1 The general framework 1
1.1.1 Why analog? Why digital? 1
1.1.2 Why low voltage? 1
1.1.3 Why low power? 2
1.1.4 Why CMOS? 2
1.2 Techniques to reduce the power consumption and the
voltage supply 2
1.2.1 Analog techniques 2
1.2.2 Digital techniques 4
1.3 Why floating gate MOS? 6
1.4 FGMOS history 7
1.5 Structure of the book 11
2 The Floating Gate MOS transistor (FGMOS) 15

2.1 Introduction 15
2.2 The Floating Gate MOS (FGMOS) device 15
2.2.1 Introducing the device 15
2.2.2 Theory 18
2.3 Designing with FGMOS: problems and solutions 22
2.3.1 Simulation 22
2.3.2 Charge accumulation 28
2.4 Minimum input capacitance: implications to the total area of
the device 35
2.5 Summary and conclusions 40
Notation 40
viii Contents
3 FGMOS – Circuit applications and design techniques 43

3.1 Introduction 43
3.2 Initial design ideas: three circuit equivalents for the FGMOS 43
3.2.1 FGMOS: a MOS transistor plus an adder 43
3.2.2 FGMOS: a MOS transistor with a controllable
threshold voltage 44
3.2.3 FGMOS in weak inversion: a current multiplier 45
3.3 Circuits applications and design techniques 46
3.3.1 A cascode current mirror 46
3.3.2 Two-stage OPAMP 50
3.3.3 FGMOS programmable inverter 58
3.3.4 An offset compensated FGMOS comparator 60
3.3.5 FGMOS D/A converters 66
3.3.6 A programmable switched-current floating gate
MOS cell 73
Notation 81
4 Low power analog continuous-time filtering based on the FGMOS

in the strong inversion ohmic region 87
4.1 Introduction 87
4.2 Transconductor basic blocks 89
4.3 The Gm -C multiple-input linear integrator. Large signal
description 91
4.3.1 Operating point – design considerations 92
4.4 The Common Mode Feedback Block (CMFB) 95
4.5 Small signal considerations 96
4.5.1 The transconductance cell (N1) 97
4.5.2 The tuning amplifier (N2) 99
4.5.3 Common mode DC response 101
4.6 Second order effects 103
4.6.1 Total harmonic distortion 103
4.7 Power Supply Rejection Ratio (PSRR) 111
4.8 Filter example 115
Notation 123
5 Low power analog continuous-time filtering based on the FGMOS

in the strong inversion saturation region 129
5.1 Introduction 129
5.2 A Gm -C integrator based on the FGMOS in the strong inversion
saturation region for audio applications 130
5.2.1 Design trade-offs. Power consumption, voltage supply
and limits for the transconductance 132
Contents ix
5.2.2 The Common-Mode Feedback Circuit (CMFB): effects

on the OTA performance 135
5.2.3 Second-order effects 143
5.2.4 Power supply rejection ratio 148
5.2.5 A filter example 152
5.3 An intermediate frequency FGMOS-based filter 157
5.3.1 The FGMOS-based IF transconductor 158
5.3.2 The FGMOS cascode load and the CMFB 159
5.3.3 Dynamic range 162
5.3.4 Rejection to supply noise 163
5.3.5 A second-order filter prototype 165
5.4 A second-order FGMOS filter in a single poly technology 168
5.4.1 Effects of the metal–poly capacitors 168
5.4.2 Experimental results 173
5.5 Comparison 175
Notation 177
6 Low power analog continuous-time Gm -C filtering using the

FGMOS in the weak inversion region 185
6.1 FGMOS translinear principle 186
6.2 A Gm -C FGMOS integrator using translinear circuits 188
6.2.1 The state-space integrator equations 188
6.2.2 FGMOS blocks 189
6.2.3 Second-order effects 193
6.2.4 A second-order filter example 202
Notation 211
7 Low power log-domain filtering based on the FGMOS transistor 215

7.1 Log-domain integration with FGMOS 216
7.1.1 Basic principle 216
7.1.2 Basic FGMOS circuits 217
7.1.3 The integrator 219
7.1.4 Advantages and disadvantages of using FGMOS
transistors 221
7.2 The second-order filter 222
7.2.1 The expander 224
7.2.2 The input stage 225
7.3 The common mode control 225
7.4 Design considerations. Second-order effects 226
7.4.1 Multiple operating points in the filter 226
7.4.2 Effect of the parasitic capacitances: qualitative study 228
7.4.3 Effect of CGD and the mismatch between the inputs:
a quantitative study 229
x Contents
7.5 A design example 235

Notation 241
8 Low power digital design based on the FGMOS threshold gate 245
8.1 Introduction 245
8.2 Threshold gates 246
8.3 The νMOS threshold gate 247
8.3.1 Theory 247
8.3.2 Practical design aspects 249
8.4 νMOS Threshold gates applications 250
8.4.1 Threshold logic based adders using floating-gate
circuits 250
8.4.2 νMOS-based compressor designs 255
8.4.3 Sorting networks implemented as νMOS circuits 258
8.4.4 A multi-input Muller C-element 267
Notation 269
9 Summary and conclusions 273
References 279
Index 297
Preface
The implantation of electronic devices into biological systems is an exciting new

development in medical science. Implantable prosthesis has the potential to provide
medical practitioners with a range of new diagnostic tools and treatment options.
These devices must be small and lightweight, which means their batteries must be
small in number and be long lasting. From the circuit design point of view this means
low voltage and low power (LV/LP). At the same time consumers are demanding
greater portability and longer battery life from their handheld and laptop devices.
Again, to the circuit designer this means LV/LP.
Partly due to the demand for smaller and faster products, there is an ongoing trend
in fabrication processes towards smaller transistors. Future deep sub-micron devices
will not be able to withstand high operating voltages, so circuit design needs to reflect
this fact. The SIA predicts that in 2010 the maximum supply voltage will be 0.8 V1 .
The current twin aims of circuit design research are low voltage and low power. The
former is required to meet the requirements of sub-micron fabrication processes and
the latter is required to meet battery longevity requirements. In digital circuits power
consumption is proportional to VDD 2 , where V
DD is the supply voltage. Most current
and future designs are likely to use the system-on-a-chip (SOC) design style. SOC
integration means that analog and digital circuits must use the same process and share
a common substrate. Whilst the performance of digital circuits is relatively robust to
transistor shrinkage, the performance of analog circuits generally degrades because
making the transistors smaller necessitates a reduction in VDD .
The drive towards very low voltage analog circuits did not gather momentum until
the mid-nineties. As a consequence, the design techniques are not well developed.
One of the most promising strategies in LV/LP analog circuit design is the floating
gate MOS (FGMOS) transistor. These devices can be fabricated in all CMOS tech-
nologies, although a double poly CMOS technology is preferred. FGMOS allows the
designer to do any or all of the following: reduce the complexity of circuits; simplify
the signal processing chain within a design; shift the signal levels and incorporate
1 The National Technology Roadmap for Semiconductors. 2002 Update. Semiconductor Industry
Association, 2003.
xii Preface
tuneable mechanisms. These new possibilities make it feasible to design circuits with
functionality that would not be possible as supply voltages get lower in comparison
to the threshold voltages of specific technologies.
This book is intended for a broad range of readers. Although the title seems to
suggest that it is a book on the FGMOS and therefore only those interested in explor-
ing this area can benefit from it, the reality is different. This is fundamentally a book
on circuit design pushing the limits of CMOS technologies. As the power and volt-
age constraints become more and more restrictive the performance of conventional
circuit topologies degrades up to the point that they stop being functional. The trade-
offs need to be established much more exhaustively and “rules of thumb” no longer
help. In some cases alternative circuit techniques have to be found. This book guides
the reader through a thorough analysis of design tradeoffs for every circuit topology
presented in it together with a quantitative and qualitative study of what the conse-
quences of shrinking the voltage levels and reducing power are. From this point of
view, it is a book that can appeal to any audience with an interest in low power and
low voltage circuit design.
But also, it is a book on the FGMOS transistor and its use in analog and digital
circuits. The text presents the device from scratch and builds up from there so no
previous knowledge is needed. Hence, a designer with previous knowledge in the
subject can probably skip parts of chapters two and three. As Section 1.5 explains
the book is mostly focused on analog because it is in this context that the FGMOS
transistor has more potential. Nevertheless some very interesting applications can also
be found in the digital area and this is why the last chapter is on low power and low
voltage digital design. Unlike the analog chapters which are much more incremental,
this chapter requires a previous basic digital knowledge.
Acknowledgments
Part of the material presented in this book is the result of the research carried out in the
Seville Institute of Microelectronics (IMSE-CNM) and was sponsored by the Spanish
government under an FPU grant. This sponsorship is gratefully acknowledged.
I am also thankful to many of the people working there at the time specially to
my then Ph.D. supervisors Dr. Alberto Yufera Garcia and Professor Adoracion Rueda
Rueda. The material presented in chapter number 8 was gathered with the help of
Dr. Jose Maria Quintana and Dr. Maria Jose Avedillo who set up the simulation
environments for the traditional implementation of the digital circuits.
Apart from those who supported me on the technical and financial side, there
were many others who played an even more important role, lifting up my mood on
many occasions I just wanted to give up electronics and become a “chocolate critic”.
Thank you to my “friends in the Himalaya”: Kike, Carlos, Antonio, Pedro, Hakim
and specially Africa (thank you for being such a wonderful friend through all these
years). Also to many others in the CNM, mostly Eduardo and Emi. Thank you to my
“friends in Circuits and Systems at Imperial”: Ganesh, Anjit, Solon, Thomas, Rohit,
David and Omeni (I miss you now that I do not see you everyday). And those in
“Signal Processing”: Jon, Thushara, Alex, David and Hugh. Thank you to Fernando
for putting up with me and the FGMOS when no one believed my “getting rid of the
charge” technique was going to work.
Also, thank you to my colleagues at Imperial College especially to Professor Peter
Cheung, Dr. David Haigh and Dr. Christos Pappavassiliou.
Thank you to Dr. Ramon Gonzalez-Carvajal for believing in me, helping me
every time I asked for help (and many other times he could read my mind), making
me laugh and showing me the best tapas bars in Seville.
Very special thank you to Mr. Phil Corbishley for being such a good student,
friend and for reading throughout the whole text.
Thank you to Dr. Patrick Degenaar since without him this book would not have
been completed.
xiv Acknowledgments
Thank you to Dr. Aleksandra Rankov for helping me with the edition.
Thank you to Mr. Vasa Curcin for being so sweet, being always there and dragging
me out for coffee when I needed it the most.
Thank you to all the people in the IEE who have worked together on publishing
this book.
And finally, thank you to my family (specially my sisters, Maria and Bea), who
lifted me in very difficult moments.
Chapter 1
Introduction
1.1 The general framework
1.1.1 Why analog? Why digital?

In recent years, digital signal processing has progressively supplanted analog signal
processing in chip design. This is due to it having lower development costs, better
precision performance and dynamic range, as well as being easier to test. The role
of analog circuits has been mostly restricted to electronic applications of interfac-
ing digital systems to the external world. Nevertheless, when precise computation of
numbers is not required (as is the case in systems designed for perception of a continu-
ously changing environment), and massively parallel collective processing of signals
is needed, low precision analog VLSI (very large scale integration) has proven to be
more convenient than digital in terms of cost, size and/or power consumption [1].
In recent years mixed signal application-specific integrated circuits (ASICs)
have become increasingly popular. The cooperative coexistence of analog and dig-
ital circuits is very beneficial since they compensate for each other’s weaknesses.
Hence, although in many aspects digital electronics is superior, in reality it requires
a symbiotic relationship with analog.
1.1.2 Why low voltage?

Since the invention of the transistor more than 50 years ago, the progress of micro-
electronics can be summarised as follows: 15 per cent decrease in feature size per
year, 30 per cent cost decrease per year, 50 per cent performance improvement and
15 per cent semiconductor market growth rate. The numbers speak for themselves.
This exponential evolution made many experts in the 1990s assert that fundamental
limits were about to be reached. Fortunately, technical innovations made it possible to
shrink the technologies to smaller dimensions than the predicted 0.3 μm. However, as
the dimensions of the devices reduce, a new constraint arises: the interconnect delays
and the fact that they directly affect the CV 2 power dissipation. In the past, this was
2 Low power low voltage circuit design with the FGMOS transistor
not a problem as the capacitances were scaled down together with the dimensions.
Recently this scaling relationship has been replaced to being proportional to the total
length of wires, L, in the circuit. The interconnect power dissipation can therefore be
rewritten as kV 2 L (where k is the dielectric permittivity). Hence, the most significant
parameter in the reduction of the interconnect power is the voltage and new strategies
are required to operate circuits at lower power supply voltages [2].
However, this is not the only motivation fuelling the eagerness of researchers
to operate circuits at lower voltages. The other one is related to the magnitude of
the electric fields in the devices. These grow proportionally as the dimensions are
scaled down, which increases the risk of dielectric breakdown. This can additionally
be compensated for by reducing the voltage differences across the devices. Hence a
low voltage power supply is beneficial.
1.1.3 Why low power?

The fast development of electronic-based entertainment, computing and communi-
cation tools, especially portable ones, has provided a strong technology drive for
microelectronics during the last ten years. System portability usually requires battery
supply and therefore weight/energy storage considerations. Unfortunately, battery
technologies do not evolve as fast as the applications demand. Therefore the chal-
lenge, derived from market requirements, is to reduce the power consumption of the
circuits.
In addition to consumer products, battery lifetime is a crucial factor in some
biomedical products which have to be either worn or implanted within the patients
for a long period of time; such systems are continuously increasing in number and in
scope. Investigation into low power biomedical systems is another interesting quest
for microelectronic designers [3–6].
1.1.4 Why CMOS?

The choice of fully integrated VLSI complementary metal oxide semiconductor
(CMOS) implementations is based on their lower cost (whenever the state-of-the-
art sub-micron VLSI processes are not required) and design portability. Furthermore,
CMOS technologies allow the possibility of integration with micro electro mechan-
ical systems (MEMS) [7]. These are the most important reasons that have directed
the semiconductor industry towards CMOS mixed signal designs, and place CMOS
technologies as the leader in the microelectronics semiconductor industry [8].
1.2 Techniques to reduce the power consumption and the

voltage supply
1.2.1 Analog techniques

Different techniques have been developed to reduce power consumption and supply
voltage in analog circuits. They can be classified into two main groups: techniques
Introduction 3
intended to reduce the voltage levels and techniques intended to reduce the current
levels.
1.2.1.1 Techniques for voltage reduction

The most popular techniques for voltage reduction are as follows:
1. Circuits with a rail-to-rail operating range: This group includes all the
techniques that are meant to extend the voltage ranges of the signals. Most of
these techniques are based on redesigning the input and output stages in order to
increase their linear range [9–13]. In these topologies the transistors have to be biased
in those regions that optimise the operating range. Since the voltage constraints are
more restrictive, in order to get the devices working in a certain region, sometimes
it is necessary to shift the voltage levels. In this book the reader will find multiple
examples of how the floating gate MOS (FGMOS) can be used to shift the voltage
levels in a very simple way, so the operating range is optimised.
2. Technique of cascading stages, instead of a single cascode stage: Conven-
tional circuit topologies stacked cascode transistors in order to obtain the high output
resistances and gains required by certain structures such as OPAMPs and OTAs.
However, stacking transistors in a given circuit branch makes the voltage require-
ments more demanding for the entire cell. The solution for this is either to reduce
the voltage requirements for the transistors, or to substitute each single stage for a
cascade of them, in such a way that the total gain at the output would be the product
of all the single ones (whenever high gain is needed). If the latter route is taken,
there is an added problem related to frequency stabilisation [14–16]. Furthermore,
as the number of branches increases, so too does the power consumption. Therefore
a compromise solution would be to still use cascode transistors, but try to minimise
their voltage requirements within the whole operating range. This book will show
how to obtain this by using cascode devices implemented with FGMOS transistors.
Different circuits will be presented, that, even with a single stage, achieve a high
performance in terms of gain under a very low power supply voltage.
3. Supply multipliers: Charge pumps can be used to scale up the power supply
voltage for certain analog cells, while still keeping the low supply voltage value for
the digital blocks [17–20]. The main drawback of this technique is that it requires
large capacitors which take a large silicon area, a considerable overhead for circuits
using this technique. In addition, the extra power consumption can be considerable.
4. Nonlinear processing of the signals: Most practical electronic systems are
designed to process signals in a linear form. However, the fact that the input/output
relationship between two variables has to be linear does not mean that internally the
system must be linear as well [21]. The fundamental devices constituting the blocks
are transistors which are inherently nonlinear. Traditional circuit techniques tried to
linearise the behavioural laws of the devices with more or less complicated topologic
solutions that in most cases unavoidably increased the power consumption. The idea
behind nonlinear techniques is to exploit the nonlinear I /V characteristics of the
transistors to process the signals more efficiently. A very powerful method to obtain
this consists of processing the input signal in such a way that for low input values
the internal signal levels are still above the maximum expected noise, whereas large
input signals are attenuated, hence reducing distortion. Ideally, a function would be
chosen that shrinks the internal dynamic range (DR) down to 0 dB and once the signal
is internally processed, the inverse is applied to recover the original DR expansion.
This would keep the DR, as well as (a) reduce the large voltage swings around the
bias point, hence relaxing the voltage constraints imposed by the lower power supply
voltage and (b) reduce the amount of power needed to charge and discharge the
parasitic capacitances.
The ‘compressing/expanding’ processing technique at the device level is known
as instantaneous companding. Other techniques at the system level are for example
syllabic companding [22–24] and time-variant compressing techniques for AGC [25].
The majority of the currently existing companding circuits base their functionality
on the nonlinear current law of the bipolar transistor, and they are not compatible with
the low power and low voltage constraints since they were originally developed to
exploit the device’s high frequency capability. Unfortunately these techniques are
not easy to translate to a MOS implementation, because of several reasons such as
asymmetric I /V curves in MOS devices and mismatch. On top of that, most of the
existing CMOS realisations cannot be generalised due to, for example, local bulks,
poor low voltage operation or redundant circuitry [26–29].
In this book, the reader will find how to overcome some of the problems inherent to
MOS realisations by using FGMOS transistors to realise externally linear, internally
nonlinear circuits. Besides, complex functionality can be implemented with very
compact designs which is also beneficial from the point of view of power.
1.2.1.2 Techniques for current reduction

The main circuit design techniques oriented towards reduction of current are
1. Adaptive biasing: This technique is based on using a non-static current bias to
optimise the power consumption according to signal demands [30,31].
2. Subthreshold biasing: Another way to reduce the current levels and hence the
power is by using MOS transistors biased in the weak inversion region driving
very low current levels. Some sections of the book will show how by employing
FGMOS transistors instead an extra power reduction can be achieved, thanks to
its better suitability to implementation of very complex functions [32].
1.2.2 Digital techniques

Power dissipation in CMOS digital circuits can be classified into static power con-
sumption, which is the consequence of the resistive paths from the power supply to
ground; and dynamic power consumption, which is due to the switching of capacitive
loads between two different voltage states.
Until relatively recently, digital designers optimised their designs to meet per-
formance and area constraint; power was a small detail to think about after all the
other requirements. The most important low power design method was simply to
avoid ‘wasting’ it, and if power was a target obvious effective solutions were either
Introduction 5
power × delay
delay
power
0 2 4 6
VDD / VT
Figure 1.1 Power and delay versus power supply voltage measured in threshold
voltage units [33]
just to reduce the power supply voltage, remove circuits that dissipate static power
or to power down inactive blocks. But using power as a metric is not completely
appropriate in the context of digital design, as power can always be reduced by
slowing down the operating frequency. A more suitable metric is the figure of merit
delay×power. Smaller delay×power means a lower energy solution at the same level
of performance. The most popular techniques to improve digital circuits performance
according to this metric can be classified as follows [33]:
1. Voltage scaling: A reduction of the power supply voltage (VDD ) brings with it
an immediate reduction of the energy per operation. But, as the capacitance (C) and
the threshold voltage (VT ) are constant, the speed of the gates2 will also decrease in
agreement with eq. (1.1) [34]:
CVDD
td ∝ (1.1)
(VDD − VT )2
As for the power, whenever the dynamic term is dominant, it can be written as
2
P ∝ C · VDD f (1.2)
These equations show that for VDD values much greater than VT , small decreases in
VDD lead to small increases in delay but much larger decreases in power consumption.
However, as VDD approaches the threshold voltage, further decreases in VDD lead to
much greater increases in delay for the equivalent decrease in power. Finally, the
variations are not significant for VDD values larger than three times the threshold
voltage. Around this point the figure of merit remains almost constant. These tradeoffs
are graphically illustrated in Fig. 1.1.
2. Transistor sizing: The main contribution to the load capacitance is the gate
capacitance of the forthcoming gate, so by making these capacitances smaller the
power consumption will also decrease. However, for a minimum transistor length, a
reduction on its width also reduces the maximum value of the current it can drive.
2 t is the delay time. f is the operating frequency.

d
In general, the best tradeoff is to design the transistors with such dimensions that the
percentage of the load due to the gate is between 20 per cent and 80 per cent of the
total.
3. Adiabatic circuits: This approach is based on resonating the loading capac-
itance with an inductor which recovers some of the energy loss in switching the
load [35].
4. Technology scaling: Ideally, if the technology is scaled in such a way that all
voltages and dimensions are also scaled in the same proportion, the delay × power
product would be reduced up to the fourth power of the scaling factor. But, this would
only be true if the static power is negligible. Otherwise it would just improve as the
square of it. However, there is a limit for this determined by the interconnections.
5. Transition reduction: This technique consists of selectively powering down
those sections that do not have any influence on the circuit response during a certain
period of time. The activation is usually done by gating the clock to the function
block.
6. Operand isolation: This technique is applicable when the circuit inputs are
changing rapidly but the output is only needed once every few clock cycles. In this
case, the input signal can be isolated from the rest of the circuit, so reducing switching
and thereby the power consumption.
7. Precomputation: This technique consists of breaking the function performed
on a large bus of data into two parts: a small and a large block. The small block is
precomputed and power is consumed in the rest of the function only if it is valid.
8. Parallelism: This technique consists of having a number N of identical func-
tional units operating in parallel. This increases the energy per operation by only the
power overhead while the delay per operation drops by N minus the delay overhead.
9. Problem redefinition: The most effective way of reducing the power consump-
tion in digital systems is to reduce the number/complexity of tasks that the operation
requires. It is at this level that the designer has a bigger task, as simplifications often
reduce both the energy and the delay to complete the operation. A reformulation of
the problem can lead to a solution that requires less computation to accomplish the
same task – it is possible to obtain a 50 per cent power reduction from architectural
changes.
The last chapter of this book shows how it is possible to develop digital circuits
topologies that considerably reduce the figure of merit power × delay by using the
FGMOS transistor as a computational block.
1.3 Why floating gate MOS?
The aim of this book is to show the reader how to use a device that can be fabricated
in all CMOS technologies, the FGMOS, to design circuits that (a) can operate at
power supply voltage levels which are well below the intended operational limits
for a particular technology and (b) consume less power than the minimum required
power of a circuit designed with only MOS devices in the same technology with the
same performance.
Introduction 7
The two main design targets will be low voltage and low power (LV/LP). They
will be achieved by pursuing four different subgoals.
1. Reduce the circuit complexity: As the circuitry gets simpler (fewer devices), fewer
current branches are required, and therefore the power consumption decreases.
This also has other benefits related to the frequency response, since the number
of internal nodes is smaller.
2. Simplify the signal processing: Complex functions are easier to implement using
FGMOS transistors. These will be used in nonlinear signal processing to reduce
the voltage demands.
3. Shift the signal levels: The devices will be biased in the most appropriate operating
region for a wider range of input signals, by shifting the effective threshold
voltages accordingly in the FGMOS transistors. This can be achieved without
the need for extra level shifters, although in some cases it can be detrimental as
well. These cases will also be discussed throughout the book.
4. Facilitate the tuning: Tuning becomes even more of an issue in a low voltage
context, where variations are more critical since they can bias the devices out
of their intended operating region. FGMOS transistors increase the number of
degrees of freedom available to tune/program the circuits.
The book will present the device as a powerful mathematic/electronic element,
which offers three very important properties in the LV/LP context:
Flexibility, Controllability, Tunability
Flexibility to implement both linear as well as complex nonlinear functions in a

compact and easy fashion, thus leading to the consequent simplification of the circuits.
Controllability, since the effective threshold voltage of every single transistor can be
controlled separately according to the operating range needs.
Tunability, because it is a multiple input device and can be designed to be tuned just
by adding extra inputs.
These properties will be extensively illustrated throughout this book as they will be
used in the design of several LV/LP analog and digital blocks.
1.4 FGMOS history
Due to the special characteristics of the FGMOS transistor, its application in both
analog and digital circuits has been very wide since the first report in 1967 [36,37].
The first well-known application of the FGMOS was to store data in EEPROMs,
EPROMs and FLASH memories. In the late 1980s, the Intel ETANN chip, employed
it as an analog nonvolatile memory element [38]. Today this technology is present
in every personal computer [39]. But this was not the only context the transistor
has been used in. During the last ten years, a number of different applications have
revealed possibilities that this device could have in many other different fields. This
results from its versatility in implementing different functions, as well as its capacity
for programmability. In the following text, several examples illustrating use of the
device in different frameworks will be cited, together with references where many
others can be found.
The FGMOS has been widely used as a trimming element for tuning purposes.
An example of this can be found in Sackinger and Guggenbuhl’s work [40]. They
proposed an application for the FGMOS transistors as tunable elements in analog
CMOS amplifiers. The idea was to use these tunable elements to correct offset errors,
linearity, gain errors and so on. Previously, this was done using laser techniques,
programming of resistor networks or dynamic compensation. It was then proposed
that the correction process be carried out by means of a continuous modification
in the accumulated charge at the floating gate. The tuning was done by means of
electric pulses, thus avoiding the need for expensive laser equipment and preventing
heating. Another advantage was, that unlike previous works on nonvolatile memories,
in which the variations in the threshold voltages could be around 40 per cent in four
days, if FGMOS transistors were used these changes could be much smaller (around
1 per cent in ten years at room temperature, in a process designed for digital memories).
Moreover, tunable circuits implemented with floating gates could be reprogrammed
many times, making this technique good not only for tuning objectives but also for
synapses, either in neural networks or in adaptive filters.
Yu and Geiger [41] suggested a similar method to design analog circuits working
at a very low voltage supply, scaling down the threshold voltage and also injecting
charge using the tunnel effect [42], but with the difference that the high voltages
were not obtained externally but internally with a charge pump, itself designed with
floating gates. In order to control the charge at the gate, Fowler–Nordheim tunnelling
current had to be generated. This required an EEPROM technology consisting of a
standard CMOS technology with an extension in the process to have an ultra-thin
oxide so that the tunnel effect could be forced, transistors which could bear high
voltages and a second polysilicon layer. All these requirements contributed to the
increase in the chip cost. Besides, these technologies were generally lagging one
or two generations behind the state-of-the-art CMOS technologies. This implied a
penalty in the integration density and also, due to the rather small size of the market
for EEPROM technologies, an effective monopoly is in place. Therefore, a circuit
with an on-chip EEPROM is expensive. The cost argument is especially valid for ICs
where the EEPROM is only a small fraction of the circuit.
Research focusing on overcoming this drawback led Thomsen et al. to fabricate a
tunnelling injector in a standard double-polysilicon CMOS process [43]. Op’t Eynde
and Zorio [44] proposed a new alternative for using FGMOS in circuits which only
had to memorise a few tens of bits, e.g. for circuit identification or for the trimming of
analog circuits by means of digital calibration. The EEPROM function was realised
in a standard CMOS technology, even when the resulting cell was larger than a
normal EEPROM. The use of even higher voltages was essential to provoke tunnelling
(around 35 V). However, in order to avoid the dielectric breakdown, time-dependent-
dielectric-breakdown (TDDB) or gate-induced-drain-leakage (GIDL) these voltages
were never applied directly to the transistor. The handicap was that these memories
Introduction 9
were useful for applications where only a few write-and-erase cycles were required
and the high programming voltage was applied externally during programming. Other
examples where the FGMOS transistor was used as a trimming element with the help
of the tunnel effect can be found from [45] to [49].
Shibata and Ohmi [50–53] proposed a different application for the transistor and
a different name: Neuron MOS (νMOS), since its functional behaviour is quite anal-
ogous to that of biological neurons: the ‘on’ and ‘off’ of the transistor is based on
the result of a weighted sum operation. The transistor was used in multivalued logic
[54–57], with the advantages this had over binary logic, since a large amount of data
can be processed per unit of area, with a reduced number of connections. This was
proved with NMOS and CMOS implementations of the hardware algorithms. Gen-
erally, although the calculations were implemented in current mode increasing the
power dissipation, the νMOS realises the sum operation in the voltage mode using a
capacitive division, and so the current flow is zero with the exception of capacitance
charges and discharges. This was very important, when the integration density was
very large [58].
Simultaneously to Shibata and Ohmi’s work, a different research was carried out,
in which the FGMOS transistor was used as a memory element in neural networks.
But it was still difficult to control the analog data stored at the gate with a high
resolution. The fine tuning of the charge injected by tunnelling was not easy, as
the tunnelling current did not have a linear dependence with the junction voltage.
Feedback circuits were an option but this would increment the chip area and the
power consumption. Fujita and Amamiya [59] developed an FGMOS device which
could be used as a precision analog memory for neural networks LSI (large scale
integration). This device had two floating gates. One was a charge-injection gate
with a Fowler–Nordheim tunnel junction and the other was a charge-storage gate that
operated as an MOS floating gate. The gates were connected by a high resistance, and
the charge-injection gate was small so that its capacitance was much smaller than that
of the charge-storage gate. By applying control pulses to the charge-injection gate, it
was possible to charge and to discharge the MOS floating gate in order to modify the
current with high resolution over 10 bits. This had a high applicability for on-chip
learning in analog neural network LSIs.
Learning on a neural network is the change in the synaptic weight according to the
learning algorithm. The use of the floating gate EEPROM technology for electronic
synapses has presented a crucial problem: the amount of incremental data change by
a constant programming pulse is not proportional to the number of pulses. The reason
is that the electric field in the tunnel oxide is determined not only by the magnitude of
a programming pulse but also by the amount of charge stored in the floating gate. The
strong nonlinear dependence on the electric field made it extremely difficult to update
the weight accurately via the number of pulses. Thus, to achieve this, complicated
control circuitry was usually employed [60]. Kosaka et al. [61,62] presented a synapse
with a good weight updating linearity under constant pulse programming. This was
realised by employing a simple self-feedback regime in each cell circuit. As it was
very simple, neither the speed nor the integration density were affected. It is also
worth noting the work of P. Hasler, C. Diorio and B. Minch, which will be delved into
deeper in the forthcoming text. In any case, a more in-depth analysis of the FGMOS
evolution in the adaptive circuits context can be carried out by reading other relevant
research [63–81].
In addition to the memory, neural networks and logic systems applications above,
the FGMOS has other potential possibilities. These result from the consequence
of the weighted sum of voltages that can be performed in a lossless node. This
allows a simplification in many blocks for analog signal processing. Yang and
Andreou proved the effectiveness of the technique by designing a multiple input
differential amplifier which functioned like a standard differential difference ampli-
fier (DDA). It had some unique advantages, including lower circuit complexity and
better input transistor matching [82], but the accumulated charge at the floating gate
was not altered. The possible residual charge that might remain trapped at the gate
during the fabrication process was an unknown parameter which had to be elimi-
nated. They accomplished this using ultraviolet light [83–85]. These authors also
developed a model to design with FGMOS transistors working in the subthreshold
region [86].
The model was then used by Hasler, Minch, Diorio and Mead’s group to carry out
their work based on the FGMOS transistor in subthreshold, which comprised several
applications for this transistor [87] – from nonlinear functions implementation in
current mode [88,89] to the design of memory cells [90,91], combining the tunnel
effect with the hot electron injection in FGMOS [92,93]. They started physically
modelling both effects following the design of a synapse with long-term weights
storage capability [66]. The single transistor synapses simultaneously performed long-
term weight storage, computed the product of the input and floating gate value and
updated the weight value according to a Hebbian or a backpropagation learning rule,
[66,67,69,70,74,92]. The charge in the floating gate was being reduced by using hot
electron injection with high selectivity for a particular synapse. It was being increased
by using electron tunnelling. As the FGMOS was working in subthreshold and the
synapse was very small, this allowed the implementation of small learning systems
with a low power consumption. Unlike conventional EEPROMs, reading and writing
were now possible simultaneously. Moreover, they also introduced the autozeroing
floating gate amplifier (AFGA) as an example of adaptive circuit based on the FGMOS
technology, which can learn continuously [68,71,78,81,92].
Other authors devoted their efforts to the design of analog blocks that take advan-
tage of the FGMOS capability to perform the sum operation at the gate. Some
examples of these blocks are multipliers, rectifiers, amplifiers and so on [94–103].
Notable is the work of Ramírez–Angulo in low voltage and low power analog design
with FGMOS. The transistor also found a niche of application in the fuzzy logic area
[104–106]. Ultraviolet light was still used to clean the floating gate.
Lande and Berg’s group used the light with a different purpose. They established
an ultraviolet conductance between the gate and the drain in the transistor. To achieve
this, the whole chip was covered with a second metal layer, except for a hole in the
boundaries between the gate and the drain [83–85]. During light exposure supplies
were interchanged and certain conditions were set for currents and transconductances,
in such a way that the equilibrium point was VDD /2 for inputs and outputs when the
Introduction 11
MOS floating gates were at the maximum power supply. This initial condition was
equivalent to a shift in the threshold voltage values, which allowed the design of
circuits working at power supply voltages as low as 0.5 V [107–116].
1.5 Structure of the book
The aim of this book is to present the FGMOS transistor as a powerful device from the
circuit design point of view. The inputs in the transistor, significantly more numerous
compared with the normal MOS device, appear as extra degrees of freedom which the
designer has to play with in order to get a desired functionality. Hence, by establishing
the right relationships between them, it is possible to achieve design tradeoffs that are
not viable with conventional MOS devices, especially in a context in which power
consumption and supply voltage are the main design constraints.
The designs and methodologies in this book are not based on exploiting the phys-
ical properties of the device, but just the mathematical ones. Therefore, the reader
will not find circuits here whose functionality is based on modifying the charge at
the floating gates by using, for example, tunnelling, hot electron injection or UV
light. If interested, the reader can find out more about such circuits in the references
mentioned in the previous section. All the circuits presented here can be realised in
any double poly CMOS process and no extra post-processing is needed.
The book will also show how, with no pretensions to being a panacea, the com-
bination of the FGMOS transistor with LV/LP design techniques provides a good
circuit design solution, mostly in situations when the limits of technologies have
been pushed far below the values recommended by the manufacturer. In this way, cir-
cuits that conventional MOS transistors would render dysfunctional will work when
some of these transistors are replaced by FGMOS devices.
The book is mostly about analog design for two reasons: (a) There is more room
for creativity in analog circuit design when extra degrees of freedom are available
and (b) it is in this context that the device provides greatest advantages. However, a
chapter on digital design has been included, because the device also turns out to be
ideal for the design of a threshold gate digital block. This can greatly simplify the
implementation of digital functions thereby reducing power.
The content is organised as follows:
Chapter 2: ‘The Floating Gate MOS transistor: FGMOS’. This chapter intro-
duces the device, starting from the physical structure, followed by some basic device
modelling. Subsequently, practical problems that designers traditionally found when
starting to work with the device are outlined, together with solutions for these
problems.
Chapter 3: ‘FGMOS-Circuit applications and design techniques’. This chap-
ter explains the advantages of using an FGMOS device over a standard MOS device
in circuits which require low power and low voltage operation. Several simple cir-
cuits are presented in order to illustrate the device functionality, explaining the design
tradeoffs together with examples.
Chapter 4: ‘Low power analog continuous-time filtering based on the

FGMOS in the strong inversion ohmic region’. This is the first of four chapters
that describe how to exploit the mathematical capabilities of FGMOS devices in more
complex circuits, taking as example the design of continuous-time analog filters. In
this chapter, the performance of the filter under the low voltage and low power con-
straints is mainly obtained due to the functionality of FGMOS devices biased in the
strong inversion ohmic region. The design methodology is discussed in this, as in the
other three chapters, together with a thorough analysis of second-order effects and
then design tradeoffs when using the transistor.
Chapter 5: ‘Low power analog continuous-time filtering based on the
FGMOS in the strong inversion saturation region’. This chapter is quite simi-
lar in structure to the previous one, the difference between the two of them being
that the topologies presented here achieve low power and low voltage operation by
exploiting the properties of FGMOS devices biased in the strong inversion satura-
tion region. Two different designs will be presented, one of them intended for very
low frequency and the other one for intermediate frequency applications. Again, the
advantages and disadvantages of using the device will be extensively discussed in
both. This chapter also shows with an example what happens in a particular design
when low quality metal/poly capacitors are used in the device, for those cases in which
neither double poly, nor high quality metal/insulator/metal capacitors are available in
the technology.
Chapter 6: ‘Low power analog continuous-time Gm -C filtering using the
FGMOS in the weak inversion region’. This chapter and the next are devoted to
exploring the potential of the FGMOS transistor operating in weak inversion for
the design of analog low voltage micropower continuous-time filters. This chapter
describes a linearised Gm -C topology in which the linearisation is achieved by means
of two ‘floating gate’ blocks. The chapter begins with a reformulation of the translinear
principle (TP), which is an ideal technique for implementing nonlinear functions using
currents in FGMOS. The advantages of an FGMOS implementation compared with
those of a normal MOS implementation in a low voltage context are then discussed.
The linearised Gm -C filter is described afterwards followed by an analysis of second-
order effects.
Chapter 7: ‘Low power log-domain filtering based on the FGMOS tran-
sistor’. This chapter will explain how to design a log domain filter using FGMOS
transistors. It starts with a review of the basic log domain principle, followed by the
description of basic FGMOS blocks that implement the required functions. Again, the
advantages and disadvantages of using FGMOS devices in these kinds of topologies
will be discussed, together with a qualitative and quantitative analysis of second-order
effects.
Chapter 8: ‘Low power digital design based on the FGMOS threshold gate’.
This chapter mainly focuses on the discussion of an efficient form of implementing
digital circuits by using the threshold gate concept. These threshold gates will be
implemented with FGMOS transistors. It will be shown how to successfully carry
out the aforementioned digital technique of redefining the problem, combined with
the utilisation of the FGMOS device, in order to reduce power. This is illustrated with
Introduction 13
several types of digital circuits grouped in combinational and sequential blocks. All
these examples will show how the figure of merit power × delay can be improved
with respect to the values obtained with traditional approaches. This chapter is not
as incremental as the previous ones, so a previous theoretical background in digital
design is recommended.
Chapter 9: ‘Summary and conclusions’. This chapter brings together the main
points of each chapter and revisits the subject relevance.
Since the number of degrees of freedom that need to be handled in all these designs
is larger than in normal approaches, the mathematics in all these chapters utilise a
considerably large number of variables. Because of this a notation section can be
found at the end of each chapter for the reader to refer to as needed.
Chapter 2
The Floating Gate MOS transistor (FGMOS)
2.1 Introduction
This chapter introduces the Floating Gate MOS transistor (FGMOS). The properties of
this device are described and a simple model for hand analysis is presented. Through-
out the chapter, the FGMOS transistor is compared with a standard MOS device, and
the main advantages and disadvantages of using an FGMOS instead of an MOS tran-
sistor are drawn. Also, some of the common initial problems a designer faces when
trying to use the FGMOS are outlined and solutions for them are provided. Hence, in
this chapter, the reader will find out how to perform accurate simulations of circuits
containing FGMOS devices without having to model them. Also, the frequently asked
question ‘how to effectively eliminate the charge accumulated at the floating gate dur-
ing the fabrication process’ is reviewed and answered in one of the sections. Finally,
a quantitative analysis of the minimum extra area required by an FGMOS when
compared with an MOS with the same channel size is presented in the final sections.
2.2 The Floating Gate MOS (FGMOS) device
2.2.1 Introducing the device

An FGMOS can be fabricated by electrically isolating the gate of a standard MOS
transistor, so that there are no resistive connections to its gate. A number of secondary
gates or inputs are then deposited above the floating gate (FG) and are electrically
isolated from it. These inputs are only capacitively connected to the FG, since the FG
is completely surrounded by highly resistive material. So, in terms of its DC operating
point, the FG is a floating node.
A possible layout of a three-input n-channel FGMOS transistor together with
three cross-sectional views are shown in Fig. 2.13 , and will be used to illustrate
3 Fig. 2.1 is not drawn to scale.

Section A
Source Drain
SiO2 SiO2
FG-Poly1
SiO
2
n+ n+
P- Well
P-Substrate
Section C
SiO2
Section C
SiO 2
P- Well
Section A
FG-Poly1
Source Drain
P-Substrate
Input2-Poly2
FG-Poly1
Section B Input1 Input2 Input3
Input2
SiO2
SiO2
Poly2 Poly2 Poly2
Section B
Input1-Poly2
In t1 Input2-Poly2
t2 Input3-Poly2
In t3 SiO2
C1 C2 SiO2 C3
FG-Poly1
SiO2
P-Well
P-Substrate
Figure 2.1 Possible layout of a 3-input n-channel FGMOS transistor (middle) and
its cross-sectional views (A, B and C)
important features of FGMOS transistors as well as to show how they can be fabricated
in existing MOS technologies. The equivalent schematic for an N-input n-channel
FGMOS transistor is given in Fig. 2.2 whilst Fig. 2.3 shows the schematic symbol
that will be used throughout this book.
It can be seen in Fig. 2.1 how the FG, fabricated using the gate electrode (poly1)
layer, extends outside the active area of the MOS transistor. The FG is surrounded
by two SiO2 insulator layers and thus electrically isolated from the rest of the device.
The floating gate MOS transistor 17
C GD D
C1
V1
D: Drain
C2 S: Source
V2 FG
B: Bulk
C3 B
FG: Floating Gate
V3
Vi: Effective Inputs, i = [1,N]
CN
VN
C GS S
C GB
Figure 2.2 Equivalent schematic for an n-channel N-input FGMOS transistor
VD
V1 V2 V 3
VN
VS S
Figure 2.3 Symbol and voltage definitions for an n-channel N-input FGMOS
The device inputs are placed on top of the upper SiO2 insulating layer and are
fabricated using another conducting layer, preferably a second layer of polysilicon
(poly2). The sizes of the input electrodes (Fig. 2.1) determine the values of the capac-
itors that connect the FGMOS inputs with the FG and they can be varied according
to the designer’s needs.
The values of the input capacitances are given by

εSiO2
Ci = Ai (2.1)
tSiO2
where εSiO2 is the permittivity of the SiO2 , tSiO2 is the thickness of the SiO2 between
the FG and the effective inputs and Ai is the area of each input capacitor plate.
2.2.2 Theory
2.2.2.1 Large signal DC behaviour
Equations modelling the operation of the FGMOS can be derived, in a very easy way,
from the equations that describe the operation of the MOS transistor used to build the
FGMOS. The derivation procedure is described in this section.
The input parameters that determine the drain-to-source current in an MOS tran-
sistor are the voltages between its terminals: gate-to-source (VGS ), drain-to-source
(VDS ) and source-to-bulk (VSB ). The relationship between these voltages and the
current has been studied thoroughly by various researchers and can be expressed
mathematically using a number of different equations that model it in several operat-
ing regimes [117,118]. In the case of the FGMOS transistor, the ‘external parameters’
that determine its drain-to-source current are the voltages at its effective inputs, as
well as the drain and source voltages, all of them referred to the bulk (Fig. 2.3).
If it is possible to determine the voltage at the FG of an FGMOS device, it is then
possible to express its drain to source current using standard MOS transistor models.
Therefore, in order to derive a set of equations that model the large signal operation
of an FGMOS device, it is necessary to find the relationship between its effective
input voltages (Vi ) and the voltage at its FG. This can be done with the help of the
equivalent circuit shown in Fig. 2.2.
The distinctive feature of the FGMOS device is the set of input capacitors, denoted
Ci where i = [1, N ], in Fig. 2.2, between the effective inputs and the FG. The parasitic
capacitances, CGD and CGS , represented using dotted lines, are the same parasitic
capacitances that would be present in an MOS transistor fabricated using the same
technology with the same active area4 . The relationship between the DC drain to
source current and the FG voltage, VFG , of an FGMOS is not affected by parasitic
capacitances. However, CGD , CGS and CGB do affect the relationship between VFG
and the effective input voltages Vi .
In summary, the equations that model the static, large signal behaviour of an
FGMOS can be obtained by combining a standard MOS model for the same technol-
ogy with the equation that relates VFG to Vi , VD , VS , Ci , CGD , CGS and CGB . This
equation can be obtained by applying the charge conservation law to the floating node
(FG) shown in Fig. 2.2. If there is an infinite resistance between the FG and all the
surrounding layers, there will be no leakage current between them, and so, the FG
will be perfectly isolated. Under this assumption the voltage at the FG is given by

N
Ci CGS CGD QFG
VFG = Vi + VS + VD +
CT CT CT CT
i=1

N
Ci CGD CGB QFG
= ViS + VDS + VBS + + VS (2.2)
CT CT CT CT
i=1
4 C
GB is not the same, which will be explained later in this chapter.
where N is the number of effective inputs. The term QFG refers to a certain amount
of charge that has been trapped in the FG during fabrication. As this term is constant,
it can be interpreted as a voltage offset at the FG, or alternatively, as an offset in the
threshold voltage of the device. The term CT refers to the total capacitance seen by
the FG and is given by

N
CT = CGD + CGS + CGB + Ci (2.3)
i=1
The equations modelling the large signal behaviour of the FGMOS can now be
obtained by replacing VGS in the equations describing the large signal behaviour of
the MOS transistor, with the expression describing the voltage between the FG and
source which can be obtained by referring VFG to the source terminal rather than
the bulk. This procedure is illustrated in Table 2.1. The left-hand column shows
typical model equations for hand calculations for an n-channel MOS transistor, and
the right-hand column shows the equivalent expressions for the FGMOS. W and L
are the effective width and length of the transistor. The undefined parameters have
their usual meaning [119].
2.2.2.2 Small signal equations

An N -input FGMOS device has N − 1 more terminals than an MOS transistor, and
therefore, N +2 small signal parameters can be defined: N effective input transconduc-
tances, gmi , for i = [1, N ], an output conductance gdsF and a bulk transconductance
gmbF . If gm , gds and gmb are the gate transconductance, output conductance and bulk
transconductance, respectively, of an MOS transistor having identical channel size,
current and VDS as the FGMOS, the FGMOS small signal parameters are then related
to the MOS parameters in the following way5 :
Ci
gmi = gm for i = [1, N ] (2.10)
CT
CGD
gdsF = gds + gm (2.11)
CT
CGB
gmbF = gmb + gm (2.12)
CT
These equations show two drawbacks of the FGMOS compared with the MOS
transistor: the reduction of the input transconductance and the reduction of the
output resistance. In order to illustrate this, let us take as an example the common
source amplifier shown in Fig. 2.4(a). Its gain (G) and output resistance (Rout ) are

(3) (2) (3) −1 (2) (3) −1
G = gm gds + gds Rout = gds + gds (2.13)
5 The term g will be employed to refer to the effective transconductance only in this chapter. For the
mi
sake of clarity in the notation, its value, Ci /CT · gm will be used instead in subsequent chapters.
Table 2.1 MOS versus FGMOS model equations
MOS FGMOS
Weak inversion saturation when VBS = 0

N
Ci ViS
CGD VDS QFG
C nUT
ID = Is e(VGS /nUT ) (2.4) ID = Is e i=1 T e CT nUT e nUT CT (2.5)

N C
i C Q
for VGS < VT , VBS = 0 for ViS + GD VDS + FG < VT
i=1 CT CT CT
VDS > 4UT VDS > 4UT , VBS = 0
Strong inversion ohmic

N
W W Ci
ID = μ0 Cox ID = μ0 Cox ViS
L
L i=1 CT

V CGB Q 1 CGD
× (VGS − VT )− DS VDS − VT − VBS − FG − − VDS
2 CT CT 2 CT
(2.6) (2.7)
N
Ci C
for 0 < VDS ≤ (VGS − VT ) for 0 < VDS ≤ ViS + GD VDS
i=1 CT CT
CGB QFG
VGS > VT + VBS + − VT
CT CT
VGS > VT
Strong inversion saturation

N
μ Cox W μ0 Cox W Ci C
ID = 0 (VGS − VT )2 ID = ViS + GD VDS
2 L 2 L i=1 CT CT
2
β CGB QFG
= (VGS − VT )2 (2.8) + VBS + − VT
2 C CT
TN 2
β Ci C C Q
for 0 < (VGS − VT ) ≤ VDS = ViS + GD VDS + GB VBS + FG −VT
2 C CT CT CT
N i=1 T
Ci CGD CGB
VGS > VT 0< ViS + VDS + VBS (2.9)
i=1 CT CT CT
Q
+ FG − VT ≤ VDS
CT
VGS > VT
where the superscript (i) refers to the transistor named Mi. If the input device is
substituted by an FGMOS, the new values for the gain and output resistance will be

(3) (2) (3) CGD (3) −1

GFG = w1 gm gds + gds + g
CT m

(2) (3) CGD (3) −1
RoutFG = gds + gds + g (2.14)
CT m
(a) (b)
M1 M2 M1 M2
Vout Vout
w1
Vin
Vin M3 M3
I BIAS
Vbias
I BIAS
w2
Figure 2.4 Common source amplifier: (a) With an MOS transistor at the input. (b)
With an FGMOS transistor at the input
where wi are the equivalent input weights (Ci /CT ) for the FGMOS transistor
(Fig. 2.4(b)). Hence if, for example, the circuit in Fig. 2.4(a) is designed with
a gain of 100 and output resistance of 1 M and the same circuit is designed
by replacing the input MOS transistor by an FGMOS device with two identical
inputs (w1 = w2 ), CGD /C1 = 0.005, CGS /C1 = 0.05 and CGB /C1 = 0.1
(w1 = w2 = C1 /(2C1 + CGD + CGS + CGB ) = 0.464), the new gain would then be
GFG = 38 and the output resistance RoutFG = 812 k. However, benefits of FGMOS
often overweight these disadvantages.
2.2.2.3 Noise and dynamic range

The sources of noise in the FGMOS are the same as in the MOS transistor since the
noise is generated by the movement of carriers in the channel and the carriers are
the same in both devices. However, the equivalent noise at the input of the FGMOS
will be different, since the signal at the FG is attenuated, or, what is equivalent, it is
amplified at the effective input. Hence, the equivalent power spectral density of the
noise at an effective input will be
2
vin
2 2
v
1
= · in (2.15)
f effective input wi f FG
where wi is the capacitive weight corresponding to the input i and (vin

2 /f )|
FG is the
noise power spectral density at the FG. Equation (2.15) shows how the equivalent
noise increases at the effective input of an FGMOS device, since the factor (1/wi )
is greater than 1. However, in the worst case scenario, the DR will not be affected
because the maximum allowable input signal for a certain level of distortion will also
be increased by the same factor (1/wi ). The reason for this is the following: if the
input capacitances are assumed to be linear, the signal seen by the FG is actually the
signal at the effective input of the FGMOS attenuated by a factor wi .
In general though, the DR will even improve since, for example, from the system
level point of view, the use of the FGMOS will greatly simplify the circuit structure.
This will consequently reduce the distortion as well as the noise levels since the
number of devices and hence the number of noise contributors will be smaller.
2.3 Designing with FGMOS: problems and solutions
Designing with FGMOS is not an easy task for a number of reasons that will be
subsequently described. Under normal conditions, a floating node in a circuit rep-
resents an ‘error’ due to the fact that the initial condition is unknown unless it is
somehow fixed [120]. This generates two problems of different nature. On one hand,
it is not straight forward to simulate these circuits. On the other hand, an unknown
amount of charge might stay trapped at the FG during the fabrication process which
will result in an unknown initial condition for the FG voltage. Some of the most com-
monly used techniques and solutions to deal with these two problems are discussed
in this section. Their advantages and disadvantages are illustrated with examples.
2.3.1 Simulation
The first problem any designer finds when dealing with FGMOS is the lack of sim-
ulation models. Because of this, the simulators do not understand the FG and report
it as a convergence error [120]. Several solutions to this problem have been offered
in literature [121–124]. The most popular ones are briefly described here putting a
particular emphasis on the technique that offers the most reliable prediction of the
FGMOS operation [124].
The first solution to the convergence problem was offered by Ramírez-Angulo
et al. in [121]. It is based on the simulation model shown in Fig. 2.5. This model
overcomes the simulation problem by connecting a very high value resistor (RG )
between the FG node and a set of voltage controlled voltage sources (VCVS) that
model the addition at the FG as described by eq. (2.2).
The gains (ai , ao and ao ) of the VCVSs are the ratios between the corresponding
input capacitances and the total capacitance seen by the FG. The simulator calculates
C1
V1 VD
C2 CGB
V2 Ci
ai =
Cn C GS + C GD + C 1 + C 2 + ·· + C n + C GB
Vn RG
VS C GS
ao =
an Vn CGS + CGD + C1 + C2 + ·· + Cn + CGB
a1 V1 CGD
a⬘o =
CGS + CGD + C1 + C2 + ·· + Cn + CGB
ao VS
R G ≅ 1000G
a⬘o VD
Figure 2.5 Ramírez-Angulo’s model for simulation of the FGMOS

the DC operating point by annulling the capacitances. This creates an open circuit at
the gate. As no current is flowing through the resistance RG , the FG DC operating
point is given by the sum of the VCVS voltages. In AC analysis, the VCVSs do not
have any influence on the simulation because the resistance connected to the gate is
very large and thus behaves as an open circuit even for very low frequencies.
Although very commonly used, this approach has several important limitations.
On one hand, the parasitic capacitances have to be known from the beginning. The
only way to obtain their values is to, prior to application of the proposed method,
carry out a set of simulations with an estimated operating point at the FG. This is
not completely correct since the voltage at the FG does depend on these parasitic
capacitances. Besides, once their values are obtained this method assumes that they
are constant for the whole operating range, which is not always true. In fact the
deviations from these values can be quite significant if the transistor has to operate
in more than one region. Also, the final values of the parasitics connected to the FG
increase once the circuit has been laid out and this can be difficult to model in the
schematics [125]. All these facts increase the risk of failure in certain topologies, and
is critical in those analog cells that have to work with reduced voltage margins, such
as, for example, the low voltage circuits described along this book. Circuits operating
in the subthreshold region are also very much affected by the variation of the parasitic
capacitances since this variation will have an exponential effect.
A very important consequence of an incorrect estimation of CGD is the miscalcula-
tion of the output resistance (see eq. (2.11)), which can cause a number of undesired
effects such as unexpected losses in a bandpass filter, large variations in a quality
factor and cutoff frequency, instability and so on.
All these problems could be minimised by oversizing the transistors in such a
way that the parasitic capacitance which varies the most is much smaller than the
minimum input capacitance (how much smaller would depend on the value of the
expected variation and how it could affect the circuit performance). However, it is
not easy to know a priori how much bigger the input capacitors need to be. Previous
assumptions need to be made that are often not trivial and could even give rise
to undesired consequences such as instability in circuits with feedback. Besides,
oversizing the input capacitors might be unnecessary and therefore a waste in terms
of area.
Yin et al. suggest a different model to simulate FGMOS transistors [122]. The
model is based on connecting resistors in parallel with the input capacitors as shown in
Fig. 2.6. The equation for the operating point is the same one as in Ramírez-Angulo’s
approach. The problems of this model are also the same.
Another technique proposed in literature [123] is based on iterative simulations
which are directly performed by the simulator with the help of a program that has to
be previously implemented. This can be done using the SKILL language functions
in Cadence [123,125]. This technique is a more accurate, but still not exact enough,
version of the method in [121].
A reliable technique to simulate FGMOS transistors is illustrated in Fig. 2.7 [124].
It is based on the use of an initial transient analysis (ITA). At the beginning of this
analysis supply voltages and circuit inputs are set to 0 V. Under these conditions,
C1
V1
R1
C2 VF
V2 1
R i = -------- ≅ 1000 GΩ
kCi
R2 C GB
Cn
Vn
R0
Rn
Figure 2.6 Model in [122] for the FGMOS
.tran
.ic v(fg1 ) = 0 .ic v(fg2) = 0 .ic v( fgn ) = 0
v(fg1)op
V1 = 0 V1
v(fg2)op
V2 = 0 V2
t = 0– t = 0+
Vn = 0 Vn
v(fgn )op
VDD = 0 VSS = 0 VDD VSS
.ic v(fg2 )=v(fg2 )op

.ic v(fg1)= v(fg1)op .ic v( fgn ) = v(fgn )op
.ac
V1
V2
Vn
VDD VSS
Figure 2.7 Simulation technique based on an initial transient analysis (ITA)
it can be guaranteed that the initial voltages at the FGs are also zero, unless that, as it
happens in some technological processes or under certain conditions, some residual
charge (QFG ) remains trapped at the FG after fabrication. If this is the case, the
initial condition at the FGs would be QFG /CT . In any case, this is an unpredictable
term common to all models and will be discussed in more detail in the next section.
The transient analysis starts with these initial conditions. If, for example, HSpice is
(a) (b)
V1
Vc
V3
VD VD
V2 Vin
Floating gate Floating gate
Figure 2.8 (a) Circuit used to illustrate the simulation technique in Fig. 2.7
(b) Circuit used to compare the techniques in Figs. 2.5 and 2.7
used, the initial conditions can be imposed with the command .ic [120]. Subsequently,
the supply voltages are set to their final values and the inputs evolve normally. Small
signal simulations can be performed extracting the operating point (v( fgi )op for i =
[1, n]) from the transient analysis and forcing it to the FGs with the help of very
high value inductances. The inductances behave as wires in DC and fix the operating
point. Besides, if they are big enough, their effect is negligible for any other operating
frequency. In the case of Hspice, for example, after completing a transient simulation,
the operating point can be found in the file input_file_name.ic (where input_file_name
is the name of the input file). In order to force the initial conditions the command
.nodeset at the beginning of the file needs to be changed to .ic.
An example of how to use this simulation technique is provided below. The netlist
describes the FGMOS inverter shown in Fig.2.8(a). First of all, the subcircuits for
generic n-channel and p-channel transistors are defined. The external nodes for these
subcircuits are the FGs, drains, effective inputs and sources. The parameters are the
values of the input capacitances and the dimensions of transistors. The reason for
adding the FG to the list of external nodes, despite the fact that it has to be completely
isolated, is that it gives the freedom to change the number of capacitive inputs without
having to define a new device. The simulation shows how to perform a parametric
DC analysis. As it was previously explained, voltage supplies and inputs are initially
set to 0 V and a zero initial condition is forced at the FG with the option .ic. When
the circuit is ‘powered up’ [vdd vdd 0 pulse(0 2 0.01 0.01 1 2 4)] , the inputs that are
going to be constant in the steady state change quickly [v1 v1 0 pulse(0 a 0.01 0.01
1 2 4), v3 v3 0 pulse(0 2 0.01 0.01 1 2 4)], whereas the input which is supposed to
be swept in DC rises very slowly [v2 v2 0 pulse (0 2 0.01 1 1 2 4)]. In subsequent
simulations the final value of the DC bias (‘a’ parameter in v1) is set to different
values thus performing the parametric variations. The input–output characteristics
obtained with these simulations are shown in Fig. 2.9.
Netlist for the circuit in Fig. 2.8(a)
Inverter
.include ’c:\simulation\modelos\NC\Typical\NC.sp’
Inverter
2
VD(V)
1
0 1 2
V2(V)
Figure 2.9 Input–output characteristic for the FGMOS inverter in Fig. 2.8(a)
.include ’c:\simulation\modelos\PC\Typical\PC.sp’
.subckt transistorn vg vd v1 v2 v3 vs w=20u l=20u

c1=210f c2=210f c3=210f
m1 vd vg vs 0 nc w=’w’ l=’l’
c1 v1 vg c1
c2 v2 vg c2
c3 v3 vg c3
.ic vg=0
.ends
.subckt transistorp vg vd v1 v2 v3 vs w=20u l=20u

c1=210f c2=210f c3=210f
m1 vd vg vs 0 pc w=’w’ l=’l’
c1 v1 vg c1
c2 v2 vg c2
c3 v3 vg c3
.ic vg=0
.ends
x1 vg vd v1 v2 v3 0 transistorn w=20u l=20u c1=210f

c2=210f c3=210f
x2 vg vd v1 v2 v3 vdd transistorp w=20u l=20u c1=210f

c2=210f c3=210f
v1 v1 0 pulse(0 a 0.01 0.01 1 2 4)
v2 v2 0 pulse(0 2 0.02 1 1 2 4)
v3 v3 0 pulse(0 2 0.01 0.01 1 2 4)
vdd vdd 0 0 pulse(0 2 0.01 0.01 1 2 4)
.tran 0.01 1 sweep a 0 10 2
.op t=1
.options post
.end
The ITA technique is even easier to apply if the Spectre simulator is used in
Cadence to design the circuit [125]. In this case the input file does not need to be
modified to perform an AC simulation. The latter can be realised just by choosing the
simulation options adequately in a transient simulation, followed by an AC analysis.
In the initial transient analysis all the sources will be automatically set to zero by the
simulator. They will then rise to their final value between the ‘starttime’ and the time
zero. At the final time the operating point will be saved in a file. The AC analysis
will then be run using this initial operating point from the file.
4.4
Floating Gate Voltage (V)

(0.2 V/div)
–0.2
0 1 2 3 4
Time (s)
Figure 2.10 Voltage at the FG of the CMOS inverter in Fig. 2.8(b) obtained with
the VCVSs model and with the ITA technique
Figure 2.10 compares the ITA technique with the technique in [121] (see Fig. 2.5)
using as an example the input–output characteristic of the inverter in Fig. 2.8(b).
The evolution of the FG voltage is shown with one of the inputs being continuously
swept and the other one changing parametrically. The dashed lines marked with spots
are the results obtained using the ITA technique. The other lines are obtained using
Ramirez-Angulo’s method. There is a variation of around 0.2 V that demonstrates
how the latter is not very reliable and depends very much on the designer’s ability to
determine the parasitic capacitances accurately.
Figure 2.11 shows the output resistance of an FGMOS transistor obtained with a
transient analysis when either (a) the VCVSs technique or (b) the ITA technique is
used. It can be seen how the VCVSs technique generates a value around five times
smaller. This is due to a bad estimation of the CGD value.
2.3.2 Charge accumulation

Until recently FGMOS transistors had only been used in digital electronics EEPROM
devices [126] and despite the important role FGMOS devices could play in analog
circuits, designers have been reluctant to work with them. The main reason for this
is the uncertain amount of charge (QFG ) that might stay trapped at the FG during
the fabrication process causing variations of the threshold voltage. QFG has different
implications on different circuits; however, the design will often not work unless
the charge is removed. Reported solutions to this problem include cleaning with
ultraviolet (UV) light; the use of tunnel effect and hot electron injection; and forcing
an initial condition with a switch [42,83,84,92,127].
UV cleaning refers to exposing the FG to UV light. It is well known that when
the surface of any semiconductor is exposed to light or any other electromagnetic
radiation, part of this radiation is reflected, part of it is absorbed and the rest is
(a) 4.012
Voltage (V)
8 kΩ
.ac
4.008
–50n 50n
Current (5 n/div)
(b) 2.705
Voltage (V)
36.9 kΩ
.tran
2.695
–50n 50n
Current (5 n/div)
Figure 2.11 Comparison between the output resistances of an FGMOS obtained

using (a) VCVSs technique (b) ITA technique
transmitted [83,84]. The number of absorbed photons is proportional to the total

number of photons, and therefore to the light intensity. It also depends on the type of
semiconductor, the wavelength of the photons and the applied electric field. When
the FG is illuminated with UV light, the electrons which are trapped at the gate are
able to travel through the potential barrier in the oxide/silicon interface. Hence, when
the light illuminates the gate/source or gate/drain region of the MOS transistor, the
excited electrons will gain enough energy to traverse the gate oxide barrier. One
of the main drawbacks of this technique is in its incompatibility with those CMOS
technologies whose passivation layers are not transparent to UV light. In such cases,
in order to enable light transmission, the passivation has to be eliminated in the
fabrication process. Unfortunately, this can affect the lifetime of the chip.
Fowler–Nordheim tunnelling is the process whereby electrons tunnel through a
barrier in the presence of a high electric field. The main drawback of the tunnelling
technique is that, in order to generate Fowler–Nordheim currents, a very high electric
field has to be generated in the region between the drain and the gate. The magnitude
of the electric field is inversely proportional to the thickness of the oxide underneath
the gate and directly proportional to the voltage across it. Therefore, unless the oxide
is very thin, the voltage needs to be extremely high. Hence, for example, for an
oxide thickness in the order of 10 nm, the voltage required is around 15 V. Besides,
extra circuitry is required in order to control the intensity and duration of the Fowler–
Nordheim currents and thus the exact amount of charge at the FG.
Hot electron injection can be generated with smaller voltages. However, the
injection grows with the current in the channel. Hence, the currents have to be very
small in order to avoid huge variations. Again, the process has to be controlled
internally with additional circuitry that will be different for each system.
Vn
V2 B
V1
V FG (0) S
Figure 2.12 Technique to set the initial conditions at the FG
A drawback common to tunnelling and hot electron injection is that they are not
modelled in most technologies. Hence, in order to use them efficiently they would
have to be modelled first, which is not a simple task.
In any case, although these techniques (tunnelling and hot electron injection)
are not recommended when the intention is to erase the initial fabrication charge
(due to the complexity they add to the overall circuit topology), they have proven
to be a very interesting alternative when adaptability or programmability is required
[73–81,127,129]. Nevertheless, these kinds of applications are far from the context
of this book and therefore will not be covered in more detail.
An alternative technique to control the charge at the FG was presented in [62].
This technique uses a switch to electrically set the initial operating point at the FG
by short-circuiting it to a certain voltage (VFG (0)) for a given combination of inputs
(Fig. 2.12). This solution deleteriously influences the device operation. For example,
if the switch is introduced in such a way as to enable discharging of the FG, then
subsequently, during normal FGMOS operation, this switch will provide a large but
finite resistance path from the FG to the substrate of the switch. So, the gate will no
longer be floating and its DC voltage will drift towards the switch substrate voltage.
This technique has been successfully used in digital and sampled-data applications,
but it is not suitable for general purpose continuous-time analog circuits.
The best results so far have been achieved with the technique in [130]. This
technique reduces QFG down to values very close to zero. It consists of adding a
series of contacts from poly1 to the top metal layer. The top metal layer is the one
that is deposited and etched last. The contacts do not connect the FG to any other
part of the circuit, so functionally, they do not alter the device. However, the added
contacts resolve the trapped charge problem. A possible explanation for this is that the
fabrication stages prior to the deposition of the top metal layer lead to an accumulation
of trapped charge on the FGs. Before etching takes place all parts of the chip in contact
with the top metal layer are connected together. Since all the FGs are connected to the
latter via the poly1 to metal contacts, the charge trapped at each gate flows to other
parts of the die. For example, if the substrate is connected to the top metal layer then
the FGs might discharge to the substrate via these contacts. The key is that during
one of the final stages of fabrication, the FGs are not floating. Then, when the top
Poly1/Metal1
Metal1/Metal2
Poly1/Metal1
Figure 2.13 Four-input FGMOS transistor layout with gate-to-metal contacts to

eliminate the charge in a two-metal layer process
6.0e.04
5.0e.04
4.0e.04
With metal
Current - IDS
3.0e.04 contacts
2.0e.04
1.0e.04 Without
metal contacts
0.0e.04
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
–1e.04
Voltage -Vi
Figure 2.14 Comparison of measured results for two identical floating gate tran-
sistors, one with discharging contacts and another without them
metal is etched away, their floating nature is restored. The layout in Fig. 2.13 shows
how this can be done.
The poly1 gate of the FGMOS is connected to the extra region of poly1 via a metal1
bridge, which is only required here because the two poly1 regions are disconnected.
The poly1 to metal1 contact in conjunction with the metal1 to metal2 contact provides
the connection from the FG to the top metal layer in this technology, i.e. metal2.
As an example of the efficiency of this technique, Fig. 2.14 shows the experimental
results for two equally sized 1-input n-channel FGMOS transistors, one with a contact
at the FG and the other without it. The transistors were fabricated in a 0.5 μm CMOS
technology. The graph shows the drain-to-source current (IDS ) versus the voltage at
the effective input. The effective threshold voltage is around 200 mV higher in the
transistor with no contacts at the FG.
Figure 2.15 Two 2-input FGMOS transistors: one with discharging contacts (right)
and another without them (left)
(a) (b) 5.00

5.00
Vout (V) Vout1(V)
Vin
500 mV Vc(V) = 0,1,2,3,4,5 500 mV

Vc Vout
div div
Vout1
FG
0.00 0.00
0.00 Vin (V) 500 mV 5.00
div
Figure 2.16 (a) Schematic of the circuit used to check the charge accumulation
(b) Input–output characteristic of the circuit in (a) – experimental
results
Figure 2.15 shows a microphotograph of two 2-input FGMOS transistors; one

with contacts at the FG on the right and another one without them on the left.
Another example which is illustrated in Fig. 2.16 shows the performance of a
simple 2-input FGMOS inverter followed by an identical inverter designed using
conventional MOS transistors. Ignoring the effects of conventional mismatch in the
transistors, the switching threshold of both inverters will be the same in the absence
of trapped charge. Fig. 2.16(b) shows experimental results obtained when sweeping
one of the inputs between 0 and 5 V and varying the other input in steps of one volt.
The difference in thresholds is within a few millivolts. These small differences can
be attributed to conventional transistor mismatch and they do not change the circuit
VD
V1
V2 VB
V3
V4
R leak
VS
V bias
Figure 2.17 Quasi-FGMOS equivalent circuit
performance. The transfer function of the same inverter fabricated without the metal k
contact would be saturated for the whole range of input voltages.
All the fabricated circuits presented along this book were designed using this tech-
nique in several technologies over a number of runs. In all the cases the experimental
results were in agreement with the expected performance. No charge accumulation
was detected, whereas circuits fabricated without using this technique did not work
at all.
2.3.2.1 Pseudo-FGMOS: when the charge is not a problem

A modified version of the FGMOS in which the initial charge is no longer an issue is
shown in Fig. 2.17 [131].
This device, called, pseudo- or quasi-FGMOS, is an FGMOS transistor whose
gate is tied to a very large value resistor that weakly connects it to the most positive
or most negative voltage value within the circuit (depending on whether it is an n- or
a p-channel device). Figure 2.17 represents a 4-input n-channel quasi-FGMOS. For
a generic N -input quasi-FGMOS, the voltage at the gate is
N
sRleak
VFG = Ci · Vi + CGS · VS + CGD · VD (2.16)
1 + sRleak CT
i=1
where all the voltages have been referred to the bulk and CT is the total capaci-
tance seen by the gate of the quasi-FGMOS. CT will be different from CT in the
normal FGMOS as in reality Rleak will be implemented using active devices. These
devices will have parasitic capacitances that contribute to the value of CT . It follows
from eq. (2.16) that the inputs are highpass filtered with a cutoff frequency which is
inversely proportional to Rleak . Hence, as long as Rleak is kept high enough, the gate
can be effectively floating for very low frequency values so that the AC operation is
unaffected. A couple of techniques to implement Rleak were originally presented in
[132] and [133]. They are illustrated in Fig. 2.18(a) and Fig. 2.18(b) for an n-channel
FGMOS.
Both techniques suffer from the same drawback: the gate voltage must not exceed
the rail by more than the cut-in voltage of the p-n body–source junction of the nMOS
transistor realising Rleak , so that it does not become forward biased. A technique that
VD V bias+
VD
V in VB V in VB
VD
VS VS
V in VB
V bias
Vbias-
(a)
(a) V DD (b (b)
)
VS
V bias 2
S1
S2
V bias (c)
(c)
Figure 2.18 (a) Implementation of the quasi-FGMOS proposed in [131] (b) Imple-
mentation of the quasi-FGMOS proposed in [132] (c) Wide range
quasi-FGMOS
improves the input range is illustrated in Fig. 2.18(c) using an n-channel FGMOS
transistor. Rleak is implemented using an nMOS transistor with its gate and bulk
connected together. The drain/source is connected to a bias voltage Vbias , whereas
the source/drain is tied to the gate. Two switches (S1 and S2 ) are used to control the
voltage at the bulk so it never exceeds the voltage at the drain. In DC the device will
operate as a normal quasi-FGMOS. The DC value of the gate voltage will be Vbias
and the DC component of the signal will be filtered out. In AC the transistor will
work as a conventional FGMOS. For negative excursions of the input signal the
source of the transistor implementing Rleak will be the terminal connected to the
quasi-floating gate. In this case, switch S1 will be on, S2 will be off and the bulk will
be connected to it. For positive excursions of the input signal, the terminal connected
to the quasi-floating gate will act as the drain, switch S2 will be on and switch S1 will
be off. In this way the p-side of the pn junctions will always be connected to the most
negative voltages and will never be forward biased. The comparison between the gate
voltage and Vbias can be performed using a simple common-source amplifier whose
threshold can be changed through the bias current. The exact value is not important
since the voltage margin before the output resistance of the diode starts dropping is a
few hundred millivolts. The output of the latter is converted to digital values by two
CMOS inverters that control the switches. For the sake of simplicity S1 , S2 and the
two inverters have been represented as ideal components in Fig. 2.18(c). However, the
results shown later on were obtained with real nMOS switches and CMOS inverters.
Figure 2.19 compares the performance of the three versions of quasi-FGMOS
transistor. The voltage at the gate changes from 0 to 2 V. The maximum value of the
leakage current for the Rleak implementation in Fig. 2.18(c) is of the order of 10−13 A
(Fig. 2.19(a)), whereas this current can reach values above mA for gate voltages
below 0.5 V (Fig. 2.19(c)) for the topology in Fig. 2.18(a) and above 1.75 V for
the one in Fig. 2.18(b) (Fig. 2.19(b)). Besides, the value of Rleak varies over several
1.5 × 10–13
1 (a)
0.5
0
–0.5
–1
–1.5
0 0.4 0.8 1 1.2 1.6 2

DC current at the gate (A)
5 × 10–3
4 (b)
3
2 (1.75,700 μA)
1
0
–1
0 0.4 0.8 1 1.2 1.6 2
16 × 10–3
14 (c)
12
10
8
6 (0.485,1.07 mA)
4
2
0
–2
0 0.4 0.8 1 1.2 1.6 2
Voltage at the gate (V)
Figure 2.19 (a) DC current versus voltage at the gate for the implementation in
Fig. 2.18(c) (b) DC current versus voltage at the gate for the imple-
mentation in Fig. 2.18(b) (c) DC current versus voltage at the gate for
the implementation in Fig. 2.18(a)
orders of magnitude within the input range for the implementations in Fig. 2.18(a) and
Fig. 2.18(b), whereas it remains in the order of 1013 for the design in Fig. 2.18(c).
Being able to use the whole voltage range is of uttermost importance in analogue
designs that have to operate at a low voltage supply. Besides, the DC value of the
signal can be shifted through Vbias which means that every device can be biased in
the most convenient operating region.
2.4 Minimum input capacitance: implications to the total area of

the device
One disadvantage of the FGMOS transistor compared with the MOS transistor is in its
much larger area due to the added input capacitors. The size of the transistor depends
on a number of factors that are discussed in the text that follows.
AE
SES1
SES
WSE1 WSE2 WSEN AMOS
LSE
A1 A2 AN
SEL
Figure 2.20 Definition of the parameters in eq. (2.18)
The total area of the transistor (AT ) has two main components: the area of the MOS
device (AMOS ) and the area occupied by the new effective inputs, that corresponds to
the extra surface of FG outside the active area (AE ):
AT = AMOS + AE (2.17)
The component AE will vary with the number of inputs N in the FGMOS:

N
N
AE = Ai + SES1 · (N − 1) · LSE + 2 · SES WSEi
i=1 i=1
+ 2 · SES1 · SES · (N − 1) + 4 · SEL · SES (2.18)
where Ai , WSEi and LSE are the area, width and length of each effective input electrode
for i = [1, N ], respectively, and SES1, SES and SEL are dimensions given by
technological rules (Fig. 2.20). It can be proven, using eq. (2.18), that the extra area
AE will be minimised when the length and width of the input capacitors are chosen
according to the following equations:

2 · SES N
Ai
LSE = Ai , WSEi =
SES1(N − 1)
i=1 (2 · SES/SES1(N − 1)) N i=1 Ai
(2.19)
The total extra area will then be

N
N
AE =
Ai + 2 2 · SES · SES1 · (N − 1) Ai
i=1 i=1
+ 2 · SES1 · SES · (N − 1) + 4 · SEL · SES (2.20)
Although in the minimum area solution the number of inputs is also minimum,
sometimes, when matching is important it is more convenient to implement capaci-
tances of different values by shorting capacitors with the same area. In this case, the
minimum extra area required is

AE = N · A1 + 2 2 · SES · SES1 · (N − 1) · N · A1
+ 2 · SES1 · SES · (N − 1) + 4 · SEL · SES (2.21)
where A1 is the area of the unit capacitor. Matching is mostly an issue between
capacitances that are used to process signals. Generally, the ratios between these
capacitances are integer numbers, whereas the values of the capacitors which are used
to bias or shift the signal levels do not keep an integer ratio. In such a case, the
unit capacitance (capacitance that is to be replicated) should be the minimum value
capacitance used to process the signal. For example, let us consider an FGMOS
transistor with three inputs and assume that due to design requirements two of these
inputs are connected to internal circuit nodes that vary with time and the weight of
one of them is two times larger than the weight of the other. Let us also assume that
the third input is connected to a constant voltage, and the ratio between its weight and
the smaller one of the other two is 1.3. The recommended layout for this transistor
would consist of four pieces of poly2 placed on top of the poly1 that forms the FG
with relative widths (1,1,1,1.3). The weight 2 would be implemented by connecting
two capacitances of weight 1 together.
The absolute value of the minimum capacitance (C1 ) depends on the dimensions
of the MOS transistor. There are two reasons for this.
1. The parasitic capacitances CGD , CGS and CGB (that depend on the size of
the transistor) reduce the value of the weights due to the effect they have on CT
(eq. (2.3)). This has direct consequences on the effective transconductance of the
device (eq. (2.10)) and hence on the circuit speed. Therefore, during the design
process, once it has been decided about the number of device inputs as well as
its minimum acceptable transconductance and hence the minimum w1 weight, that
depends on the minimum area A1 , the implemented weight will deviate from the ideal
value by a percentage:
100 · (CGS + CGD + CGB ) 100 · (CGS + CGD + CGB )
ε≈ N = (2.22)
i=1 Ci A1 CSiO2 Ni=1 (Ai /A1 )
where CSiO2 is the capacitance per unit area of the poly2 on top of the poly1 with SiO2
in between, that forms the effective inputs and the FG, respectively (εSiO2 /tSiO2 ).
2. The second reason is related to the way in which the current in the FGMOS
transistor is affected by the ratio between CGD and CT . When a transistor is operating
in saturation region (either in weak or strong inversion), a high output resistance is
desired. However, in general, in FGMOS devices, the value of the output conductance
is dominated by the CGD /CT term in eq. (2.11). If this term is not small enough it
could cause unwanted effects such as distortion or gain loss. These effects could
be minimised by keeping VDS constant which can be achieved by using some circuit
techniques such as cascoding. However, the term CGD VDS /CT could still be a problem
because of the shift of the operating point at the FG with respect to the weighted sum
of the voltages at the effective inputs (see eq. (2.2)).
Capacitance
C2 + 2C5
CGS
C1 + 2C2/3
CGS , CGD
C1 + C2/2
vDS = constant
CGS, CGD CGD vBS = 0
C1,C3
CGB
2C5
0 Non-
Off Saturation VGS
saturation
VT VDS + VT
Figure 2.21 Capacitances CGS , CDS and CGB as a function of VGS (VDS = constant
and VBS = 0) [119]
CGD and CGS have two contributors: the overlap capacitance due to the overlap
between the gate and drain/source, respectively; and the intrinsic capacitance which
depends on the operating mode of the device. CGB has three contributors: the overlap
capacitance, which occurs between the gate and bulk at the edges of the channel;
the intrinsic charge-storage capacitance; and the parasitic capacitance between the
portion of poly1 layer outside the channel that constitutes the major part of the FG
and the substrate (proportional to AE ). The first two components are common to MOS
devices. The third component is typical of FGMOS transistors and, in general, is much
larger than the others. Figure 2.21 shows how CGD and CGS vary as the transistor
enters different operating regions. A simplified mathematical model for CGD , CGS
and CGB is shown in eq. (2.23) to eq. (2.31) [119]. This model does not predict the
smooth transitions as shown in Fig. 2.21 but is good enough for our derivations.
Weak inversion saturation

εSiO2
CGB = Cox · W · L + CGBO · L + F
· AE (2.23)
tSiO2
CGS = CGSO · W (2.24)

CGD = CGDO · W (2.25)
Strong inversion saturation
εSiO2
CGB = CGBO · L + F
· AE (2.26)
tSiO2
CGS = CGSO · W + (2/3)(Cox · W · L) (2.27)

CGD = CGDO · W (2.28)
Strong inversion ohmic

εSiO2
CGB = CGBO · L + F
· AE (2.29)
tSiO2
CGS = CGSO · W + 0.5 · Cox · W · L (2.30)

CGD = CGDO · W + 0.5 · Cox · W · L (2.31)
Parameters CGBO, CGSO and CGDO depend on the technology and account for
the overlap capacitances per unit of effective width (or length). Cox is the value of the
capacitance per unit of area in the active region (εSiO2 /tox ) and tSiO
F
2
is the thickness
of the field oxide. For example, for a transistor operating in the strong inversion
saturation region the deviation of the value of the minimum w1 weight with respect to
the ideal value will depend on the transistor dimensions and technology parameters
in the following way:
100·(CGBO·L+CGSO·W +CGDO·W +(2/3)(Cox ·W ·L)+(εSiO2 /tSiO

F )A )
E
ε≈ 2
CSiO2 AE
(2.32)
Hence, if for a device with N inputs ε has to be smaller than a certain value, e,
due to design constraints such as a minimum allowed effective transconductance, the
smallest extra area required, Amin
E would be
(CGBO · L + CGSO · W + CGDO · W + (2/3)(Cox · W · L))

Amin
E =
(0.01 · e · CSiO2 − εSiO2 /tSiO
F )
2
(2.33)
As an example, let us consider an FGMOS transistor in a typical technology with

three identical effective inputs, effective width and length W = 10 μm, L = 10 μm,
respectively, and CGBO = 3.45×10−10 F/m, CGDO = 1.38×10−10 F/m, CGSO =
1.38 × 10−10 F/m, tox = 10−8 m and tSiO F
2
= 310 × 10−10 m. If the maximum desired
variation of the minimum weight with respect to its ideal value is 15 per cent, the
minimum extra area would be 2130 μm2 .
If ε is not an issue, but the value of the output resistance, which is normally the
case, then, the term CGD /Ci has to be smaller than a certain value, k . In this situation,
a design equation that can be used to estimate the area is
CGD N · CGDO · W N · CGDO · W
≈ < k ⇒ Amin
E = (2.34)
Ci CSiO2 · AE CSiO2 · k
In order to illustrate this, let us assume the same device as in the previous exam-
ple. If the output resistance of the MOS transistor that is used to build the FGMOS is
sufficiently high, as to neglect the first term in eq. (2.11), and the effective transcon-
ductance is required to be at least 100 times larger than the output conductance, then
the minimum extra area required would be 372 μm2 .
2.5 Summary and conclusions
The FGMOS transistor can, in general, be treated as a normal MOS transistor. It can
be fabricated in any MOS technology, although for a good performance, a double
polysilicon layer is recommended. The physical models used to describe the MOS
transistor can be adapted for the FGMOS just by applying a change of variables into
the equations. Similarly, the same simulation models can be used, as long as the
simulations are set up properly.
The main disadvantages of the FGMOS are the reduction of the output resistance
and the effective transconductance, the increased area and the uncertain amount of
charge that might remain trapped at the FG during fabrication. However, with a
careful design and layout, the extra area can be minimised and the effect of the charge
made negligible. The reduction of the output resistance can be compensated for with
design techniques that will be discussed in the following chapters.
Notation
ε Deviation between the implemented and the ideal weight in

FGMOS device (eq. (2.22))
εSiO2 Dielectric constant of the SiO2
A1 Area of the unit capacitor (eq. (2.21))
AE Area occupied by new effective inputs in FGMOS (extra surface
of FG outside the active area) (eq. (2.18), Fig. 2.20)
Amin
E The smallest extra area required for an FGMOS device with N
inputs and ε (eq. (2.22)) smaller than a certain value e
(eq. (2.33))
Ai for i = [1, N ] Area of the i-th effective input electrode (Fig. 2.20)
ai ; ao ; ao Gains of VCVSs in Fig. 2.5
AMOS Area of the MOS device in an FGMOS transistor (eq. (2.17),
Fig. 2.20)
AT : AMOS + AE Total area of an FGMOS transistor (eq. (2.17), Fig. 2.20)
C1 Value of the minimum input capacitance (see Section 2.5)
CGB Gate-to-bulk parasitic capacitance
CGBO Overlap bulk capacitance per unit of length
CGD Gate-to-drain parasitic capacitance
CGDO Overlap drain capacitance per unit of width
CGS Gate-to-source parasitic capacitance
CGSO Overlap source capacitance per unit of width
Ci for i = [1, N ] i-th FGMOS input capacitance
Cox Capacitance per unit of area in the active region (εSiO2 /tox )
CSiO2 Capacitance per unit area of the poly2 on top of the poly1 with
SiO2 in between that form the effective inputs and the FG,
respectively (εSiO2 /tSiO2 ) (eq. (2.22))
CT Total capacitance seen by the FG (eq. (2.3))
CT Total capacitance seen by gate of the quasi-FGMOS in

Fig. 2.17 (different from CT in the normal FGMOS due to
Rleak )
FG Floating Gate
G Gain of the common source amplifier in Fig. 2.4(a)
GFG Gain of the common source amplifier in Fig. 2.4(b)
gds MOS transistor output conductance
(i)
gds for i = [1, 3] Mi output conductance (Fig. 2.4)
gdsF FGMOS transistor output conductance (eq. (2.11))
gm MOS transistor gate transconductance
(i)
gm for i = [1, 3] Mi transconductance (Fig. 2.4)
gmb MOS transistor bulk transconductance
gmbF FGMOS transistor bulk transconductance (eq. (2.12))
gmi for i = [1, N ] Effective i-th input transconductance of an FGMOS with N
inputs (eq. (2.10))
k Number of metals in the technology
k Minimum value of CGD /Ci required by design constraints
(eq. (2.34))
L Effective length of transistor
LSE Length of the i-th effective input electrode (the same length
has been assumed for all inputs) (Fig. 2.20, eq. (2.19))
QFG Charge trapped in the FG during fabrication
RG Resistance connected to the FG in Ramirez-Angulo’s model
(Fig. 2.5)
Rleak Quasi-FGMOS resistance implemented using active devices
(Fig. 2.17)
Rout Output resistance in common source amplifier in Fig. 2.4(a)
RoutFG Output resistance of the common source amplifier in
Fig. 2.4(b)
SEL Dimension given by technological rules – related to extra
surface of FG (Fig. 2.20)
SES Dimension given by technological rules – related to extra
SES1 Dimension given by technological rules – related to extra
tSiO2 Thickness of the SiO2 between the FG and the effective
inputs Vi
F
tSiO Thickness of the field oxide below the FG outside the active
2
area
VFG Voltage at the FG
v(fgi )op Operating point at the FG for simulation purposes (Fig. 2.7)
vi for i = [1, N ] FGMOS effective inputs
2 /f |
vin Noise power spectral density at the effective input (eq. (2.15))
effective input
2 /f |
vin Noise power spectral density at the FG (eq. (2.15))
FG
VT Threshold voltage (Table 2.1)
W Transistor effective width
WSEi for i = [1, N ] Width of the i-th effective input electrode (Fig. 2.20,
eq. (2.19))
w1 The minimum weight in an FGMOS – depends on
the minimum area A1 of its corresponding input
electrode
wi for i = [1, N ] : Ci /CT FGMOS equivalent i-th input weight
Chapter 3
FGMOS – Circuit applications and
design techniques
3.1 Introduction
This chapter explains the advantages of using an FGMOS device over a standard
MOS transistor in circuits which require low power and low voltage operation. The
first section introduces three circuit equivalents for the FGMOS device which are
then used in subsequent chapters to design more complex circuits with a higher
functionality. Several simple circuits are then presented in order to give the reader
a flavour of various new design paths that the FGMOS transistor opens, in addi-
tion to illustrating the device functionality. The design tradeoffs are also explained
together with examples. Novel, state-of-the-art building blocks and architectures can
then be developed from the presented structures. They are left to the reader to opti-
mise and refine. More advanced circuits are introduced and analysed in subsequent
chapters.
3.2 Initial design ideas: three circuit equivalents for the FGMOS
This section introduces three circuit equivalents for the FGMOS transistor: an adder;
a variable VT FET and a current multiplier, which present the device as a tunable,
controllable and flexible element. Seeing the FGMOS as one of these equivalent
blocks can help the designer to anticipate when the transistor might be useful in a
specific design.
3.2.1 FGMOS: a MOS transistor plus an adder

As eq. (2.2) shows, the FGMOS could be seen as a normal MOS transistor with a
weighted sum of voltages applied to its gate (Fig. 3.1). The weights are given by the
∑
ID C
VFG i
V1 VFG≈ ------ . V
CT i
∑ i=1
VN
Figure 3.1 Equivalent circuit for the FGMOS
ratios between each input capacitance and the total capacitance seen by the FG. Since
the addition can be performed without having to use any extra circuitry other than a
number of capacitors, the use of this device can simplify certain circuit topologies
and have very positive implications on the circuit in terms of area, noise, DR and
power consumption. Besides, simpler designs can be of crucial importance in circuits
that have to operate by pushing the limits of the technology. For example, if a cir-
cuit requiring addition has to operate at a maximum voltage supply well below the
technological limit, not having to design an analog adder block can make the design
process easier.
3.2.2 FGMOS: a MOS transistor with a controllable threshold voltage

The FGMOS device can also be seen as a MOS transistor with a controllable threshold
voltage. In order to understand this, let us rewrite the equation for the current in
an n-channel FGMOS transistor, operating in the strong inversion saturation region
(eq. (2.9))6 assuming as a first approximation that the intrinsic parasitic capacitances
are much smaller than the input capacitances:
⎡ ⎛ ⎞⎤2
⎢ ⎜ CT
N
Cj CGD CGB QFG ⎟⎥
ID = βFG ⎣ViS − ⎝ VT − VjS − VDS − VBS − ⎠⎦
Ci j=1
Ci Ci CT Ci
j=i
⎡ ⎛ ⎞⎤2
⎢ ⎜ CT CGB
N
Cj QFG ⎟⎥
≈ βFG ⎣ViS − ⎝ VT − VBS − VjS − ⎠⎦
Ci Ci j=1
Ci Ci
j=i
2
CT
= βFG ViS − VTFG
Ci
= βFG (ViS − VT )2 (3.1)
6 A similar derivation can be carried out for other operating regions.

FGMOS – Circuit applications and design techniques 45
where
2
μo Cox W Ci μo Cox W 2
βFG = = w (3.2)
2 L CT 2 L i
CT CGB Cj QFG
N
CT
VT = VT − VBS − VjS − = VTFG (3.3)
Ci Ci j=1
Ci Ci Ci
j=i
The equation shows how the FGMOS behaves as an MOS transistor with a βFG
parameter that depends on the weight wi at the effective input (eq. (3.2)), and a
threshold voltage parameter, VT (eq. (3.3)) that can be controlled electrically by
changing the voltages at the additional inputs (VjS , for j = [1, N ] ( j = i)). This is
one of the biggest advantages of the FGMOS transistor as it enables programming
of the signal levels individually in each device according to the needs of the specific
circuit. Besides, eq. (3.3) shows that it is possible to reduce the effective threshold
voltage, make it zero or even invert its sign. This makes FGMOS devices ideal for
implementation of very low voltage circuits, or circuits that require a high degree
of tunability/programmability. Hence, they can, for example, compensate for big
variations in performance due to mismatch.
3.2.3 FGMOS in weak inversion: a current multiplier

This section will explain how an FGMOS transistor biased in the weak inversion
saturation region can behave as a nonlinear circuit block and perform multiplication
of rational powers of currents. In order to understand this, let us start by considering
a MOS transistor operating in the weak inversion saturation region with a current Ii
flowing through it. From eq. (2.4), if VDS > 4UT and VBS = 0
Ii = Is eVGS /nUT (3.4)
And if the gate voltage is Vi
Ii
eVis /nUT = (3.5)
Is
A similar expression can be obtained for N equally sized MOS transistors with
currents Ij (for j = [1, N ]) and gate voltages Vj (for j = [1, N ]).
Let us now consider an FGMOS transistor with the same channel size, same
voltages, Vj for j = [1, N ], at its effective inputs and also the same voltage VS at
the source, as in the case of the N equally sized MOS transistors. Using eq. (2.5) the
current through this FGMOS transistor will be
N
Cj /CT
VjS CGD VDS QFG
ID = IS exp exp × exp (3.6)
nUT CT nUT CT nUT
j=1
The voltages Vj can be regarded as voltages generated at the gates of the N MOS
transistors by their respective drain currents Ij (for j = [1, N ]). Equation (3.6) can be
VD I’o
ID IN I1
I1 I2
I2
VN
V2 V1 VN IN
V1
N
wj
VS
I '×
o ∏ (I j )
VS j=1
Figure 3.2 Equivalent circuit for the FGMOS in weak inversion: a current multiplier
rewritten as a function of these currents:

N wj
Ij QFG CGD VDS
ID ≈ Is × × exp × exp (3.7)
Is CT nUT CT nUT
j=1
where wj = Cj /CT . Assuming that the gate-to-drain parasitic capacitance is much

smaller than the input capacitances and hence (CGD VDS /CT nUT ) 1, eq. (3.7) can
be further simplified resulting in the expected nonlinear relationship between currents:
N
ID ≈ Io × (Ij )wj (3.8)
j=1
N
(1− wj )
where Io = Is j=1
× exp(QFG /CT nUT ).
Therefore, the FGMOS transistor working in the subthreshold saturation region
could be seen as a multiplier of currents with rational or integer exponents. The
currents being multiplied are those flowing through N different MOS transistors that
have (a) the same source voltage as the FGMOS and (b) their gates voltages are the
same as the voltages at the N FGMOS effective inputs. The equivalent circuit as well
as the schematic symbol for the FGMOS operating as a current multiplier is shown
in Fig. 3.2.
3.3 Circuits applications and design techniques
The aim of this section is to illustrate the performance of FGMOS devices using very
simple circuit blocks. The design procedure is explained so that the reader can apply
it to the design of other circuits with FGMOS.
3.3.1 A cascode current mirror

One of the simplest circuits in which replacing an MOS transistor by an FGMOS can
be beneficial is current mirrors. This would be used, for example, when the power
(a) + (b) +
VLoad Iin Iout VLoad Iin Iout
- -
VDD1 Vout VDD1 Vout
M2 M4 M2 M4
V1 VD3
V1
M1 M3 M1 M3
Vbias1 Vbias2
Figure 3.3 (a) Conventional cascode mirror (b) Cascode current mirror with
FGMOS
supply voltage is one of the main design constraints. In order to show this, let us
consider the current mirrors in Fig. 3.3. Figure 3.3(b) is the FGMOS realisation of
the MOS current mirror in Fig. 3.3(a).
The bottom MOS transistors have been replaced by FGMOS devices with two
effective inputs. One of the inputs acts as the MOS gate in the conventional MOS
current mirror, whereas the other is connected to a constant bias voltage, Vbias1 for
device M1 and Vbias2 for M3. Let us assume that Vbias1 = Vbias2 = VDD , where VDD
is the maximum supply voltage. The voltage VDD1 can be written as a function of M2
gate-to-source voltage, VGS2 , and M1 drain to source voltage, V1 as
VDD1 = VGS2 + V1 (3.9)
For a proper operation of the current mirrors, the transistors should operate in
the strong inversion saturation region. Using eq. (2.8), eq. (3.9) can be rewritten as a
function of the current
!
2Iin L2
VDD1 = VT + + V1 (3.10)
μo Cox W2
where Wi and Li are the effective width and length of transistor Mi, respectively. As a
first approximation all threshold voltages are assumed to have the same value, VT .
Hence, the minimum voltage supply VDD min the current mirror would need to drive a
max
maximum input current Iin would be
!
min 2Iinmax L2
VDD = VLoad + VT + + V1 (3.11)
μo Cox W2
where VLoad is the voltage across in the biasing/input part of the circuit (Iin ).
Equation (3.11) can now be rewritten for both MOS and FGMOS current mirrors:
MOS
! ! !
min 2Iinmax L2 L1
VDD = VLoad + 2VT + + (3.12)
μo Cox W2 W1
FGMOS
! "! ! #
min CT 2Iinmax L2 CT L1
VDD = VLoad + VT + VTFG + + (3.13)
C1 μo Cox W2 C1 W1
where C1 is value of the capacitance connected to V1 , and, VTFG is the effective

threshold voltage for the FGMOS transistor (see eq. (3.3)). Substituting eq. (3.3) into
eq. (3.13) and assuming that the ratios between the parasitic capacitances and the
total capacitance as well as QFG are small enough as to be negligible, the difference
between the minimum voltage supply needed by these two circuit cells is
⎡ ! ⎤
1 2Iinmax L2
VDD ≈ ⎣VLoad + VT + ⎦
(1 + C1 /CVDD ) μo Cox W2
⎡ ! ⎤
1 Iinmax
= ⎣VLoad + VT + ⎦ (3.14)
(1 + C1 /CVDD ) β2
where CVDD is the capacitance connected to VDD and βi is the β parameter for
transistor Mi. Equation (3.14) represents how much smaller the minimum voltage
supply can be for an FGMOS current mirror than for its equivalent MOS current
mirror. Hence, the constraint for the minimum supply voltage can be less restrictive
in an FGMOS design if the capacitances ratios are adequately chosen for a given
current range and aspect ratios.
Hence, for example, if the current mirrors in Fig. 3.3 are designed in a 0.35 μm
technology, the minimum value of the power supply voltage that the MOS current
mirror needs for the correct operation is 1.15 V, whereas the FGMOS current mirror
works at 0.85 V. The values of the input capacitances used for the FGMOS current
mirror are given in Table 3.1.
Figure 3.4 shows the input–output characteristics of both the MOS and the
FGMOS current mirrors operating at a power supply voltage of 0.85 V. The figure
shows how the output current of the FGMOS current mirror perfectly follows the
Table 3.1 Design parameters for the

FGMOS current mirror in
Fig. 3.3
Device Parameter Value (fF)
M1 C1 225
M1 CVDD 175
M2 C1 225
M2 CVDD 175
1.20u
900n
Iout (A)
I out (FGMOS)
600n
300n Iout (MOS)
0.00
–200m 200m 600m 1.00
I in (μA)
Figure 3.4 Input–output characteristic of the cascode current mirrors in Fig. 3.3
with VDD = 0.85 V. The bottom line is the output current of the MOS
cascode current mirror. The top line is the output current of the FGMOS
current mirror
16.0n
12.0n
I ou t (A)
8.00n
4.00n
0.00
0.00 400n 800n 1.20u
I in (A)
Figure 3.5 Output current versus input current for the FGMOS cascode current
mirror in Fig. 3.3(b) for a mismatch of 0.194 V in threshold voltages of
the FGMOS transistors, when Vbias1 = Vbias2 = 0.85 V
input current along the whole range of input values, whereas the output of the MOS
current mirror saturates at around 140 nA.
In addition to reducing the minimum voltage requirement, FGMOS transistors
can also be beneficial for a different reason. In general, they can be very useful
in circuits fabricated in technologies that cannot guarantee a good level of matching
between adjacent devices (an example could be polysilicon thin film transistors (TFT)
technologies in which the variations between the threshold voltage of two identical
devices can be worse than 1 V). In order to illustrate this let us assume that the
difference between the threshold voltages of the bottom transistors in Fig. 3.3 is
0.194 V. Figure 3.5 shows the input–output characteristic for the FGMOS current
mirror in this situation. In the MOS circuit it would be impossible to compensate
for this mismatch by using external tuning signals. In the FGMOS, however, the
mismatch can be corrected by applying different bias signals, Vbias1 and Vbias2 , to the
1.10u
I ou t (A) 800n
500n
200n
–100n
0.00 400n 800n 1.20u
I in (A)
Figure 3.6 Output current versus input current for the FGMOS cascode current
mirror in Fig. 3.3(b) for a mismatch of 0.194 V between the thresh-
old voltages of the FGMOS transistors, when Vbias1 = 0.425 V and
Vbias2 = 0.85 V
CVDD inputs in the FGMOS pair. The mismatch is corrected because the effective
threshold voltages of FGMOS transistors are functions of the voltage values at the
extra device inputs (see eq. (3.3)). Hence, by creating a difference between the values
of the voltages at the extra inputs the final effective threshold voltages will be the
same. Figure 3.6 illustrates the input current-output current characteristic for the
FGMOS circuit with Vbias1 = 0.425 V and Vbias2 = 0.85 V. The graph shows that
the variation has been totally compensated.
This example shows a way to ‘adapt’ the input–output characteristic of an FGMOS
circuit after fabrication, without having to provoke tunnelling and/or hot electron
injection within the circuit itself [73–81,92]. The challenge is to determine the com-
pensation voltage for each transistor and design the extra circuitry required by every
specific topology.
3.3.2 Two-stage OPAMP

This section describes how to add flexibility and reduce the overall power consump-
tion in a CMOS operational amplifier by replacing some of the MOS transistors with
FGMOS devices. The disadvantages of doing this are also discussed, taking as an
example the circuit in Fig. 3.7.
The two parameters that determine the power consumed by the circuit in Fig. 3.7
are the current flowing through the devices and the voltage supply needed for the
circuit operation. A current reduction would affect negatively the performance of
the circuit in terms of unity gain bandwidth (GBW) and slew rate (SR). If the voltage
supply is decreased instead, the performance of the circuit (in terms of GBW and SR)
would not be affected as long as the values of the currents are maintained together
with the operating modes of the devices. Hence, the voltage supply would be the
best variable to alter in order to reduce the power consumption. However, the voltage
supply reduction (VDD if VSS = 0) usually results in an undesired behaviour of
the devices as a consequence of their operation in non-saturated regions. Forcing the
VDD VDD VDD
Ms1
M11 M10
Vout
Mc
V+ V–
M2 M1 Ccomp
V DD C load
Vref
M9
Ms2
Figure 3.7 Two-stage compensated MOS opamp
transistors to work in strong inversion saturation region becomes impossible for a

value of VDD under a certain limit, and the only way to achieve this is to come up
with new, innovative and complex circuit topologies. The use of FGMOS devices
can enlarge the operative range for a lower value of VDD without compromising the
complexity of the circuits. In the opamp in Fig. 3.7, the minimum voltage supply
is determined by two saturation voltages (VDSAT(M 9) + VDSAT(M 1/2) ), the threshold
voltage of a p-transistor (|VTP |) and also, indirectly, by the common mode at the input,
unless a level shifter is used. The option of adding a level shifter to change the
common mode at the input is not always a good idea since it would increase the
power consumption. Besides, the equivalent offset at the input would have an added
component due to the offset of the level shifter. Hence, the common mode voltage
indirectly sets a constraint to the minimum voltage supply if the voltage levels are
wanted to be centred around VDD /2, the input range maximised, and extra circuitry
avoided.
In order to illustrate this, the two-stage opamp in Fig. 3.7 is designed in a
0.35 μm technology for the following specifications: a power consumption smaller
than 12 μW, GBW larger than 0.8 MHz for a load of 5 pF, phase margin (PM) greater
than 60◦ , gain over 60 dB and a maximised input range. The minimum voltage supply
is in this case determined by the common mode input range that requires a minimum
value of 1.5 V. The performance is summarised in Table 3.2.
In order to reduce the voltage supply and hence the power, the opamp is redesigned
using FGMOS devices. A first possibility is to substitute the input transistors, M1 and
M2 for two-input FGMOS transistors, as shown in Fig. 3.8. The values of the input
capacitances for the new circuit are collected in Table 3.3. A summary of performance
is shown in Table 3.4.
Table 3.2 Summary of performance for the

opamp in Fig. 3.7
min
VDD 1.5 V
Ccomp 3 pF
Voltage gain 80 dB
Unity gain frequency (GBW) 1.38 MHz
Phase margin (PM) 75◦
Power 10.8 μW
Slew rate (SR) 1.2 MV/s
Input rangea 0.8 Vpp = 0.53 VDD
a The input range has been determined with the opamp con-
nected as a buffer, and a 1 kHz sinusoidal input signal,
allowing a maximum THD at the output of 1%.
VDD VDD VDD
Ms2
Ms1
M11 M10
C2 V out
C2 Mc
VDD M2
x1 M1
x2 VDD Ccomp
V+C V–
1 C1
VDD C load
V bias M9
Ms2
Ms
1
Figure 3.8 Two-stage FGMOS opamp: first version
Table 3.4 shows a 27 per cent reduction in the power consumption and almost
50 per cent improvement in the input range relative to VDD for the FGMOS cell.
However, the price to pay for this is a reduction of the gain and the GBW. The
reduction of the GBW is due to the following: the GBW of the MOS opamp in
Fig. 3.7 is given by7
gm1
GBW = (3.15)
Ccomp
7 The subscripts refer to the name of the corresponding transistor.

Table 3.3 Input capacitances for

the FGMOS transistors
in Fig. 3.8
M1 C1 89.7
M1 C2 175
M2 C1 89.7
M2 C2 175
Table 3.4 Summary of the performance for the

opamp in Fig. 3.8
min
VDD 1.1 V
Ccomp 1.5 pF
Voltage gain 62 dB
Unity gain frequency (GBW) 0.84 MHz
Power 7.9 μW
Input rangea 0.8 Vpp = 0.72 VDD
nected as a buffer, and a 1 kHz sinusoidal input signal,
In the floating gate version of the opamp, using the expression for the effective input
transconductance in eq. (2.10) together with eq. (3.15)
C1 gm1
GBWFG = (3.16)
CT Ccomp
The compensation capacitance Ccomp is two times smaller in the FGMOS version
of the opamp. This on its own would double the GBW. However the ratio C1 /CT is
always smaller than one, in this case 0.3, which explains the overall reduction of the
GBW by a factor of 1.6:
GBWFG C1 3
= · (3.17)
GBW CT 1.5
The effect on the gain is more serious. The reason for this is in equations (2.10)
and (2.11). The total DC gain of the MOS amplifier (Ao ) is
gm1 gms1
Ao = × (3.18)
gds1 + gds10 gdss1 + gdss2
And for the FGMOS opamp
C1 gm1 gms1
AoFG = × × (3.19)
CT gdsF1 + gds10 gdss1 + gdss2
The FGMOS reduces the gain because, on one hand the factor C1 /CT is smaller than
one and on the other hand, the output conductance of the input transistor is dominated
by the term CGD1 gm1 /CT (see eq. (2.11)). Hence, in general, the gain is going to be
limited to
C1 gms1
AoFG ≈ × (3.20)
CGD1 gdss1 + gdss2
unless the total capacitance is significantly increased or, design strategies such as
cascoding are used.
Let us analyse now why a smaller compensation capacitance is needed. The
increase of the output conductance in the first stage will also shift the first pole in a
non-compensated design towards a higher frequency value. This, on its own, would
be detrimental for the PM since the first pole would be closer to the second one. As a
consequence of this effect only, a higher value of the compensation capacitance would
be required. However, the reduced gain will have a completely opposite effect on the
circuit. In the case of this specific design, the reduction of the gain compensates for
the increase of the first pole frequency in terms of the PM. This is why a smaller
compensation capacitance is needed in spite of the expected higher frequency of
the first pole without compensation. Analytically, this can be explained using the
following equation for the PM [119]:

◦ GBW GBW
PM ≈ ±180 − atan − atan (3.21)
|wp1 | |wp2 |
where wp1 and wp2 are the angular frequencies of the first and second poles, respec-
tively. The frequency of the second pole is the same for both the MOS and FGMOS
designs whereas wp1 is larger in the FGMOS circuit, and thus increases the denomi-
nator of the second term in eq. (3.21). However this is compensated by the reduction
of the GBW in the FGMOS topology (eq. (3.16)).
In summary, the advantages of using FGMOS transistors at the input of the opamp
circuit are the reduction of the power consumption as well as power supply voltage
and, the increase of the input range that now goes almost rail to rail since the topology
is operating at 1.1 V. In this particular case, the addition of FGs does not increase the
area since the compensation capacitance is smaller. The disadvantages are smaller
gain and GBW.
A design of the two-stage opamp in which a further reduction of the power supply
voltage can be achieved is shown in Fig. 3.9. The idea is to shift the effective threshold
voltages (eq. (3.3)) of all those transistors whose operating modes would be affected
V DD V DD V DD
C2 V C2 C2
SS V SS V SS
M11
x11 M10
x10 xs1
Ms1
C1 C1 C1
V SS V SS
mc4
Mc4 Mc3
mc3 V SS
Mc5
mc 5
Mc2
mc 2 Mc1
mc1
V DD V DD
Rc
C comp V out
C2 C2
Vc Vc
M2
x2 M1
x1
V+ V−
C1 C1
Cloa d
V bias M9
m9
Ms2
ms 2
V SS V SS V SS
Figure 3.9 Two stage opamp with FGMOS: second version
by the reduction of the supply voltage. These are M1, M2, M10, M11 and Ms1
(Fig. 3.8). The new effective threshold voltages are shown in Table 3.5.8
The new opamp works with a power supply voltage of 0.9 V. In order to improve
the performance, cascode transistors are added. The cascode transistors compensate
for the degradation in output resistance and hence in gain caused by the (CGD gm )/CT
term added to the output conductances of the FGMOS transistors. Without cascode
transistors, the gain of the amplifier would be9

C1 C1 gml gms1
AoFG = × × (3.22)
CT 1 CT s1 gdsF1 + gdsF10 gdsFs1 + gdss2
8 The input connected to either V = V

C DD or VSS is used for this purpose.
9 The subscripts in the capacitances ratios refer to the names of the transistors they correspond to.
Table 3.5 New effective threshold volt-

ages (eq. (3.3)) for the FGMOS
transistors in Fig. 3.9
M1 & M2 VTn − 334 mV = 225 mV

M10 & M11 |VTp | − 318 mV = 335 mV
Ms1 |VTp | − 446 mV = 213 mV
Note: VTn and |VTp | will be used from now on to

distinguish between the threshold voltages of n- and
p-channel transistors, respectively. Also, their values
will be supposed to be the same for all same type
devices unless the opposite is said.
Using eq. (2.11), eq. (3.22) can be rewritten as

C1 C1 1
Ao ≈ × ×
CT 1 CT s1 (CGD /CT )1 + (CGD /CT )10 μp W10 L1 /μn L10 W1

CT
× (3.23)
CGD s1
μp , and μn refer to the mobility of p- and n-channel transistors, respectively.
Equation (3.23) shows that, unless the size of the transistors is such that the cor-
responding CGD is much smaller than the total capacitance seen by the FG (in the
order of 100 times smaller), the gain is going to be limited in value and is almost
independent of biasing conditions. However, if cascode transistors are added, the
new output conductance of each pair cascode/FGMOS is going to be gmc /gdsc times
smaller than the value given by eq. (2.11), being gmc the transconductance of the
cascode transistor and gdsc its output conductance, hence increasing the gain. This is
used in the design of this opamp. Without cascode transistors the maximum achiev-
able DC gain is in the order of 40 dB, whereas the addition of cascode transistors
increases it to 78 dB. A summary of the performance for the amplifier in Fig. 3.9 is
shown in Table 3.6. The values of the input capacitances are collected in Table 3.7.
As a summary of the above, FGMOS transistors add a number of advantages and
disadvantages to the design. The decision to use or not to use them will ultimately
depend on the design tradeoffs.
The negative points are
1. A larger compensation capacitance is needed due to the addition of the cascode
transistors in the first stage. Since the gain of the latter increases the compensation
capacitance has to be larger in order to achieve an acceptable PM.
2. The SR is reduced because of the larger compensation capacitance.
Nevertheless, the transistor also makes it possible to
1. Reduce the power supply voltage,
2. Reduce the power consumption,

opamp in Fig. 3.9
min
VDD 0.9 V
Ccomp 5 pF
Rc 400 k
Voltage gain 78 dB
Unity gain frequency (GBW) 800 kHz
Power 7.9 μW
Input rangea 0.77Vpp = 0.85VDD
nected as a buffer, and a 10 Hz sinusoidal input signal,
Table 3.7 Design parameters for

the FGMOS opamp in
Fig. 3.9
M1 & M2 C1 89
M1 & M2 C2 320
M10 & M11 C1 89
M10 & M11 C2 300
Ms1 C1 79
Ms1 C2 320
3. Increase the gain,

4. Increase the input range and
5. Adjust externally the threshold voltage of each individual device by changing
the bias voltage at the extra FGMOS input, if required. This can be used for
the compensation of process variations since the devices can be tuned externally
after fabrication. The kind of circuitry required for this purpose would depend
on the specific topology and the type of variations. As an example, an offset at
the output of the opamp in Fig. 3.9 could be corrected by applying two different
biasing voltages (VC ) to the input FGMOS transistors. Another possibility would
be to apply different biasing voltages to the input capacities C2 in transistors
M10 and M11 (in this case the opamp would have to be designed with two
disconnected FGs for these transistors). A final possibility would be to change
the biasing voltage (connected to C2 ) in transistor Ms1.
Vin
Vc Vou t
Vout1
VFG
Figure 3.10 Two-input FGMOS inverter followed by a MOS inverter
3.3.3 FGMOS programmable inverter

This section describes how the flexibility/programmability of a CMOS inverter can
be increased by replacing MOS with FGMOS transistors as shown in Fig. 3.10. The
programmable FGMOS inverter can be used in a number of interesting applications
that will be explained in subsequent chapters.
In a conventional CMOS inverter, the switching threshold Uinverter is given by

VDD − |VTp | + βn /βp × VTn
Uinverter ≈ (3.24)
1 + βn /βp
where again subscripts n and p refer to the n- channel and p-channel MOS transistors,
respectively. If both, n-channel and p-channel MOS transistors are replaced with
n-channel and p-channel 2-input FGMOS devices with common FG, and if one input
of these new devices acts as the effective input and the other is connected to a bias
voltage, the switching threshold would remain the same. This is, of course, assuming
that the channel width and length of these transistors are the same as in the MOS
implementation. However, in the new design, the value of the switching threshold
does not have any practical meaning, since it represents the value of the voltage at
which the FG and the output share the operating point (VFG = Vout1 ), but the FG
is not the input anymore. It is therefore, more appropriate to find the value of the
voltage at the effective input for which this happens. This voltage can be obtained by
combining eq. (2.2) and eq. (3.24), which yields

CT (CGDn + CGDp ) VDD − |VTp | + βn /βp × VTn
Vin(threshold) = − ×
Cin Cin 1 + βn /βp
Cc (CGSp + CGBp )
− Vc − VDD (3.25)
Cin Cin
where Cc and Cin refer to the FGMOS capacitances connected to Vc and Vin , respec-
tively. Equation (3.25) shows that it is possible to change the switching threshold of
the FGMOS inverter just by varying the value of Vc . Experimental results for this
circuit are shown in Fig. 3.11. Vout1 represents the output of the FGMOS inverter
obtained by sweeping Vin between 0 and 5 V and varying Vc in steps of one volt. Vout
is the output of an MOS inverter that is connected in series with the FGMOS inverter
5.00 5.00
Vout (V) Vout1(V)
V c (V) = 0,1,2,3,4,5
500 mV 500 mV
div div
0.00 0.00
0.00 V in (V) 500 mV 5.00

div
Figure 3.11 Input–output characteristic of inverter in Fig. 3.10
and whose transistors have the same aspect ratios as the corresponding devices in the
FGMOS implementation. The purpose of the second inverter is twofold: on one hand
it regenerates the digital levels; on the other hand, it allows to compare the perfor-
mances of the MOS and FGMOS inverters as well as to estimate the matching. This
is next explained by showing how this cell can be used to approximately determine
the experimental value of QFG .
Ideally, as both inverters have the same aspect ratios their switching thresholds
should also be the same. However, if a difference exists between them it would give
us an idea of how much the threshold voltages of the FGMOS transistors vary with
respect to the threshold voltages of their MOS counterparts. This would also be an
indication of the possible amount of charge that remains trapped after fabrication.
Assuming that other sources of error are negligible10 , the variation of Uinverter
(eq. (3.24)) caused by threshold voltage variations would be
1
Uinverter ≈ × −|VTp | + βn /βp × VTn (3.26)
1 + βn /βp
Equation (3.26) is valid for both, the MOS and the FGMOS inverter. However, again,
in the case of the FGMOS inverter it should be rewritten referred to the effective
10 This is justified by the fact that the main extra source of mismatch between FGMOS and MOS is
QFG . If QFG is erased the mismatch would be of the same order of magnitude as in two identical MOS
devices.
input, which gives

CT Cc
Vin(threshold) = · (Uinverter + Uinverter ) − Vc
Cin Cin
(CGDn + CGDp )Uinverter (CGSp + CGBp )
− − VDD (3.27)
Cin Cin
From eq. (3.27)

Cin Cc CGDn + CGDp
Uinverter = · Vin(threshold) + Vc − Uinverter · 1 −
CT Cin CT
(CGSp + CGBp )
+ VDD (3.28)
Cin
Hence, if QFG is supposed to be the only source of threshold voltage variation in
eq. (3.26) its value could be obtained from the output of this circuit, just by measuring
the value of the voltage at the effective input when Vout = Vin = Uinverter and using
these values in eq. (3.28):

QFG Cin Cc CGDn + CGDp
≈ × Vin(threshold) + Vc − Uinverter × 1 −
CT CT Cin CT
(CGSp + CGBp )
+ VDD (3.29)
Cin
As it was already mentioned in Chapter 2, in a design using metal contacts on the
FG, the experimental value obtained for eq. (3.29) turns out to be of the same order
of magnitude as the expected variations due to mismatch, so it is difficult to know, if
this is due to (a) the simulated values of the couplings capacitances used to estimate
QFG , which are different from the real ones; (b) mismatch between MOS transistors
in the two inverters or (c) the still remaining charge. In any case, the variation is not
critical and can be easily taken into account during the design process. It can be just
treated as either normal mismatch or process variations.
3.3.4 An offset compensated FGMOS comparator

This section shows how the use of FGMOS transistors in a comparator circuit can (a)
reduce the minimum required power supply voltage; (b) increase the input range and
(c) compensate for offset variations.
Figure 3.12 shows a simplified schematic of the fully differential comparator.
It consists of two n-channel FGMOS input transistors (M1 and M2 ) loaded by two cross
coupled PMOS transistors (M3 and M4 ) that act as regenerative comparator elements.
Each FGMOS device has three input gates. One of them represents the effective input
and is connected to Vin± (the differential input is given as Vind = Vin+ − Vin− ). A second
input is connected to the comparator’s clock Vclk . This input plays a double role: on
one hand, when a comparison is being performed, the high value of the clock signal
reduces the effective threshold voltage and so, the transistors turn ON even for low
M3 M
4
–
Vout Vout+
+ –
V off Voff
V clk M1 M2 Vclk
–
V +in Vin
gnd
Figure 3.12 FGMOS comparator schematic
values of the input signals. On the other hand, transistors can be turned OFF when
this input is low which will reset the comparator. Analytically this can be written as

+ CT Cclk Cc
Input transistors ON if : Vin > VTn − Vclk − Voff (3.30)
Cin CT CT

CT Cc
Reset if : Vin+ < VTn − Voff (3.31)
Cin CT
where Cin is the value of the capacitance connected to the effective input, Cclk is
the value of the capacitance connected to the clock signal and Cc is the value of the
capacitance connected to the offset compensation signal. The third input is therefore
used for offset compensation. For the sake of simplicity, Voff has been assumed to
+ −
take the same value for both differential branches (Voff = Voff = Voff ), which is
true in the absence of offset. However, in reality it will take a different value for each
branch, which will consequently give rise to two different equations for eq. (3.30) as
well as for eq. (3.31). This will be explained in more detail later on. Equations (3.30)
and (3.31) show that the adequate choice of the capacitances ratios makes it possible
to turn the input transistors ON for the whole range of input values when Vclk is high
or, to turn them off and reset the comparator when Vclk is low.
3.3.4.1 The offset compensation

The main sources of offset in the comparator in Fig. 3.12 are explained in this section
along with the strategies used for its compensation.
1. Mismatch in the threshold voltage of the input transistors: One of the main
drawbacks of FGMOS devices is that any variation of the threshold voltage in the MOS
transistor used to build the FGMOS is multiplied by the term CT /Cin at the effective
input. As the term CT /Cin is always greater than one, the effects of the equivalent
variation will always be greater than in a normal MOS device. This is critical in this
comparator topology since the mismatch between FGMOS input transistors affects
V DD vdd
vbia s
S3 S4
phi2 phi2
M3 M4
S5 S6
S1 phi2_b phi2_ b S2
phi1 Vout− Vout+ phi1
+ –
V off V off
V clk M1 M2 V clk
–
V in + V in
gnd
Figure 3.13 FGMOS comparator with offset compensation
the offset directly. Hence, if other sources of mismatch are ignored, the offset at the
input caused by the variation between the threshold voltages of transistors M1 and
M2 , VT = (VTn2 − VTn1 ) (subscript i refers to device Mi ), is
CT CT
voff VT = VT = (VTn2 − VTn1 ) (3.32)
Cin Cin
In order to compensate for this and other sources of offset the comparator in
Fig. 3.12 is modified as shown in Fig. 3.13. The basic operation of the compara-
tor is performed in two phases: offset compensation phase and comparison phase
(Fig. 3.14). During the offset compensation phase switches S5 and S6 are opened
while S1, S2, S3, S4 are closed. Switches S3 and S4 connect the gate of the load
PMOS transistors to a constant voltage, Vbias , which makes them operate as current
sources. As a first approach, ignoring the effects of mismatch between these two
devices, the current flowing through them is
βp
Ibias = (VDD − Vbias − |VTp |)2 (3.33)
2
Switches S1 and S2 establish a negative feedback by connecting one input of each

of input FGMOS transistors with their respective drains (outputs of the comparator).
The same value of the voltage, Vcm , is applied to Vin+ and Vin− during this phase.
Assuming that the threshold voltages of both input transistors are different (VTn1 for
+ −
transistor M1 and VTn2 for transistor M2 ), the voltage generated at Vout = Voff and
offset
transition
compensation comparison
phase phase
1 2 3
phi1
phi2
1: 2: 3:
• S1 -S 2 o n • S1- S 2 off • S1 -S 2 off
• S3 -S 4 o n • S3- S 4 on • S3 -S 4 off
• S5 -S 6 o ff • S5- S 6 off • S5 -S 6 on
Figure 3.14 Timing diagram for the comparator in Fig. 3.13
− +
Vout = Voff by the negative feedback is
!
− CT 2Ibias Cin Cclk
Voff = VTn2 + − Vcm − Vclk (3.34)
Cc βn Cc Cc
!
+ CT 2Ibias Cin Cclk
Voff = VTn1 + − Vcm − Vclk (3.35)
Cc βn Cc Cc
During the comparison phase S1, S2, S3 and S4 are open (S1 and S2 must open
− +
slightly before S3 and S4 to avoid variation of Voff and Voff during the transition),
while S5 and S6 are closed. A positive feedback is thus created between the PMOS
couple. A differential input voltage is applied to the input transistors. A small capac-
− +
itance connected at nodes Voff and Voff helps maintain the value of the voltages
sampled during the compensation phase. Thus, based on (3.34) and (3.35), the voltage
at the FGs of transistors M1 and M2 (VFGI and VFG2 , respectively) will be
!
Cin + 2Ibias
VFG1 = (V − Vcm ) + VTn1 + (3.36)
CT in βn
!
Cin − 2Ibias
VFG2 = (V − Vcm ) + VTn2 + (3.37)
CT in βn
Hence, the effective differential signal seen by the gates is
Cin +
VFG1 − VFG2 = (V − Vin− ) + (VTn1 − VTn2 ) (3.38)
CT in
And dividing by Cin /CT in order to refer it to the effective input gives
(effective) CT
Vind = (Vin+ − Vin− ) + (VTn1 − VTn2 )
Cin
= (Vin+ − Vin− ) − voff V T (3.39)
In this way, the main source of offset is automatically subtracted from the effective
input.
2. Mismatch in the input capacitances: The compensation mechanism described

above also compensates for any mismatch in the values of input capacitances, as long
as the common mode of input signals is also Vcm during the comparison phase.
3. Feedthrough and effect of the parasitic capacitance CGD : During the com-
− +
parison phase the voltages at nodes Voff and Voff can vary with respect to the values
they had at the end of the compensation phase. This could happen, for example, due
to the feedthrough. In this case eq. (3.35) and eq. (3.36) are rewritten including these
variations as
!
Cin + 2Ibias Cc +
VFG1 = (Vin − Vcm ) + VTn + − Voff (3.40)
CT βn CT
!
Cin − Ibias Cc −
VFG2 = (Vin − Vcm ) + VTn + − Voff (3.41)
CT βn CT
+ −
where the terms Voff and Voff account for them. The same threshold voltage is
assumed for both input transistors in order to simplify derivations and also because
the case with different threshold voltages has been previously discussed. Also, in the
derivations above it has been assumed that the term (CGD /CT )VD , in the expressions
for the FGs voltages, is negligible, which is true only if CT
CGD . However, if the
total capacitance is made too big, it will have a detrimental effect on both the area and
speed of the circuit, and hence it will not represent a realistic case. Therefore, if the
term corresponding to the parasitic coupling is included in previous derivations, the
final expressions for (3.35) and (3.36), are
!
Cin + 2Ibias Cc +
VFG1 = (V − Vcm ) + VTn + − Voff
CT in βn CT
CGD − −
+ [V − Vout(compensation) ] (3.42)
CT out(comparison)
!
Cin $ − % 2Ibias Cc −
VFG2 = Vin − Vcm + VTn + − Voff
CT βn CT
CGD + +
+ [V − Vout(compensation) ] (3.43)
CT out(comparison)
This on its own generates an equivalent input referred offset given by
voff (CGD /feedthrough)

CGD − +
= [V − Vout(comparison) ]
Cin out(comparison)
CGD − + Cc + −
− [V − Vout(compensation) ]− (Voff − Voff )
Cin out(compensation) Cin
(3.44)
In general, the second and the third terms in (3.44) are negligible. The second
term is proportional to a small variation, with a proportionality factor considerably
smaller than one. In the third term, the proportionality factor Cc /Cin does not need
to be much smaller than one, since the variation due to the feedthrough is small. The
first term, however, dominates as its absolute value is proportional to VDD , that is,
(CGD /Cin )VDD , which is much larger, and thus, a compensation mechanism for this
term is required. The compensation is performed by adding a small extra input, Cf ,
of value comparable with CGD to M1 and M2 and applying a positive feedback by
connecting it to the transistors opposite outputs (see Fig. 3.13). In this way, from
(3.44) the new equivalent offset at the input is
(CGD − Cf ) −
voff (cGD/feedthrough ) ∼
= +
[Vout(comparison) − Vout(comparison) ] (3.45)
Cin
And this term can be made very close to zero, as proven by the design simulation
results.
4. Offset due to different threshold voltages in the PMOS: Mismatch between the
threshold voltages in the top PMOS transistors will result in different Ibias currents,
+ −
Ibias and Ibias , flowing through the positive (M1 and M3 ) and negative (M2 and M4 )
comparator branches during the compensation phase. This causes an equivalent offset
at the input:
& '
+ −
CT 2Ibias − 2Ibias
voff (Ibias ) = √ (3.46)
Cin βn
This term, however, is much smaller than the previously discussed sources of offset
and can be controlled by making the PMOS transistors sufficiently large.
In order to illustrate the performance of this cell, the comparator is designed
in a 0.35 μm process with threshold voltages around 0.55 V. It operates at
0.9 V supply voltage. The power consumption measured at 11 MHz is under
6.5 μW, which compares favourably with other LP comparator realisations.
The equivalent offset at the input, determined with Monte Carlo simulations is
5 mV with 3σ standard deviations of device variation. The operation is shown
in Fig. 3.15.
V out+
V out+/V out- (V )
700m
300m
V out -
–100m
xxx V in+
505m
V in+/V in- (V)
500m
V in-
495m
phi2
xxx
phi1
2.0u 3.0u 4.0u 5.0u 6.0u

time (s)
Figure 3.15 Comparator transient simulation
3.3.5 FGMOS D/A converters

This section illustrates how to use the FGMOS transistors to realise digital to analog
conversion. The circuits presented here offer innovative techniques for facilitating
D/A conversion by simplifying the topologies and hence reducing the required power
supply voltage. It is left to the reader to further develop these ideas and produce novel
LP/LV D/A converters.
3.3.5.1 Converter design 1

Figure 3.16 shows the schematic of a very simple D/A converter based on the operation
of the FGMOS transistor M1. M1 has N + 2 inputs, where N is the number of bits. N
of the inputs are binary weighted. The other two inputs are weighted accordingly to
keep transistors biased in the right operating region. The converter works as follows:
if it is assumed that both n-channel devices have the same channel size, the p-channel
transistors have the same dimensions and all four of them are operating in the strong
inversion saturation region, then, using eq. (2.8)
2
Cin (n−1)
N
βn Cc
IM1 = 2 Vi + VDD − VTn (3.47)
2 CT CT
n=1
βn
IM4 = (Vout − VTn )2 (3.48)
2
where IM1 and IM4 are the currents flowing through transistors M1 and M4 respec-
tively, Cin is the minimum value of the binary weighted input capacitances and Cc is
a biasing capacitance connected to Vc = VDD . As a first approximation the parasitics
have been neglected from the equation for IM1 . Transistor M2 mirrors the current
IM1 to transistor M3, which makes IM1 = IM4 . Hence, the voltage generated at the
VDD VDD
M2 M3
VD1
V1 I I
M1 M1 M4 M4
VN Vout
Vc=VDD
Vr=gnd
Figure 3.16 N-bit D/A converter based on the FGMOS
output is
Cin (n−1)
N
Cc
Vout = 2 Vi + VDD (3.49)
CT CT
n=1
Equation (3.49) is the function of a D/A converter if the inputs Vi are digital
values [134].
Let us now explain the need and function of the biasing capacitances Cc and
Cr 11 . The previous derivation has assumed that the four transistors are biased in the
strong inversion saturation region. In order to achieve that, the following boundary
conditions have to be maintained for transistors M1 and M4 (see Table 2.1):
Transistor M1
Cc
VDD > VTn (3.50)
CT

Cin (n−1) Cc
N
VD1 > 2 + VDD − VTn (3.51)
CT CT
n=1
Transistor M4
Vout > VTn (3.52)
Also, for the p-channel transistors to operate in the strong inversion saturation
region:
Transistor M2
VDD − VD1 > |VTp | (3.53)
Transistor M3
Vout < VD1 + |VTp | (3.54)
11 C is the value of the input capacitance connected to V = ground in Fig. 3.16.

r r
Equation (3.50) shows that by having an extra input capacitance Cc connected to VDD
it is possible to shift the voltage operating point at the gate of M1 by (Cc /CT )VDD .
In this way, the transistor can be operating in the strong inversion saturation region
even with the all zero digital word connected to the input. However, from eq. (3.54),
in order to keep M3 in the saturation region, the voltage Vout (eq. (3.49)) has to remain
below

VDD + ( βn /βp )VTn
Vout < $ % (3.55)
1 + βn /βp
where βp is the β parameter of transistors M2 and M3.
Hence, for the highest digital word, or what is the same, the most restrictive
condition

Cr βn /βp VTn
>$ % 1− (3.56)
CT 1 + βn /βp VDD
where as the first approximation the effect of the parasitics has been neglected thus
yielding a value for the total capacitance

N
CT = Cr + Cc + Cin 2(n−1) (3.57)
n=1
Equation (3.56) shows how in order to keep transistor M3 saturated for all input
combinations Cr has to be different from zero. Rewriting now eq. (3.50), the boundary
conditions for Cr and Cc can be established:

Cin (n−1)
N
βn /βp VTn Cr VTn
$ % 1− < <1− 2 − (3.58)
1 + βn /βp V DD C T CT
n=1
VDD

Cin (n−1)
N
VTn Cc βn /βp VTn
< <1− 2 −$ % 1− (3.59)
VDD CT CT
n=1
1 + βn /βp VDD
Hence, for example, for βn /βp = 0.1, if the n-channel transistor threshold voltage
is 0.5 V, and VDD = 1 V, the required values of the compensating capacitances,
normalised to the total capacitance, would be
Cin (n−1)
N
Cr
0.045 < < 0.5 − 2 (3.60)
CT CT
n=1
Cin (n−1)
N
Cc
0.5 < < 0.9545 − 2 (3.61)
CT CT
n=1
In order to maximise the step size for a given number of bits, the ratio Cin /CT should
be maximum, and hence, the minimum values of the compensating capacitances that
fulfil the boundary conditions should be chosen. If Cc /CT = 0.5, and for example
FGMOS DAC 1
Boundary condition (surfaces) for Cr
N = 8bits
0.6
0.4
Cr /CT
0.2
0
0.6
0.7
0.8
0.9 2
1
1.1 1.5
× 10–3
1.2 1
Volume between the two surfaces 1.3
0.5 Cin /CT
satisfies the boundary condition V DD 1.41.5
0
Figure 3.17 Boundary condition for Cr in the 8-bit DAC (Fig. 3.16)
500m
400m
300m
Vout (V)
200m
100m Transistor M4 in weak inversion
0.00
0.00 400u 800u 1.20m
time (s)
Figure 3.18 Output voltage of a 4-bit converter (Fig. 3.16)
(n−1) = 0.45, then for 0.045 < C /C < 0.05 both equations (3.60)
(Cin /CT ) N n=1 2 r T
and (3.61) would be met.
Figure 3.17 illustrates the design tradeoff represented by eq. (3.58). A value of
Cr /CT that meets all the design constraints has to be contained between the two
surfaces. Once that value is chosen, Cc /CT can be found from eq. (3.57).
An example of the design of this circuit is shown in the following. Figure 3.18
shows the output of a 4-bit converter designed using a 0.35 μm technology with
VDD = 1 V, when the input changes in time from the maximum to the minimum
digital word (starting at time = 500 μs). It can be seen how for low values of the
digital words the circuit does not have enough time to realise the conversion properly
and the output appears distorted. This happens because for low values of the input,
the output transistor M4 is in weak inversion.
Transistor M3 in the triode region

900m
600m
Vout (V)
300m
0.00
0.00 400u 800u 1.20m
time (s)
Figure 3.19 Output voltage of a 4-bit converter with Cc added (Fig. 3.16)
800m
600m
Vout (V)
400m
200m
0.00
0.00 400u 800u 1.20m
time (s)
Figure 3.20 Output voltage of a 4-bit converter with Cc and Cr added (Fig. 3.16)
Figure 3.19 illustrates the performance when Cc is added. It can be observed how
for high input values the output is saturated. This is because transistor M3 enters the
triode region and therefore the current mirror does not operate properly anymore.
Finally, the operation of the same converter with the two biasing capacitances
is shown in Fig. 3.20. The figure shows how the converter is linear for the whole
operating range.
Examples of the operation of a 6-bit and an 8-bit converter are also shown in
Fig. 3.21 and Fig. 3.22. The performance of these blocks is summarised in Table 3.8.
The Integral Non Linearity (INL) and Differential Non Linearity (DNL) of the 6-bit
A/D converter is illustrated in Fig. 3.23.
Also, Monte Carlo simulations were run for both designs. They showed that the
main problem this topology might have would be in the form of an output offset. This
could be compensated just by changing the value of the bias voltage connected to Cc .
Another possibility would be to design some kind of self-compensation strategy,
similar to the one used for the comparator described in the previous section.
These designs are not optimised and could be improved by, for example, using
cascode transistors. They would increase the output resistances of the devices, and
this would allow the use of a smaller Cin capacitance in the FGMOS device and would
improve the speed and linearity. The disadvantage would be a more limited output
range.
600m
500m
Vout (V)
400m
300m
400u 600u 800u 1.0m 1.2m 1.4m 1.6m
time (s)
Figure 3.21 Output voltage of a 6-bit converter (Fig. 3.16)
580m
560m
540m
520m
500m
Vout (V)
480m
460m
440m
420m
400m
380m
400u 700u 1.00m 1.30m 1.60m 1.90m
time (s)
Figure 3.22 Output voltage of an 8-bit converter (Fig. 3.16)

converter in Fig. 3.16
6-bit 8-bit
LSB 3 mV 0.75 mV
Non linearity and noise <1LSB <1LSB
Maximum sampling rate 11 MS/s 11 MS/s
Average power consumption 0.5 μW 0.5 μW
1 INL
LSB 0.5
–0.5
–1
0 10 20 30 40 50 60
code
DNL
1
0.5
LSB
–0.5
–1
0 10 20 30 40 50 60
code
Figure 3.23 Integral Non Linearity (INL) and Differential Non Linearity (DNL) for
the 6-bit DAC (Fig. 3.16)
Finally, for very low voltage and low speed applications the converter could
also operate with the transistors biased in the weak inversion saturation region. The
advantage of doing this would be a lower power consumption and a lower area due
to the fact that the required biasing capacitances would be much smaller. The main
disadvantage would be a much lower speed.
3.3.5.2 Converter design 2

Another example of D/A converter based on the FGMOS transistor is shown in
Fig. 3.24. In this topology, the circuit operation is based on the transistor M1 working
in the strong inversion ohmic region. Again, a biasing capacitor, Cc , connected to
VDD is used to bias the input transistor in the strong inversion region. The other input
capacitances are binary weighted. The opamp fixes the voltage at the drain of M1 to
Vref . Hence, the current flowing through M1 and M2 is given by

N
Cin Cc
IM1 = βn Vref 2(n−1) Vi + βn VDD − VTn Vref (3.62)
CT CT
n=1
And again if the inputs are digital voltages, this represents the function of a D/A
converter. This current is mirrored to an output branch where it can be converted
back to a voltage just by using a resistance. The latter can be implemented using
active components, although, in this example, a passive resistor has been considered
V DD V DD
M2 M3
OPAMP
x + I out R1
V1
VN M1
VDD Vref
Figure 3.24 N-bit D/A converter based on an FGMOS transistor operating in the
ohmic region
for the sake of simplicity. The second term in eq. (3.62) represents an offset at the
output. The offset can be cancelled out by using, for example, a differential topology.
Another way to compensate for it is by using a tuning technique which either adds a
current at the output, or connects the biasing capacitor to a variable bias voltage.
The performance of 6-bit and 8-bit D/A converters (as in Fig. 3.24) is shown in
Fig. 3.25. The average power consumption of this converter is around 10 μW. It is
higher than the power consumption of the converter described in the previous section
due to the additional power required by the opamp. Also, the voltage supply used
is 1.5 V as opposed to 1 V in the previous topology. For both topologies the total
capacitance is around 7 pF. This converter is again designed in a 0.35 μm technology
with a value of VTn for the n-channel devices of around 0.5 V. The maximum sampling
rate for the converter in Fig. 3.24 is 2 Ms/s.
3.3.6 A programmable switched-current floating gate MOS cell

This section explains how to use the FGMOS transistor in a switched-current (SI)
memory cell as a means to increase its programmability capability.
SI circuits perform discrete time processing with analog circuits. Applications
include filters and various types of ADC and DAC converters. The constituent parts
of these circuits, SI cells, may be designed to have a gain, α, such that the z-domain
transfer function of the cell is
iout = α × z −1/2 × iin (3.63)
The gains of SI cells determine the overall transfer function of an SI circuit. Generally,
the gain coefficients within SI cells are created by using current mirrors in various
(a) 4.0u
3.0u
Iout (A)
2.0u
1.0u
0.0
400u 700u 1.0m 1.3m 1.6m 1.9m
time (s)
(b) 4.0u
3.0u
Iout (A)
2.0u
1.0u
0.0
–1.0u
0.0 1.0m 2.0m 3.0m 4.0m 5.0m
time (s)
Figure 3.25 Output current for the converter in Fig. 3.24: (a) 6 bits (b) 8 bits
forms. In a first generation cell, Fig. 3.26(a), the current mirror is part of the cell.
In a second generation cell, a single transistor is used for both input and output
currents so a current mirror after the output is required to provide the necessary gain,
as shown in Fig. 3.26(b) [135]. It is often useful, however, to adjust coefficients
within switched current circuits after manufacture, to program them. This allows the
circuit characteristics to change over time, for example, to provide adaptive filtering
or to tune the circuit accounting for device mismatch and process variation.
A number of possible programming schemes are in the literature [136–142].
Programmability may be achieved with an array of different gain circuits that are
switched between [136–138]. A discrete number of gains are possible by switching
in and out parts of the array. This technique, however suffers from the limited size of
the array. To increase the number of discrete values the gain may take, or resolution,
requires increasingly more circuitry. Therefore this technique is only useful for a
considerable range of gain and continuous programmability. A four quadrant multi-
plier built into a current cell [139] and a separate multiplier [140] after the SI cell
are both found in the literature. However a multiplier is required for every SI cell
(a) (b)
J α.J J α.J
iin i out i in iout α.iout
1: α 1: α
Figure 3.26 First generation SI cell (a) and second generation cell (b)
Iα
iin i out
-gm α .g m
Figure 3.27 A first generation switched transconductance cell
V1 V2 V3
iin T2 T3 T4
S1 S2
S3 iout
T1
Figure 3.28 The PSIFG cell

thus requiring considerable circuitry for each cell. Finally, transconductance ampli-
fiers can be used to program a circuit [141,142]. An example is the first generation
switched transconductance cell, Fig. 3.27, where the gain, α, is varied by current Iα .
This solution requires considerable circuitry, with an amplifier for each cell, and, the
programmability is of limited range. All these techniques either offer a limited reso-
lution of gain or require considerable circuitry per current cell for programmability.
In low power applications, where programmability is necessary, the large amounts
of circuitry consume substantial power. This is undesirable. A simple way to add
programmability without increasing the circuitry and hence reducing the power is
by using the FGMOS transistor in a configuration that from now on will be called
Programmable Switched Current Floating Gate Cell (PSIFG).
Figure 3.28 shows the basic PSIFG cell. T1 provides the sample and hold element
necessary for SI cells while T2 and T3 form a floating gate current mirror whose
V1
V2
V3
T2 T3 T4 T5 T6 T7
i in1 iin2
S7 S8 S1 S2 S4 S5 S10 S11
iout1 iout2
S3 S6
T1 T8
S9 S12
Figure 3.29 The differential PSIFG cell
current gain is programmed by the bias voltages, V1 and V2 . Switches S1 and S3 are
closed and S2 open during the sampling phase of the clock signal. The voltage on the
gate of T1 is held during the second phase when S1 and S3 are opened and S1 closed.
This clocking strategy is similar to a second generation SI cell. During the hold phase,
T1 drain current is steered through the floating gate current mirror. Assuming two-
input (V1 and V2 ) FGMOS devices, with input capacitances C1 and C2 (for V1 and
V2 , respectively), biased in the weak inversion saturation region, applying eqs. (2.5)
and (3.63), neglecting parasitics as a first approximation, the overall transfer function
for the circuit in Fig. 3.28 is

−1/2 k(V2 − V1 )
iout = z × iin × exp (3.64)
ηUT
where k = C2 /(C2 + C1 ). Thus, by altering V1 and V2 the gain of the cell may be
changed.
The basic PSIFG cell may be extended to a differential version, solid lines of
Fig. 3.29. The cell is formed from two basic PSIFC cells: T1 , T2 and T3 form one cell
and T6 , T7 and T8 form another. Transistors T4 and T5 extend the respective floating
gate current mirrors giving the differential PSIFG cell its principle advantage: the gain
may be both positive and negative, unlike the simple PSIFG cell with only positive
gain. This may be observed in the mathematics, where the expressions for the output
currents, iout1 and iout2 , are

k(V2 − V1 ) k(V3 − V1 )
iout1 = z −1/2 × iin1 × exp + z 1/2 × iin2 · exp
nUT nUT

−1/2 k(V3 − V1 ) 1/2 k(V2 − V1 )
iout2 = z × iin1 × exp + z × iin2 · exp
nUT nUT
(3.65)
x(n) x(n–1) x(n–2)

z –1 z –1
a0 a1 a2
y(n)
Figure 3.30 Block diagram of a second-order FIR filter
iin–
x1 + + +
iin+ x1 x1
–
PSIFG PSIFG
+ + +
xA xA
xA – – –
iout–
iout+
Figure 3.31 Realisation of the filter from PSIFG cells

Hence the differential output, iout , is given by

−1/2 k(V2 − V1 )
iout = iout1 − iout2 = z × (iin1 − iin2 ) × exp
nUT

k(V3 − V1 )
− exp (3.66)
nUT
Thus, when V2 > V3 the cell gain is positive and when V2 < V3 it is negative. Also,
two extensions can be added to the differential PSIFG cell as shown by the dotted
sections of Fig. 3.29: additional outputs may be produced by extending the floating
gate current mirror; and an extra switched circuit may be added at the inputs to create a
full unit delay, z −1 from the z −1/2 already present. This solution for programmable SI
cells is advantageous as it does not require any form of transconductance or operational
amplifier for each cell, reducing power consumption and it provides a continuous gain
factor unlike array switching.
The concept of PSIFG is illustrated in the following with the design of a second
order FIR filter. The equation for the filter is
y(n) = a0 x(n) + a1 x(n − 1) + a2 x(n − 2) (3.67)
This may be seen in Fig. 3.30 in the form of a block diagram. Two PSIFG cells are
required, both with a full unit delay and one with a double set of outputs, Fig. 3.31.
In addition to the SI cells, a floating gate current mirror is used at the input of the
system to provide the coefficient a0 , without any delay added. In the z domain the
1 3
0.8 7 2
4 6
0.6
8 11 10
0.4 12
Imaginary Part
0.2
1 5 9 9 5 1
0
–0.2
12
–0.4 10
8 11
–0.6 4 6
–0.8 7 2
–1 3
–1 –0.5 0 0.5 1
Real Part
Figure 3.32 Zero placement for all filter setups
transfer function for a second-order filter is

Y (z)
T (z) = = a0 + a1 z −1 + a2 z −2 (3.68)
X (z)
Thus, for a2 , a1 , a0 to be real, the roots of T(z) are either a complex conjugate pair
or both real, the roots of T(z) being the zeros for the system. For a pair of complex
conjugate roots of radii r and at angles ±ϕ, the coefficients are
a1 = −2a0 rcos(ϕ) (3.69)
a2 = a0 r 2 (3.70)
where a0 is chosen as a suitable scaling factor given the range of gains from the
floating gate current mirrors. The two poles of the system are located at the origin
making the filter inherently stable.
To illustrate the filter performance, a range of values for r and ϕ are chosen,
r = 0.5, 0.8 and 1, ϕ = 45, 90, 135 degrees. In addition three points for the real-
valued zeros on the real axis are chosen, with zero pairs at 1, −1; 0.8, −0.8; and 0.5,
−0.5. All the zero positions may be seen in Fig. 3.32. The numbers correspond to
each pair of zeros, for identification.
With a0 set to 0.5, Table 3.9 shows the values of a1 and a2 where the zeros position
number corresponds to the numbers in Fig. 3.32.
The filter is designed in a 0.35 μm CMOS process and simulated using a 1 V
power supply. The clock frequency is 33.3 kHz. Figure 3.33 is a timing diagram for
the various clocks signals. Switches controlled by ϕ1 and ϕ4 , S1 , S2 , S4 , S5 , S7 , S8 ,
S10 and S11 of Fig. 3.29, are single NMOS FETs. The other switches have dummy
transistors included, of half the width of the switching transistor. Dummy switches
Table 3.9 Values of the coefficients corresponding to the zeros in Fig. 3.32
Zero position number Radius (r) ϕ a1 a2
1 1 0 (and 180) 0 0.5

2 1 45 −0.707 0.5
3 1 90 0 0.5
4 1 135 0.707 0.5
5 0.8 0 (and 180) 0 0.32
6 0.8 45 −0.566 0.32
7 0.8 90 0 0.32
8 0.8 135 0.566 0.32
9 0.5 0 (and 180) 0 0.125
10 0.5 45 −0.354 0.125
11 0.5 90 0 0.125
12 0.5 135 0.354 0.125
0 13 14 15 28 29 30 μs
φ1
φ2
φ2_dummy
φ4
φ3
φ3_dummy
Figure 3.33 Timing diagram for the simulations
switch slightly after the main switch, so as to not interfere with its operation. Signal
ϕ2 controls switches S9 and S12 , signal ϕ3 controls switches S3 and S6 .
Transfer functions for the FIR filter with the various coefficients are shown in
Fig. 3.34. The figure shows transfer functions for the values of coefficients chosen.
The x axis is frequency in kHz and the y axis is amplitude in nA. The ideal response
is dotted, the simulated response is a continuous line. The numbers in each plot
correspond to the numbers of each pair of zeros on the z domain plot, Fig. 3.32.
Columns show the zeros with radii, 1, 0.8 and 0.5 from top to bottom. Rows show
the zeros as they move around the unit circle, 0 (and 180), 45, 90 and 135 degrees
from left to right. Other characterisation results are shown in Table 3.10.
The FGMOS transistor can be seen either as a MOS transistor with an analog adder
circuit connected at its gate, or as an MOS transistor with an electrically tunable
threshold voltage, or even as a complex nonlinear multiplier divider block. Besides,
=0 = 45 = 90 = 135
nA
100 1 100 2 100 3 100 4
50 50 50 50
r =1
0 0 0 0
0 10 0 10 0 10 0 10
nA
100 5 100 6 100 7 100 8
50 50 50 50
r = 0.8
0 0 0 0
0 10 0 10 0 10 0 10
nA
100 100 100 100
9 10 11 12
50 50 50 50
r = 0.5
0 0 0 0
0 10 0 10 0 10 0 10
kHz kHz kHz kHz
Figure 3.34 AC analysis results for the filter in Fig. 3.31
Table 3.10 Summary of performance

for the filter in Fig. 3.31
Modulation index 0.6

Input amplitude 30 nA
THD Under 0.5%
Supply voltage 1V
Power consumption 1 μW
Clock frequency 33.3 kHz
the degrees of freedom the designer now has to play with increases from 3 to N +2 for
a single device, where N is the number of inputs in the FGMOS. This opens up a new
range of possibilities for the designer in terms of design tradeoffs. Although this is
very promising, designing becomes even more complex. This chapter has illustrated
with simple examples how to obtain the benefits of the FGMOS. In addition, an
overview has been given on how to foresee when the device might or might not be
useful in different contexts, all of them having in common the low power constraint.
Notation
α Gain coefficient in a SI memory cell (Fig. 3.27)

βi Mi β parameter
βFG FGMOS equivalent β parameter (eq. (3.2))
βn β parameter of n-channel MOS device
βp β parameter of p-channel MOS device
Uinverter Variation of Uinverter caused by threshold voltage variations
(eq. (3.26))
VDD Difference between the minimum voltage supply needed by two
current mirror cells, a conventional one and the FGMOS based
cell (Fig. 3.3, eq. (3.14))
+ − + −
Voff , Voff Variation of the voltages at nodes Voff , Voff during the
comparison phase with respect to the values they have at the end
of compensation phase in the FGMOS comparator (Fig. 3.13)
VT VTn2 − VTn1 difference between the threshold voltages of M1 and
M2 in the FGMOS comparator (Fig. 3.12, eq. (3.32))
VTn Threshold voltage variation between n-channel devices in
Fig. 3.10 (eq. (3.26))
|VTp | Threshold voltage variation between p-channel devices in
Fig. 3.10 (eq. (3.26))
ϕ Angles of complex conjugate roots of T(z)
Ao Total DC gain of the two-stage MOS opamp in Section 3.3.2
AoFG Total DC gain of the two-stage FGMOS opamp in Section 3.3.2
ai Coefficients of the filter in eq. (3.67)
C1 Input capacitance connected to VI in the FGMOS current mirror
cell (Fig. 3.3). Also input capacitances in several of the FGMOS
transistors in the opamps in Fig. 3.8 and Fig. 3.9 (their values are
in Table 3.3 and Table 3.6).
(C1 /CT )i Mi input weight (eq. (3.22), Fig. 3.9)
C2 Input capacitance used for biasing purposes in several FGMOS
devices in the opamps in Fig. 3.8 and Fig. 3.9 (their values are in
Table 3.3 and Table 3.6).
Cc Value of the capacitance connected to Vc in Fig. 3.10. Also, value
of the capacitance connected to the offset compensation signal in
the comparator in Fig. 3.12 and of the biasing capacitance
connected to Vc = VDD in the N-bit D/A converters in Fig. 3.16
and Fig. 3.24.
Cclk Value of the capacitance connected to the clock signal in the
FGMOS comparator in Fig. 3.12
Ccomp Compensating capacitance for the opamps in Section 3.3.2
Cf Small extra input capacitance added to compensate the offset due
to CGD and feedthrough in the FGMOS comparator in Fig. 3.13
(eq. (3.45))
CGB Gate-to-bulk capacitance
CGBp Gate-to-bulk capacitance of p-channel FGMOS device in

Fig. 3.10. (Eq. (3.25). Used when it is necessary to
differentiate between n-channel and p-channel devices.)
CGD Gate-to-drain parasitic capacitance
CGDn Gate-to-drain capacitance of n-channel FGMOS device in
CGDp Gate-to-drain capacitance of p-channel FGMOS device in
(CGD /CT )i Ratio between gate-to-drain capacitance and the total
capacitance for transistor Mi (eq. (3.23))
CGSp Gate-to-source capacitance of p-channel FGMOS device in
Ci for i = [1, N ] FGMOS input capacitances
Cin Input capacitance connected to the effective input of the
FGMOS comparator in Fig. 3.12 and also in the inverter in
Fig. 3.10 (used when there is only one effective input) (In the
case of the N-bit D/A converter in Fig. 3.16 it is also the
minimum value binary weighted input capacitance)
Cr Input capacitance connected to Vr = ground in the N -bit D/A
converter in Fig. 3.16
CT Total capacitance seen by the FG (the same value will be
assumed for all the devices unless the opposite is said)
CVDD Input capacitance connected to VDD
GBW Gain-bandwidth product of a two-stage MOS opamp
GBWFG Gain-bandwidth product of a two-stage FGMOS opamp
gdsFi Equivalent output conductance for the FGMOS transistor Mi
gdsi Transistor Mi output conductance
gmi Transistor Mi gate transconductance
Iα Current controlling the gain α in switched transconductance
cell (Fig. 3.27)
Ibias Current flowing through PMOS in the comparator transistors
M3 and M4 in the absence of mismatch between them when
Vbias is applied (Fig. 3.13)
− +
Ibias , Ibias Different bias currents flowing through the negative and
positive comparator (Fig. 3.13) branches, respectively, due to
the mismatch between the threshold voltages in top PMOS
transistors during the compensation phase
Iin− Current in the input branch of the current mirror in Fig. 3.3
Iinmax Maximum value of input current Iin in the current mirror in
Fig. 3.3 (eqs. (3.11)–(3.14))
Ij for j = [1, N ] Current flowing through j-th MOS transistor in Fig. 3.2
(eq. (3.7))
IMi for i = [1, 4] Current flowing through Mi in the N-bit D/A converters in
Fig. 3.16 and Fig. 3.24
Iout Output current
Io See Fig. 3.2
iin Input current in the SI cells in Fig. 3.26 and Fig. 3.28
iin− Single input current for the differential filter built with PSIFG
cells (Fig. 3.31)
iin+ Single input current for the differential filter built with PSIFG
cells (Fig. 3.31)
iini for i = [1, 2] Input currents in the differential PSIFG cell in Fig. 3.29
iout− Single output current for the differential filter built with PSIFG
cells (Fig. 3.31)
iout+ Single output current for the differential filter built with PSIFG
cells (Fig. 3.31)
iouti for i = [1, 2] Output currents in the differential PSIFG cell in Fig. 3.29
J Bias current source in Fig. 3.26
k C2 /(C2 + C1 ) (eqs. (3.64)–(3.66))
Li Effective length of transistor Mi
phi1 Clock signals for S1 and S2 in FGMOS comparator in Fig. 3.13
phi2 Clock signals for S3 and S4 in FGMOS comparator in Fig. 3.13
phi2_b Clock signals for S5 and S6 in FGMOS comparator in Fig. 3.13
QFG Charge trapped at the FG
Rc Compensating resistance in the two-stage opamps
r Radius of complex conjugate roots of T(z) (zeros of the
system) (eqs. (3.69) and (3.70), Table 3.9)
T (z) Transfer function for the second-order filter in eq. (3.68)
Uinverter Switching threshold of CMOS inverter (eq. (3.24))
V1 M1 drain-to-source voltage in Fig. 3.3 (eqs. (3.9)–(3.11)). Also
one of the voltages used to tune the PSIFG cell (Fig. 3.28)
V2 One of the voltages used to tune the PSIFG cell (Fig. 3.28)
Vbias Bias voltage of PMOS transistors M3 and M4 in FGMOS
comparator in Fig. 3.13
Vbias1 Bias voltage for M1 in the current mirror with FGMOS in
Fig. 3.3(b) (assumed to be equal to VDD )
Vbias2 Bias voltage for M3 in the current mirror with FGMOS in
Fig. 3.3(b) (assumed to be equal to VDD )
Vc Input voltage of the two-input FGMOS inverter in Fig. 3.10 –
used for changing the switching threshold (eq. (3.25))
Vclk Comparator clock voltage
Vcm Common mode voltage applied to Vin+ during the offset
compensation phase (Fig. 3.13)
VD1 Drain voltage of transistor M1 in N -bit D/A converter in

Fig. 3.16
VD3 Drain-to-source voltage of transistor M3 in Fig. 3.3(b)
VDD1 Voltage at the drain (gate) of diode connected transistor M2
in the cascode current mirror in Fig. 3.3 (eq. (3.10))
min
VDD Minimum voltage supply needed by the current mirror in
Fig. 3.3 to drive a maximum input current Iinmax
(eqs. (3.11)–(3.14))
VFG Voltage at the FG (Fig. 3.1)
Vi for i = [1, N ] FGMOS effective inputs
Vin FGMOS effective input voltage (when there is only one)
(effective)
Vind Effective differential input signal when there is an offset
caused by mismatch between the input transistors
(Fig. 3.13, eq. (3.39))
Vin(threshold) Voltage at the effective input of FGMOS inverter for which
the switching of the inverter happens
Vin± Effective inputs of the FGMOS comparator in Fig. 3.12
VLoad Voltage across the biasing/input part of the cascode mirror
circuit in Fig. 3.3
Voff Input of the FGMOS comparator in Fig. 3.12 used for offset
+ −
compensation. It is Voff = Voff when there is no mismatch
+ − + −
Voff , Voff Voltage generated at Vout and Vout respectively (by the
negative feedback) in the FGMOS comparator in Fig. 3.13
during the offset compensation phase (eqs. (3.34) and
(3.35))
voff (CGD/feedthrough ) Equivalent input referred offset due to the effect of the
parasitic capacitance CGD and feedthrough (eq. (3.45))
voff (Ibias ) Equivalent offset at the input of the comparator in Fig. 3.13
due to the difference in bias currents in the positive and
negative branches of the comparator (eq. (3.46))
voff VT Offset at the input caused by threshold voltage variations in
the FGMOS comparator (Fig. 3.12, eq. (3.32))
±
Vout Single outputs of the FGMOS comparator in Fig. 3.13
−
Vout(comparison) , Voltage at the ‘−’ or ‘+’ output node of FGMOS
+ comparator in Fig. 3.13 during the comparison phase
Vout(comparison)
−
Vout(compensation) , Voltage at the ‘−’ or ‘+’ output node of FGMOS
+ comparator in Fig. 3.13 during the compensation phase
Vout(compensation)
Vout1 Output of the FGMOS inverter in Fig. 3.10
Vref Voltage at the drain of M1 in the D/A converter in Fig. 3.24
fixed by the opamp (eq. (3.62))
VT (CT /Ci )VTFG . FGMOS effective threshold voltage
parameter (eq. (3.3))
N
VTFG VT − j=1 (Cj /CT )VjS − QFG /CT . Different value for
j=1
different designs within the chapter
VTni for i = [1, 2] Mi threshold voltage when there is mismatch in the
FGMOS comparator (Fig. 3.12)
VTn n-Channel MOS transistors threshold voltage (the same for
all the devices unless the opposite is said)
|VTp | p-Channel MOS transistors threshold voltage (the same for
all the devices unless the opposite is said)
Wi Effective width of transistor Mi
wj Cj /CT . Input weight corresponding to the j-th FGMOS
input
wpi for i = [1, 2] Angular frequency of the i-th pole in the MOS/FGMOS two
stage opamp (eq. (3.21))
x(n) See eq. (3.67) (FIR filter)
y(n) See eq. (3.67) (FIR filter)
Chapter 4
Low power analog continuous-time filtering
based on the FGMOS in the strong inversion
ohmic region
4.1 Introduction
The following four chapters describe how to exploit the mathematical capabilities
of FGMOS devices in more complex circuits taking as an example the design
of continuous-time analog filters. Filters are inevitable building blocks in high-
performance low power (LP) and low voltage (LV) electronics [143,144]. A filter
circuit is a two-port unit and its function is to process the magnitude and/or phase of
an input signal in a desired way. There are several filter realisations and their choice
depends on the needed signal processing as well as the required circuit performance
[145]. Once the mathematical function describing a filter transfer characteristic is
derived, its final realisation will be determined by technological limitations.
The first decision to be made when opting for a filter topology is whether it
is going to be a discrete or a continuous-time system [146]. In many cases, the
filter input signal comes from the real world and therefore is continuous in time. At
this point, the designer has to evaluate whether it is beneficial for the whole system
performance to implement it all in discrete time or not. The best performing technique
for designing discrete time filters is the one known as switched capacitor. Switched
capacitor circuits simulate filters transfer functions only in terms of capacitances
ratios and so component tolerancing is not a major problem. However, switching
speed limitations make this a fairly low frequency technique.
Yet, the main focus of this book is on continuous-time circuits12 . There are three
main approaches for the design of CMOS continuous-time active filters: MOSFET-C
filters, RC active filters and OTA-based filters [147].
12 Some discrete time realisations can be found in Chapter 3.

MOSFET-C filters are based on the use of an equivalent CMOS tunable resistor in
an RC structure [148,149]. Their main drawback is that they require high-performance
OPAMPs which limits their use to low-frequency applications. Besides, if the voltage
supply is scaled down it becomes very difficult to keep an acceptable performance
since the OPAMP functionality seriously degrades.
The difference between RC filters and MOSFET-C filters is that the former use
passive instead of MOSFET-based resistors. This adds some extra problems as they
require a technological process with linear, stable resistors which also sometimes
need to be large and well matched. Furthermore, it is more difficult to tune them.
The basic elements of OTA-C filters13 are Operational Transconductance Ampli-
fiers (OTA) and capacitors (C). The major advantage of OTA-based continuous-time
filters versus MOSFET-C realisations is their wider frequency response [150–159].
In MOSFET-C blocks the frequency range is limited to one-thirtieth of the OPAMP
gain-bandwidth frequency (GBW), whereas in OTA-based integrators which are
implemented in open loop configuration the only limitation to the frequency response
is the OTA gain-bandwidth product [160]. The main drawback of OTA realisations is
the smaller dynamic range which is limited by the distortion introduced by nonlinear
terms in the mathematical functions that describe the current. Linearised topologies
can be used in order to reduce higher order harmonic components [153,161–178].
Owing to all these reasons, linearised OTA-C topologies seem to appear as the most
suitable choice when the predominant design constraints are the low voltage and low
power [143].
The performance of low voltage continuous-time filters is directly related to
the quality of the integrator building block measured in terms of power consump-
tion, linearity, noise, phase response and tunability as main parameters [160].
There are different approaches in terms of the integration method, which can be
either the traditional internally linear-externally linear (ILEL) or externally linear-
internally nonlinear (ELIN) [21,22,179–188]. The latter has appeared as a very
popular option for the low-voltage high-frequency operation in bipolar or BiCMOS
technologies [189–193] while MOS realisations are more suitable for audio frequen-
cies [26,194,195]. An important subset of ELIN circuits are logarithmic filters.
Their operation is based on the exponential law that models the behaviour of
devices such as bipolar transistors or MOS working in the weak inversion region
[169,170,179,182,189–205].
The following four chapters describe a different approach in designing continuous-
time analogue filters, which combines some of the previously reported techniques
with the use of FGMOS devices within the building blocks. This design approach
has proven to be very useful when re-establishing the technological limits in order
to reduce the power [94,101,174,177,178,206–210]. The aim of these chapters is to
teach the reader how to take advantage of the FGMOS device when the main design
constraints are the power supply voltage and power consumption and to analyse the
negative implications of using this device.
13 Also called G -C filters.

m
Filtering with FGMOS in the strong inversion ohmic region 89
The first chapter of this series of chapters focuses on the properties of an FGMOS
transistor operating in the strong inversion ohmic region. An LP/LV multiple-input
linear transconductor is designed and subsequently used for the implementation of
two circuits: a fully differential Gm -C integrator and a filter.
The FGMOS transistor is also used, although not biased in the ohmic region, in
other sections of the filter circuit with different aims. Examples are the use of FGMOS
devices (instead of normal MOS) as: a) cascode devices with reduced effective thresh-
old voltages; or b) in a common mode feedback circuit to perform averaging at a very
low power cost.
The side effects of using the FGMOS in each individual case are also thor-
oughly analysed. Studies of Power Supply Rejection Ratio (PSRR), Common Mode
Rejection Ratio (CMRR) and Total Harmonic Distortion (THD) are provided.
Experimental results supporting the conclusions are shown at the end of the
chapter.
4.2 Transconductor basic blocks
This section introduces basic building blocks that are later used in the implementation
of a Gm -C integrator and a filter. Two circuit topologies are presented, namely N1 and
N2. N1 represents a multiple input transconductor which uses an FGMOS transistor
biased in the strong inversion ohmic region while N2 is a single stage FGMOS-based
OPAMP required to linearise the operation of N1. The operation of both blocks, N1
and N2, is explained below.
The current law for an n-channel FGMOS transistor biased in the strong inversion
ohmic region, as given by eq. (2.6), can be rewritten in a simpler form if the drain to
source voltage is assumed to be very small, that is
VDS (VGS − VTn ) (4.1)
In this situation, the quadratic term can be neglected from the equation, resulting in
" N #
Ci
ID ≈ βn · Vi − VTn VD (4.2)
CT
i=1
where the n subscript in the threshold voltage as well as in β parameter is used to

refer to the n-channel nature of the transistor. It has also been assumed that the source
and bulk are connected to ground, there is no charge accumulation and the parasitic
capacitances are much smaller than the input ones. Equation (4.2) shows that if the
drain voltage, VD , is kept constant, the current ID will depend linearly on the input
voltages with slopes βn (Ci /CT ).
An FGMOS-based implementation of eq. (4.2) is shown in Fig. 4.114 . M1 is
the FGMOS transistor biased in the strong inversion ohmic region. Transistor M2
14 This basic cell was also the basis of a D/A converter as explained in Chapter 3, Section 3.3.5.
VDD
M2 Vout(N2)
Vinr
Vb N2
V1
M1
VN
VDD
N1
Figure 4.1 N1 – Basic circuit implementing eq. (4.2) (N1)
VDD
IB2
Vinr Mp Mp Vb
VR Vout(N2)
Mn Mn
VDD
Figure 4.2 N2 – Tuning amplifier
performs a current to voltage conversion and will be used for the mirroring purposes.
N2 is a high gain differential input/single output amplifier that sets the drain voltage
of M1 to a constant reference: VD = Vb . An FGMOS-based realisation of N2 is
shown in Fig. 4.2. It consists of a pMOS differential pair with n-channel FGMOS
transistors acting as a load. The input Vinr is compared with the reference voltage
Vb and if their values are not equal the difference is amplified at the output Vout(N2) ,
which is connected to the gate of the pMOS transistor M2 (in N1 block). This would
unbalance the currents flowing through M1 and M2 and would have to be corrected
by the output resistance which changes the Vinr value in the opposite direction. The
feedback forces the operating point of Vinr to be equal or almost equal to the voltage
Vb . From now on this Vb voltage will be referred to as the tuning voltage.
Voutcm
1:k k:1
VDD VDD
VDD VDD VDD VDD
Ma3 Ma3r
Vs3p Vs3n
M2 M3 M3r M2r
Mc3 Mc3r
Vb N2 gnd C gnd N2 Vb
Voutp Voutn
V1p Iout V1n
M1 Mc1 Mc1r M1r
VNp M4 VDD VDD M4r VNn
VDD Vs1p Vs1n VDD
Ma1 Ma1r
N3
N1 N4
k:1
1:k
Figure 4.3 Fully differential multiple-input integrator15
4.3 The Gm -C multiple-input linear integrator. Large signal

description
Figure 4.3 shows the schematic of a fully balanced linear transconductor built by
replicating and combining the blocks described in the previous section.
The schematic shows two identical N1 blocks on the right and left side of
the figure. The positive inputs are connected to the left side transconductor block
(Vjp , j = [1, N ]), whereas the negative ones are applied to the transconductor
on the right side (Vjn , j = [1, N ]). The respective output in each N1 cell is
connected in a mirror configuration to a pMOS device (M3/M3r). The latter gen-
erates a copy of the current flowing through M1/M1r and sources that current to
a diode connected nMOS transistor (M4/M4r) which, acting as the input device
of a mirror structure, produces a negative/positive copy of it. A positive and a
negative scaled (1 : k) replica of the negative and positive currents16 , respectively,
are subtracted at the output nodes (Voutp /Voutn ) generating the output current Iout ,
described by eq. (4.3). This equation shows how the constant term including the
threshold voltage in eq. (4.2) disappears thanks to the differential nature of the
15 The subscript r is used to distinguish between different devices with the same dimensions. Hence,
for example, transistor M1 and transistor M1r will be equally sized. Also, from now on the parameters
will have an added subscript in agreement with the transistor name.
16 The positive current is the current generated by the p inputs while the negative current is the current
generated by the n inputs.
structure:

N
Cj
Iout = kβ1 Vb (Vjp − Vjn ) (4.3)
CT
j=1
Equation (4.3) represents the function of a linear multiple input fully differential
transconductor whose transconductance can be adjusted by changing the value of the
reference voltage Vb in the amplifier N2. The large signal transconductance referred
to one of its differential inputs (i) is given by eq. (4.4). Again, this value does not
depend on common mode signals. Hence, the only restriction for the common mode
input refers to the operation of the transistors in the strong inversion region:
Ci
Gmi = kβ1 Vb (4.4)
CT
The negative replica of the output stage in the schematic provides a negative copy
of the output current flowing from node Voutn to Voutp . The capacitor added between
these two terminals integrates this output current. Cascode transistors are added at
the output stage in order to improve the value of the output resistance.
4.3.1 Operating point – design considerations

The biasing of certain devices in Fig. 4.1 and Fig. 4.2 can be of crucial importance
for the proper operation of the integrator in Fig. 4.3, especially when the maximum
voltage supply is pushed down. The design equation derived in this section can be
used to predict the deviation of the real tuning voltage Vinr , from the chosen value Vb
as a function of the supply voltage as well as the common mode and the gain of the
feedback amplifier.
The current flowing through transistors M1 and M2 is the same. Transistor M2
has to be biased in the strong inversion saturation region for the proper operation.
Under these conditions, using equations (2.7) and (2.8)
β2
β1 (VFG − VTn )Vb ≈ (VDD − Vout(N2) − |VTp |)2 (4.5)
2
where Vout(N2) is the output of the tuning amplifier and as such it is a nonlinear
function of Vb − Vinr 17 . However, as a first approximation the first two terms of a
Taylor series expansion are good enough to represent this function since Vinr ∼= Vb :
Vout(N2) = f (Vb − Vinr ) ≈ Vout(N2)Q − gmf Routf (Vb − Vinr ) (4.6)
The subscript f refers to the feedback block N2. gmf Routf represents its input/output
gain18 . Vout(N2)Q is the operating point at the output when its differential input is zero,
17 Neglecting common mode effects.

18 The tuning amplifier N2 is just a differential pair. Its gain is the product of its transconductance and
the output resistance.
i.e. Vb = Vinr . Combining eq. (4.5) and eq. (4.6), the expression for Vinr is
1
Vinr = Vb + (D − BVb ) (4.7)
gmf Routf
where
D = VDD − Vout(N2)Q − |VTp | (4.8)
2.β1
B= (VFG − VTn ) (4.9)
β2
Equation (4.7) shows how the function which relates the tuning voltage, Vb , with the
voltage, Vinr , fixed by the N2 block is not an exact equality, especially when Vb
decreases, which might increase the complexity of the tuning mechanism for low
transconductance values. This effect can be made less significant by increasing the
gain of N2; however, doing this could cause instability as will be shown in the next
chapter.
4.3.1.1 The threshold voltage shift

All the discussion and equations derived in previous sections are based on the assump-
tion that all transistors are operating in the strong inversion region, which means that
the condition |VGS | > |VT | is satisfied (see eqs. (2.6–2.9)) for all the devices. If
applied to transistor M2 in each N1 block, this means:
Vout(N2) < VDD − |VTp | (4.10)
As an example, in a technology with |VTp | ≈ 0.8 V, and for a supply voltage of 1.5 V,
eq. (4.10) would restrict the voltage at the output of the tuning amplifier to a maximum
value of 0.7 V. But, besides, nodes VR and Vout(N2) in Fig. 4.2 should have close oper-
ating points for the sake of accuracy. Also, the n-channel MOS transistors Mn should
work in the strong inversion region for matching and speed reasons, which means that
their gate voltage has to be above their threshold voltage (VG > VTn ≈ 0.8 V). Both
conditions cannot be met simultaneously if the mirror is implemented with normal
n-channel MOS transistors. However, if FGMOS devices are used instead, as shown
in Fig. 4.2, the design constraint for VR is
CT CVDD
VR > VTn − VDD (4.11)
CR CR
where CR is the value of the capacitance connected to the node VR and CVDD is
the capacitance of the other FGMOS input. This input is connected to the maximum
voltage supply, VDD , in order to shift the threshold voltage and thus relax the voltage
requirements. Equation (4.11) neglects the effect of the parasitic capacitances on the
FG19 voltage. This approach will be followed from now on, unless the opposite is
asserted. The reader can find out more about the design of the current mirror with
FGMOS in Chapter 2.
19 This is valid whenever they are much smaller than the input capacitances.
The threshold voltage has also been shifted in the input transistor of the N1 block
by connecting one input capacitance in M1 to VDD . This is done with the aim to
improve the input range of the circuit for the chosen common mode at the input so
that the condition of having the transistors working in the strong inversion region is
met for a wider range of input voltages. A more exhaustive study on how to choose
this capacitance can be found in Chapter 3, Section 3.3.5.
4.3.1.2 The cascode output stage. Output swing

The low frequency performance of the integrator in Fig. 4.3 can be improved by
adding two-input FGMOS cascode transistors at the output stage (Mc1 , Mc1r , Mc3 ,
Mc3r ). The reason for choosing FGMOS instead of normal MOS devices is twofold:
on one hand, one input of each transistor can be used as part of the common mode
feedback mechanism, needed because of the fully differential nature of the topology;
on the other hand, the second input can be used to shift the effective threshold voltage
hence making the low voltage constraint less restrictive. The common mode feedback
strategy will be explained in the next section. Let us now explain the benefits of using
the FGMOS instead of MOS devices in this particular topology.
The cascode transistors have to operate in the saturation region, preferably in
strong inversion, in order to maximise the speed. The conditions for this are (according
to eqs. (2.6–2.9))20
VFGcn − Vs1p > VTn (4.12)
Vs3p − VFGcp > |VTp | (4.13)
where VFGcn and VFGcp are the voltages at the FGs of the n-channel and p-channel
cascode transistors, respectively. The previous equations can be rewritten as

Voutcm > VTn + Vs1p (4.14)

Voutcm < Vs3p − VTp (4.15)
where Voutcm is the voltage at the output of the common mode feedback circuit,
Vs1p and Vs3p are the voltages at the sources of the n-channel and p-channel cascode
, V are their effective threshold voltages which in
transistors, respectively, and VTn Tp
this case are given by

CT CVDDf
VTn = VTn − (VDD − Vs1p ) (4.16)
CCMn CT
CT Cgndf
VTp = |VTp | − Vs3p (4.17)
CCMp CT
20 A similar expression can be obtained for the negative branch.

CVDDf , Cgndf are the capacitances connected to VDD and ground in the n-channel and
p-channel cascode transistors, respectively, and CCMn , CCMp are the capacitances
connected to Voutcm .
According to equations (4.14) and (4.15) Voutcm has to be high enough in order
to keep the n-channel transistor in the strong inversion region, but at the same time
low enough, so the p-channel device remains in that region as well. However, as the
voltage supply decreases the limits for the bias voltage move closer to the supply
voltages. Because of this the generation of a suitable bias signal becomes a more
complex task. The FGMOS can simplify it by shifting the effective threshold voltages
with adequate values of CVDDf and Cgndf according to eqs. (4.16) and (4.17) These
expressions show how the reduction of the effective threshold voltages is directly
proportional to the value of the weights CVDDf /CT and Cgndf /CT for the n-channel
and p-channel devices, respectively. Nevertheless, the designer should bear in mind
that there is a price to pay for this, since the input transconductances of the FGMOS
transistors also decrease proportionally, and hence the output resistance. Still, as long
as the effective input transconductance is kept larger than the output conductance for
both FGMOS devices, it will be beneficial to use them at the output stage.
4.4 The Common Mode Feedback Block (CMFB)
The characteristic of fully differential structures is that signals are differential instead
of being referred to ground. This property, however, is also the origin of one of their
more important drawbacks: the common mode voltage of fully differential structures
cannot be stabilised by the differential feedback. Thus, in order to stabilise the operat-
ing point an additional Common-Mode Feedback circuit (CMFB) has to be included
in the topology [160].
The design of a CMFB can be difficult and it increases the complexity of the
transconductor as the final topology has two signal paths: one defining the differential
transfer function of the system and another one for the common mode signals. The
circuit must detect the common signals as fast as possible in order to prevent them
from affecting the differential ones. This should be achieved with a simple block,
that ideally would not affect the overall integrator performance. Besides if the total
power is a limiting factor, the circuit should consume as little power as possible.
Several CMFB implementations have been reported. Some of them use a
capacitor–resistor network to detect the common mode level [148]. Although this
technique is very efficient because of its zero power consumption and the high linear-
ity, it requires a large area and it might also reduce drastically the output resistance
of the transconductor. An alternative to this is to use transistors biased in the ohmic
region [211]. The problem of this technique is that it reduces the transconductance
and the bandwidth of the loop. The most common technique is the use of a differen-
tial pair in which the transistors at the output are saturated, and for low differential
voltages, the AC-signal at the sources is equal to the common mode input signal. The
problem with this structure is that it can be strongly nonlinear when the differential
signal exceeds a certain value. It also degrades the differential impedance.
VDD
Mf3 Mf4
Voutcm
Voutp VDD
Voutn Mf1 Mf 2
VDD gnd
Vbias3
Figure 4.4 Common Mode Feedback Block (CMFB)
The circuit in Fig. 4.3 is a fully differential structure and as such it requires
a CMFB in order to correct variations in the common mode at the output of the
transconductor. An example of a block that could be used as CMFB in this par-
ticular design is shown in Fig 4.4. It consists of a differential pair with FGMOS
type transistors at the input and a MOS current mirror as a load. The mean value
of the single positive and negative outputs is sensed at the gate of one of the input
FGMOS transistors and compared to a reference voltage generated at the gate of
the other FGMOS device. If they are not equal the difference is amplified at Voutcm
and fed back to both cascode transistors. This unbalances their currents in opposite
directions, which forces the common mode at the output to move towards the ref-
erence value in order to equalise the sinked and sourced currents within the same
branch.
The reference voltage in Mf2 (see Fig. 4.4) can be generated either by directly
connecting the voltage supplies to two properly weighted inputs or with a biasing
circuit. The first option would affect the PSRR as it will be shown later on.
Also, an extra input can be added to each of the input FGMOS devices with
the purpose of reducing the effective threshold voltages in order to keep the input
transistors in the strong inversion region (for matching and speed reasons) even
in cases when the common mode at the output is below the nominal threshold
voltage.
4.5 Small signal considerations
Small signal analysis of the cells forming the transconductor is presented in this
section and several issues, important for the design of this circuit, are brought to the
reader’s attention.
C⬘1 vinr C⬘3
vFG AV vinr
gm1vFG gds1 gm2 AV vinr gds2 C⬘2
Figure 4.5 Small signal equivalent circuit for N1
Hence, for example, the effect of the gate to drain parasitic capacitance will be
described and quantified. It will be shown how the latter introduces an error in the
transconductance, causes a zero in the transfer function and affects the circuit stability.
Also, small signal analysis will prove how the zero of the N2 block is transmitted
to the output current. The frequency at which this zero happens depends on the values
of the input capacitances for the FGMOS transistors in N2.
Furthermore, the ratio between the gate to drain parasitic capacitances and the
total capacitance in N2 FGMOS transistors influences N2 dominant pole and this,
together with the value of the capacitance used to shift the effective threshold voltage
(connected to VDD ), can again affect the transconductor stability.
Finally, the CMRR is going to suffer from the mismatch between the FGMOS
transistors input capacitances. This effect will also be analysed for low frequencies.
4.5.1 The transconductance cell (N1)

A small signal equivalent circuit for Fig. 4.1 is shown in Fig. 4.5. AV is the small-signal
gain of the feedback block N2. The parasitic capacitances are given by
C1 = CGD1 (4.18)
C2 = CDB1 + CDB2 + CLN2 (4.19)
C3 = CGD2 (4.20)
CGDi , CDBi are the gate to drain and drain to bulk parasitic capacitances for tran-
sistor Mi. CLN2 represents the capacitive load due to the connection with the tuning
amplifier N2. The small signal voltage at the FG is
Cin C
VFG = vin + 1 vinr (4.21)
CT CT
where vin is the effective input21 and Cin is the value of its corresponding input
capacitance. Hence, the transfer function is given by
−(Cin /CT )(gm1 − sC1 )vin
vinr = & '
gm2 AV +gds2 +gds1 +gm1 (C1 /CT )+s C1 +C2 +(1 − AV )C3 −C 21 /CT
(4.22)
And the current flowing through a transistor M3 connected, as a mirror, to M2:
(Cin /CT )gm3 AV (gm1 −sC1 )vin
iout = & '
gm2 AV +gds2 +gds1 +gm1 (C1 /CT )+s C1 +C2 +(1 − AV )C3 −C 21 /CT
(4.23)
Modelling the frequency response of the feedback circuit (AV ) as a single pole –
single zero function, eq. (4.23) can be rewritten as
(Cin /CT )gm3 A0 (1−s/z)(gm1 −sC1 )vin
iout = ( ( ) )
(gm2 −sC3 )A0 (1+s/z)+ gds2 +gds1 +gm1 (C1 /CT )+sCeq (1+s/p)
(4.24)
where A0 , p and z are the DC gain, pole and zero of the tuning block (N2), respectively.
Also:
C 21
Ceq = C1 + C2 + C3 − (4.25)
CT
The feedback circuit has a couple of important effects on the frequency response of
the transconductor core block. On one hand, it affects the value of the DC gain which
now is given by
iouti
ioutr = (4.26)
1+ε
where the subscripts r and i refer to the real and ideal DC responses and ε quantifies
the DC error:
gm1 C1 (gds2 + gds1 )
ε= + (4.27)
gm2 A0 CT A0 gm2
Equations (4.26) and (4.27) reveal how the larger the DC gain of the feedback circuit
is the smaller is the error. Also, the first term in eq. (4.27) is proportional to the ratio
between the gate to drain parasitic (C1 ) and the total capacitance, which means that
if this ratio is smaller the error will be smaller too.
On the other hand, the numerator of eq. (4.24) shows how the zero of the feedback
amplifier is transmitted to the output current, so this amplifier should be designed in
such a way that its zero remains far away from the operating frequency band of the
circuit. Apart from this zero, there is another one caused by the coupling between M1
21 The input for which the transfer function is calculated.

drain and the FG. This means that in order to achieve the desired operating frequency
range a trade-off between the maximum transconductance, the power and the area
has to be met. On one hand, increasing the transconductance will increase the power
as well as the error in eq. (4.27), but it will also improve the speed. On the other
hand, the error can be reduced by increasing the total capacitance seen by the input
FG (without changing the Cin /CT ratio), but this would also increase the circuit area.
Another conclusion can be extracted from the previous results analysing the sec-
ond order polynomial in the denominator of eq. (4.24): the circuit is only conditionally
stable, this is whenever any of the conditions shown in eq. (4.28) or eq. (4.29) happen,
the circuit will oscillate. Hence, the tuning amplifier must be designed in such a way
that none of them ever occurs:
" #
gm2 A0 C 21 C1 (gds1 + gds2 )
+ C1 + C2 + (1 − A0 )C3 − + gm1 + <0
z CT CTp p
(4.28)
⎛ ⎞
C1 + C2 + C3 − C 21 /CT C3 A0
⎝ − ⎠<0 (4.29)
p z
The analysis of the tuning amplifier frequency response is carried out in the
following section.
4.5.2 The tuning amplifier (N2)

The small signal equivalent circuit for the tuning amplifier (N2) is shown in Fig. 4.622 ,
where
C02 = CVDD + 2CGSn + 2CGBn (4.30)
C22 = CDBn + CDBp + CGDp + CL (4.31)
C42 = CDBn + CDBp (4.32)
Subscripts n and p refer to the name of the device. gdsbias accounts for the real
conductance of the current source that generates IB2 . CL refers to a capacitive load at
the output:
vout(N2) (gmp /2)(s/p1 + 1)
= (4.33)
vinr A1 s 2 + B 1 s + D 1
22 The pMOS transistors are assumed to be in separate wells. Also, for the sake of simplicity, only the
most important capacitances have been considered. Therefore, the following analysis will only be valid
for mid-frequencies, but this is good enough to obtain the most important conclusions.
gdsbias
vs
gmpvS gdsp gdsp gmp(vS-vinr)
CGDn (CR+CGDn)
vout(N2) vR
gmnvFGn gdsn C22 C02 C42 gdsn gmn vFGn
Figure 4.6 Small signal equivalent circuit for N2
1 C2 (CR + CGDn ) a0
A1 ≈ (C22 + CGDn − GDn ) − CGDn (4.34)
p1 CT CT z1
(gmn (CGDn /CT ) + gdsn + gdsp /2) gmn (CR + CGDn ) a0

B1 ≈ +
p1 CT z1

C2 CGDn gdsp a0
+ C22 +CGDn − GDn −a0 (CR +CGDn )− (4.35)
CT CT 2 z1
(CR + CGDn ) CGDn gdsp gdsp

D1 ≈ gmn a0 + gmn + gdsn + − a0 (4.36)
CT CT 2 2
where in this case, the total capacitance seen by the FG is
CT ≈ 2CGDn + C02 + CR (4.37)
a0 , p1 and z1 , are the gain, transmission pole and zero from vout(N2) to vR , respectively.
The values of these parameters are collected in equations (4.38), (4.39) and (4.40):
gdsn + 2gmn (CGDn /CT )

a0 = (4.38)
gdsn + 2gmn ((CGDn + CR )/CT )
gdsn + 2gmn (CGDn /CT )

z1 = (4.39)
(C22 + CGDn ) − CGDn (CR + 2CGDn )/CT
gdsn + 2gmn ((CR + CGDn )/CT )

p1 = (4.40)
(C42 + CR + CGDn ) − (CR + CGDn )(CR + 2CGDn )/CT
Therefore, N2 DC gain will be given by

gmp /2
A0 ≈
gmn a0 ((CR +CGDn )/CT )+gmn (CGDn /CT )+gdsn +gdsp /2−(gdsp /2)a0
(4.41)
As expected the increased output conductance of the load FGMOS devices affects
the DC gain, A0 , which is now reduced if compared to the gain of the same topology
built with normal MOS devices instead.
N2 dominant pole can be calculated using the following assumptions:
(CGDn + CR )
2gmn
gdsp + gdsn (4.42)
CT

CR
C22
1− CGDn (4.43)
CT
CR
CGDn (4.44)
which gives the approximate value
2[gmn a0 ((CR +CGDn )/CT )+gmn (CGDn /CT )+gdsn +gdsp /2−gdsp a0 /2]
p≈
3C22
(4.45)
The value of the zero is given by
2gmn
z≈− (4.46)
CT − CR
Hence, the oscillation conditions in eq. (4.29) will never happen, because of the
negative value of z. Also eq. (4.28) can be rewritten as

C 2
1 C1 (gds1 +gds2 ) gm2
C1 +C2 +C3 − +gm1 + < C3 + (CT −CR ) A0
CT CTp p 2gmn
(4.47)
The right-hand side in eq. (4.47) shows how in order to minimise the risk of oscillation
the term relating the transconductor transconductance, the capacitance connected to
VDD in the current mirrors and N2 transconductance should remain as small as possible
and the relationship between the values of the parasitic and load capacitances should
be such that eq. (4.47) never occurs.
4.5.3 Common mode DC response

The integrator in Fig. 4.3 is a fully differential structure. Therefore, ideally, its
common mode gain should be as small as possible. By replicating the small signal
equivalent circuit shown in Fig. 4.5 for both differential branches, and performing a
small signal analysis, the equation for the gain referred to one of the output branches
(p) is23
νoutp
ν =0
νinCM ind
[(Cin /CT )p (gmf Routf )p gm1 gma3 −(Cin /CT )n (gmf Routf )n gm1r gm3r (gmal /gm4r )]Routp
≈
[gm2 (gmf Routf )p +gds2 +gds1 +gm1 (C1 /CT )p ][1+ACMp ]
(4.48)
Routp is the name used for the output resistance in the output branch (p). Assuming that
the p-channel FGMOS transistors are in separate wells, the value of Routp is given by

−1 gdsFc3 gdsa3
Routp = Goutp ≈
[(CCMp /CT + Cgndf /CT )p gmc3 + gdsFc3 ]

−1
gdsFc1 gdsa1
+
[(CCMn /CT + CVDDf /CT )p gmc1 + gmbFc1 + gdsFc1 ]
= [gdsFc3 · R1p + gdsFc1 · R2p ]−1 (4.49)
ACMp depends on the CMFB. Its value is given by24

(CCMp /CT )p (gmc3 · R1p ) + (CCMn /CT )p (gmc1 R2p ) 2gmf 1 (Cin /CT )
ACMp ≈
Goutp gdsf 4 + gdsFf 2 p

ACM1p 2gmf1 (Cin /CT )

= (4.50)
Goutp gdsf 4 + gdsFf 2 p
The differential gain can be derived by assuming that both branches are identical
−1 25
and Routp = Routn = Rout = Gout :

νout
≈ Gm Rout (4.51)
νind νinCM =0
The Common Mode Rejection Ratio (CMRR) is given by the ratio between the
differential gain in eq. (4.51) and the difference between the common mode gains of
both output branches assuming mismatch (eq. (4.48)):
[νout /νind ] |νinCM =0
CMRR = ( ) (4.52)
((νoutp /νinCM ) |νind =0 − (νoutn /νinCM ) |νind =0 )
23 The subscript p in g
mf Routf is used to make reference to the N2 amplifier corresponding to the p
branch. Also, in general when either the subscript p or n is added to the name of a parameter it will refer
to the p or n branch, respectively.
24 The value of the input capacitance C has been assumed to be the same as in the CMFB.
in
25 For the sake of clarity. As its value is supposed to be very high this will not affect much the final
result.
The main conclusion drawn from this analysis is that the use of FGMOS transistors
can have a detrimental effect on the CMRR. This is because the common mode gain at
the output increases as the gain of the CMFB decreases. Having FGMOS transistors as
cascode devices will reduce ACM(n,p) due to the reduction of the transconductance and
the increase of the total output conductance through gdsFci . However, if the FGMOS
parasitic capacitances in the CMFB are small enough (compared with the total capac-
itance), the gain of the block can still be large enough to meet the specifications and
compensate for the compression of the signals at the FGs. It is also worth pointing
out that the mismatch between the input capacitances in the FGMOS transistors is
going to affect the value of the common mode output gain. As an example, for a
1 per cent mismatch in the input capacitances, ACM(n,p) = 10, k = 1, and values of
conductances much smaller than the values of the transconductances, the magnitude
of the CMRR will be higher than 60 dB.
4.6 Second order effects
Two typical analog filter design specifications are linearity and noise. Linearity limits
the value of the largest useful signal that can be handled by the filter while noise
establishes the limit for the smallest. DR is the ratio between these two magnitudes,
and it is a crucial figure in LV/LP analog design. The maximum input signal that can
be applied to a circuit depends directly on the Total Harmonic Distortion requirement
(THD) for the system [212]. This section presents a study of both the THD and the
noise in the transconductor block together with a discussion of the repercussions of
using the FGMOS transistors in the values of these parameters.
4.6.1 Total harmonic distortion

When a sinusoidal waveform is applied to a linear time-invariant system, the output
will also be a sinusoidal waveform at the same frequency, although, possibly with
different magnitude and phase values. If this input is applied to a nonlinear system the
output will have extra frequency components (harmonics) at frequency values which
are multiples of the input frequency. The THD of a signal is defined as the ratio of
the total power of the second and higher order harmonic components to the power of
the fundamental harmonic:

(HD22 + H2 + H2 + · · · )
D3 D4
THD = × 100 (4.53)
HD1
HD1 , HDi are the amplitudes of the fundamental and ith-harmonic components,
respectively.
A problem related to the use of THD as a filter characterisation parameter is
that often harmonic components fall in the stopband of the filter, and thus the THD
value can appear as falsely improved. In fact, the filter linearity is worse when higher
input signal frequencies are applied due to nonlinear capacitances or nonlinear signal
cancellation. A better way of measuring the linearity near the upper passband edge is
by using the intermodulation test (IM3). This test is actually the only one that works
for bandpass filters [212]. This section analyses the main sources of non-linearity in
the integrator circuit. Also, an approximate equation for both, the THD and IM3 as a
function of different design parameters is derived.
Equation (4.2) was obtained assuming that the drain current of the input FGMOS
transistor in N1 was a perfect linear function of VDS . Also, β1 was supposed to be
a signal-independent parameter. However, none of these assumptions is really true,
and if non-ideal effects are taken into account the output of the transconductor block
will not be an exact linear function of the voltage.
An equation that fits better the real drain current in the transistor is
" #
VD2
ID ≈ β1 (VFG − VTn )VD − (4.54)
2
Also the mobility μ and consequently β1 depend non-linearly on both VFG and VD ,
the dependence with VFG being dominant. A function that models this approximately
is [117]26
μ0
μ= (4.55)
1 + θ (VGS − VTn )
where μ0 represents the zero-field mobility of carriers and θ accounts for the effect
of the vertical electric field.
An expression for the drain voltage can be obtained by considering that the currents
flowing through transistors M1 and M2 in Fig. 4.1 are the same27 :

−B + B 2 + 2(β2 /β1 )k42 [CGD1 /CT − 1/2 − (β2 /2β1 )gmf 2 R2 ]
outf
VD =
2[CGD1 /CT − 1/2 − (β2 /2β1 )gmf
2 R2 ]
outf
(4.56)
where:
CVDD Cin Cin Vind
B = V + VinCM − VTn +
CT DD CT CT 2
β2
+ gmf Routf (VDD − Vout(N2)Q + gmf Routf Vb − |V Tp |) (4.57)
β1
k4 = VDD − Vout(N2)Q + gmf Routf Vb − |V Tp | (4.58)
Assuming that the differential input signal is sinusoidal with an amplitude A for
each branch, putting these expressions into the equations describing the currents
26 V
GS is, as usual, the gate to source voltage. In this case the gate is floating.
27 V stands for the drain voltage that was called V
D inr in Fig. 4.1. VinCM is the common mode at the
input.
flowing through M1 and M2 and simplifying them using Taylor series expansions,
the nonlinear eq. (4.59) results:
Iout ≈ (c0 d1 + c1 d0 )A + (c0 d3 + c1 d2 + c2 d1 )A3 (4.59)
Where coefficients are defined as follows:

CVDD Cin
c0 ≈ β1i 1 − θ VDD + VinCM − VTn (4.60)
CT CT
Cin
c1 ≈ −β1i θ (4.61)
CT
2
2 Cin
c2 ≈ β1i θ (4.62)
CT
β1i represents the ideal β1 and

CVDD Cin 1 CGD1 2
d0 = VD(Vind =0) VDD + VinCM − VTn − − VD(Vind =0)
CT CT 2 CT
(4.63)

Cin VD(Vind =0) CVDD Cin Cin
d1 = − VDD + VinCM − VTn k5
CT CT CT CT

1 CGD1 Cin
−2 − k5 VD(Vind =0)
2 CT CT

−1
CGD1 1 β2 2 2
× 2 − − g R (4.64)
CT 2 2β1 mf outf
⎛ ⎞
⎜ ((CVDD /CT )VDD + (Cin /CT )VinCM − VTn + (β2 /β1 )gmf Routf k4 ) ⎟
k5 = ⎝1− ( )⎠
2
B(V + (2β /β )k 2 C /C − 1/2 − (β /2β )g 2 R2
ind =0) 2 1 4 GD1 T 2 1 mf outf

B(V =0)
=1− √ind (4.65)
k6

CVDD Cin
d2 = B4 × VDD + VinCM − VTn
CT CT

#
1 CGD1 Cin 2 & '
3/2 −1
−2 − VD(Vind =0) 8k7 k6
2 CT CT

(Cin /CT ) k5 Cin (1/2 − CGD1 /CT ) (Cin /CT ) k5

− + (4.66)
2k7 CT 2k7

CGD1 1 β2 2 2
k7 = − − g R (4.67)
CT 2 2β1 mf outf
β2
B4 = 2 · k42 · · k7 (4.68)
β1

B4 (Cin /CT )2 Cin (1/2 − CGD1 /CT ) (Cin /CT ) k5

d3 = 3/2
+
8k7 k6 CT k7

B4 (Cin /CT )3 × B(V ind =0) CVDD Cin
− 5/2
× VDD + VinCM − VTn
4k7 k6 CT CT

1 CGD1
−2× − × VD(Vind =0) (4.69)
2 CT
Since the structure is fully differential, it can be assumed that the THD is
dominated by the third-order harmonic. If this is the case the THD is given by
1 (c0 d3 + c1 d2 + c2 d1 ) 2
THD ≈ A (4.70)
4 (c0 d1 + c1 d0 )
And, the third-order intermodulation value is

3 (c0 d3 + c1 d2 + c2 d1 ) 2
I M3 ≈ A (4.71)
4 (c0 d1 + c1 d0 )
The following figures (Fig. 4.7 to Fig. 4.9) represent the THD as a function of
various parameters, according to eq. (4.70). Figure 4.7 shows the THD versus two
parameters – input amplitude and input weight28 . As expected (see Chapter 2, Section
2.3.3), the THD decreases as the signal is further attenuated at the FG. However, even
for a relatively high value of the input weight (Cin /CT ) = 0.5, the value of the THD
is still lower than 0.6 per cent.
Also, Fig. 4.8 shows how the increase of the drain to gate capacitive coupling
contributes slightly to the reduction of the THD as Cin /CT increases. The reason
for this is that the nonlinear term in the current expression is now multiplied by
(0.5 − CGD /CT ), instead of 0.5, which reduces the partial derivative of the output
current with respect to VDS2 .
Figure 4.9 shows how the THD increases with the input signal amplitude for
very low values of the tuning voltage29 . This is due to the finite gain of the tuning
amplifier N2. The same happens for large values of Vb . This is because the input
transistor approaches the saturation region and the behaviour of the current becomes
more nonlinear. Hence, the maximum tuning voltage will be limited by the operating
region of the input transistor while the minimum by the tuning amplifier.
28 The values of the parameters used to obtain these surfaces are the same as for the design example
described at the end of the chapter.
29 Always under the assumption that the input transistor is operating in the ohmic region.
0.8
0.6
THD (%)
0.4
0.2
0
1
0.5
0.5 0.4
A(V 0.3
) 0.2
0 0.1 C in/C T
Figure 4.7 THD as a function of an input amplitude and Cin /CT
0.6
0.5
0.4
THD (%)
0.3
0.2
0.1
0.2
0.5
0.1 0.4
C 0.3
GD
/C 0.2
T 0 0.1 C in /C T
Figure 4.8 THD versus CGD /CT and Cin /CT
The previous analysis was carried out under the assumption that the fully differ-
ential structure is completely ideal, which, in general, is not true due to the mismatch
between both differential branches. Mismatch will increase the THD since the second-
order harmonic will not be zero anymore. The effect can be quantified using the same
derivations as before but considering different coefficients for each branch. This will
0.25
0.2
0.15
THD (%)
0.1
0.05
0
0.1
1
0.05
V( 0.5
b V
)
0 0 A(V)
Figure 4.9 THD versus the amplitude of the input signal and the tuning voltage
add the following second-order term to the expression of the current30 :

HD2 ≈ [(c0p d2p + c1p d1p + c2p d0p ) − (c0n d2n + c1n d1n + c2n d0n )]A2 (4.72)
The second-order distortion term defined as the ratio between the second- and first-
order harmonic will then be
HD2 [(c0p d2p + c1p d1p + c2p d0p ) − (c0n d2n + c1n d1n + c2n d0n )]
HD2 = ≈ A
HD1 2(c0n d1n + c1n d0n + c0p d1p + c1p d0p )
(4.73)
HD2 is represented in Fig. 4.10 for an input capacitance weight of 0.25 and a drain
to gate parasitic capacitance ratio of 0.05, as a function of the input signal amplitude
and the percentage of mismatch in the input capacitances of the n and p branches.
Although mismatch in other parameters could also have an effect on the THD, these
are the most significant ones. Table 4.1 shows the value of parameters used in the
calculations. Experimental results of THD are presented from Fig. 4.18 to Fig. 4.22.
4.6.1.1 Noise
The lower limit of the signal that can be handled by the circuit is determined by the
noise. The noise of the transconductor in Fig. 4.3 has three main contributors: the
main cell, the tuning amplifier and the common mode feedback circuit.
30 Subscripts p and n refer to the positive and negative branches.

0.8
0.6
HD2 (%)
0.4
0.2
0
1.5
1 0.1
A(V 0.5 0.05

)
0 /C inn
0 1–C inp
Figure 4.10 Second-order harmonic versus the amplitude of the input signal and
the percentage of mismatch
Table 4.1 Design

parameters
θ 0.234
gmf Routf 76
CGD 8 fF
VinCM 0.75 V
If both single branches in the circuit are assumed to be identical, the noise can
be obtained by analysing the small signal equivalent circuit for one of them and
multiplying the result by a factor of two. As a first approximation, the small signal
circuit with all noise sources included is shown in Fig. 4.11. Although the subscript
r in certain transconductance parameters makes reference to the other single branch,
the effect of the mismatch will be ignored, which is equivalent to assuming that
gmi ≈ gmir . The schematic for the CMFB has not been drawn for the sake of clarity,
| = |S | = S and |S | =
as it is also similar to the N2 block. In the N2 block |Sp1 p2 p n1
| = S and in general for transistor Mi
|Sn2 n

ρgmi
Si = S i = 4KTgmi
2
+γ (4.74)
Wi L i f
where ρ and γ depend on the technology and are different for n- and p-channel devices
[117,213]. An approximate expression of the equivalent noise at the input valid for
gds2 S⬘2 -gm2 vout(N2)

gdsa3 S⬘a3 -kgm2 vout(N2)
vinr
vs3
gdsF1 S⬘1 CCMp C
gnd f
---------------- g
mc3 s 3
v –voutcm – ----------------- g
mc3 vs3
gdsFc3 S⬘c3 CT CT
N1
voutp
C
CMn
-----------------
C
g mc1 voutcm –vs1
gdsFc1 S⬘c1 T
C
VDDf
gds3 S⬘3 -gm2vout(N2) – ------------------- g
C
v –g v
mc1 s1 mbF c1 s1
T
vs1
v D3
gdsa1 S⬘a1 kgm4 vD3r
gds4 S⬘4 gm4 vD3
N3 N4
gdsbias S⬘bias1
vs
gmpvS gdsp S⬘p1 S⬘p2 gdsp gmp vS
vR vout(N2)
(CR/CT)gmnvR gdsFn S⬘n1 S⬘n2 gdsFn (CR/CT)gmnvR
N2
Figure 4.11 Small signal equivalent circuit for noise calculations
low frequencies is given by
vind (Sa1 + Sa3 ) k 2 (S3 + S4 ) Sc3 (R1 )2 + Sc1 (R2 )2

≈2 + +
f Gm 2 Gm2 Gm2
" #
2
2k 2 gm2 2 /(g
(gmp dsFn + gdsp ) )(S1 + S2 )
2
+ (4.75)
Gm 2 (gdsF1 + gds2 + gm2 (gmp /(gdsFn + gdsp )))2
The noise generated by the CMFB is a common mode signal that ideally gets
rejected if both positive and negative branches are identical. Because of this its value
has been neglected from eq. (4.75).
The noise transfer functions for the tuning amplifier and the CMFB are given by
2
vout(N2) 2(Sp + Sn )
≈ (4.76)
f (gdsp + gdsFn )2
2
voutcm 2(Sf 1 + Sf 3 )
≈ (4.77)
f (gdsf 3 + gdsFf 1 )2
The conclusions that can be extracted from the analysis are the following:
1. The FGMOS in the tuning amplifier affects its noise performance in the follow-
ing way: the noise power spectral density is reduced as a consequence of the
capacitive coupling between the drain and gate in the load current mirrors. Nev-
ertheless, this is just an added term in the denominator of eq. (4.76), so ultimately
its effect will depend on how small the output conductances are in comparison
to it.
2. The same is applicable to the CMFB.
3. Regarding the total equivalent noise at the input of the transconductor, eq. (4.75)
shows that the main noise contributors are those transistors that realise the
transconductance function together with their mirrored counterparts (M1, M2,
M3, M4, Ma1 and Ma3, and their symmetrical devices in the other branch). A
strategy to deal with this could be to increase their area. However, this has to be
done with precaution for the input FGMOS devices, since their parasitic capaci-
tances would increase as well which might have other negative consequences (as
it is being described along the book) and affect the operating range. As expected
the noise is proportional to (CT /Cin )2 .
4.7 Power Supply Rejection Ratio (PSRR)
As previously shown in a number of sections of the book, the supply voltages can
directly be used to shift the voltage at the gate of FGMOS devices, or what is the same,
to change the effective threshold voltages. However, this might have a negative effect
on the Power Supply Rejection Ratio (PSRR) of the circuit which is next analysed.
Figure 4.12 shows a small signal equivalent circuit used to calculate the PSRR.
In order to simplify the notation for this circuit:
CM
gmi = (CVDD /CT )gmi , gmc1 = (CCMn /CT )gmc1 ,

gmc1 = (CVDDf /CT )gmc1 , gmc3 = (Cgndf /CT )gmi ,
*
N
CM
gmc3 = (CCMp /CT )gmc3 , gmgi = CVDD gmi + Ci gmi CT (4.78)
i=1
vDD
vDD
g ds2 gm2(vDD-vout(N2))
gdsa3 gma3(vDD -vout(N2))
vinr
vs3
gdsF1 g⬘m1 vDD -gmg1vss
CM CM
gdsFc3 (g⬘mc3+ g )v -g⬘ v – g v
mc3 s3 mc3 SS mc3 outcm
N1
vSS
voutp
vDD
CM
gdsFc1 g⬘mc1vDD + g v
mc1 outcm
-(g⬘mc1+gmbFc1+ gCM )v -v g
mc1 s1 s mbc1
gds3 gm3 (vDD-vout(N2))
vs1
vD3
gdsa1 gma1 (vD3r -vSS)
gds4 gm4(vD3-vss )
vSS
vSS N3 N4
Figure 4.12 Small signal equivalent circuit for PSRR calculations
and vDD , vSS refer to the variations of the positive and negative supply voltage,
respectively. Assuming that the single branches are not identical due to the mismatch,
the following expression for the PSRR+31 results:
Gm Rout
PSRR+ ≈ (4.79)
( vout /vDD |Vind =0 )

vout gm3r gm3
≈ −Routp gma1 (1 − a0n ) + Routn gma1r (1 − a0p )
vDD vind =0 gm4r gm4

− (ACM2p Routp − ACM2n Routn ) − Routp gma3 a0p + Routn gma3r a0n

voutcm
− (ACM1p Routp − ACM1n Routn )
vDD vout CM =0
+ gma3 Routp − gma3r Routn (4.80)
31 R
outp and Routn refer to the output resistance of the positive and negative branch, respectively. Their
expression can be obtained from eq. (4.49) by interchanging the subscripts.
ACM1p was previously defined in eq. (4.50), and ACM2p is given by

CVDDf
ACM2p = gmc1 R2p (4.81)
CT p
a0p depends on the tuning amplifier as

(gm2 − gm1 + gds2 )(gmf Routf )p vout(N2)
a0p = + (4.82)
[gdsF1 + gds2 + gm2 (gmf Routf )p ] vDD vinr =0
ACM1n , ACM2n and a0n can be obtained from equation (4.50) and equation (4.81) just
by replacing the subscripts by those corresponding to the other branch. The second
term is related to VDD rejection ratio as

vout(N2) CVDD gmn gmn 1
= =− ×
vDD vinr =0 CT gdsFn + gdsp gdsFn + gdsp PSRR+(N2)
(4.83)
Equation (4.83) shows how N2 PSRR+ is seriously degraded if VDD is used to shift
the threshold voltage of the load current mirror with the FGs connected together.
A way to compensate for this would be not to connect the FGs, which would dra-
matically improve this value. Nevertheless, this would affect the accuracy and the
output resistance of the mirror. Another solution, if the PSRR+ is an issue, would
be to use a separate voltage source, although, the latter might increase the power
consumption.
A small signal equivalent circuit for PSRR calculations in the CMFB is shown in
Fig. 4.13. In this case, the output gain with respect to the positive voltage supply is32

voutcm
vDD voutCM =0
Cin gmf 1 1
≈ ×
CT gdsFf 1 + gdsf 4 PSRR+CMFB

gmf 4
= (gdsFf2 + gdsf4 )−1 [gmf 3 − gmf 4 ] + 1 +
gmf 3

(gmf 2 +gmbf 2 )gmf 1 (CVDD /CT )f 1 −(gmf 1 +gmbf 1 )gmf 2 (CVDD /CT )f 2
×
gmf 2 +gmbf 2 +gmf 1 +gmbf 1
(4.84)
where again the mismatch has been neglected in the expression of the differential gain.
This is justified for two reasons: first, it increases the clarity, and second, because
the differential gain is expected to be high and thus this assumption is not going to
affect much the final result. CVDD /CT is the weight associated with VDD in the input
32 Assuming g
dsFf (CVDD /CT )gmf .
vDD
gmf3(vDD-vout1) gdsf3 gdsf4 gmf4(vDD-vout1)
vout1 voutcm
(CVDD/CT)f1gmf1vDD-gmsf1vS gdsFf1 gdsFf2 (CVDD/CT)f2gmf2vDD-gmsf2vS

+g mbFf1vSS +( gmbFf2+(Cgnd /CT)f2 gmf 2)vSS
vS
gdsbiasc
C VDD 2C in C VDD C gnd
g msf1 = --------------- + ------------ g mf1 + g mbFf1 g msf2 = --------------- + ------------ g mf2 + g mbFf2
CT CT CT CT
vSS
Figure 4.13 Small signal equivalent circuit for PSRR calculations in the CMFB
transistors. The weights are expected to be different because of the mismatch, and
that is why subscripts f1 and f2 have been added in their names.
In order to have an idea of the order of magnitude predicted by eq. (4.84), and
also isolate the contribution of the FGMOS to the PSRR+ in the circuit, let us assume
only mismatch between the input capacitances. In this case, the PSRR+ predicted for
the CMFB is

−1
Cin CVDD CVDD
PSRR+CMFB ≈ − (4.85)
CT CT f1 CT f2
If for example CVDD = Cin and the mismatch is 1 per cent, the PSRR+CMFB
value would be 40 dB. This value would increase if VDD is also used to gener-
ate the reference voltage, in which case the second term in eq. (4.85) would be
larger.
Going back to the transconductor circuit, if all these expressions are substituted
into eq. (4.79) and again only mismatch between input capacitances is assumed,
equal to 1 per cent, the transconductances and input weights are assumed to have
the same value, and N2 DC gain is 10, the PSRR+ for the whole circuit would be
just 14 dB. This very low value is caused by the tuning amplifier N2, because of
the aforementioned reasons. However, if the FGs are not connected together, or the
threshold voltage shift is performed without using VDD , the PSRR+ would increase to
36 dB and it would mainly be caused by the connection to VDD in the transconductor
FGMOS input transistors33 .
33 In the design example at the end of the chapter the FGs were not connected together.
The same analysis performed for the PSRR− yields

gm3r gm3
PSRR− ≈ Gm Rout −Routp gmal 1 − a1n + Routn gmalr 1 − alp
gm4r gm4
− (ACM3p Routp − ACM3n Routn ) − Routp gma3 alp + Routn gma3r aln

voutcm
− (ACM1p Routp − ACM1n Routn )
vss voutCM =0
−1
+ gmal Routp − gma1r Routn (4.86)
Cgndf
ACM3p = (gmc3 × R1p ) + (gmbFc1 × R2p ) (4.87)
CT

(gmg1 + gdsF1 )(gmf Routf )p vout(N2)
a1p = + (4.88)
[gdsF1 + gds2 + gm2 (gmf Routf )p ] vSS v
inr =0

voutcm
vss voutCM =0

gmf 4
≈ 1+
gmf 3
$ %
(gmf 2 + gmbf 2 )gmbFf 1 − (gmf 1 + gmbf 1 ) gmbFf 2 + (Cgnd /CT )gmf 2
×
(gmf 1 + gmf 2 + gmbf 1 + gmbf 2 )(gdsf 4 + gdsFf 2)
(4.89)
where Cgnd /CT is the weight associated with the ground connected input in the CMFB.
In this case, the main contribution to the PSRR− comes from the CMFB, due
to the connection to ground used to generate the reference voltage. However, as the
gain from the common mode feedback nodes to the output is much lower than from
the input to the output, the degradation of the PSRR− caused by the FGMOS is not
going to be as critical as the degradation of the PSRR+. Hence, for example, under the
same assumptions as before and also, considering that the output conductances have
the same values and, the same for the transconductances, and the mismatch between
both branches capacitances connected to the CMFB is 1 per cent, the PSRR− would
be over 36 dB. Improvements can be achieved by generating the reference voltage
without using the negative voltage supply, but again this would compromise the power
consumption.
4.8 Filter example
This section illustrates the operation of the previously described integrator when it
is used as the main building block in a second-order lowpass/bandpass filter. The
V2 V3
U1
C C
U2
V1 V4
Figure 4.14 Biquad circuit
starting point for the design is the state space equations

ẋ1 = −2ko x1 − ko x2 + ko u (4.90)
ẋ2 = ko x1 (4.91)
The lowpass (LP) and bandpass (BP) transfer function are given by X2 (s) and X1 (s),
respectively. The filter has a fixed quality factor of 0.5 and a cut-off frequency
ω0 = k0 34 :
ko2
X2 (s) = U (s) (4.92)
s2 + 2ko s + ko2
sko
X1 (s) = U (s) (4.93)
s2 + 2ko s + ko2
The block diagram of a circuit that implements the state space equations is shown
in Fig. 4.14, where x1 = (V1 − V2 ), x2 = (V3 − V4 ) and u = (U1 − U2 ). If the values
of the input weights are the same for all inputs, this is Ci = Cin for i = [1, N ], ko is
given by
kβ1 Cin Vb
ko = (4.94)
CT C
The filter was designed in a 0.8 μm CMOS technology [214] with nominal thresh-
old voltages of around 0.8 V. The transistor sizes and values of the capacitances used
in the design are collected in Table 4.235 .
A microphotograph of the fabricated circuit is shown in Fig. 4.15. The total area
is 0.13 mm2 . Experimental frequency responses for different LP and BP functions are
illustrated in Fig. 4.16 and Fig. 4.17. They are obtained for different tuning voltages,
Vb , in the range from 20 to 150 mV. The THD measured for different input levels is
illustrated from Fig. 4.18 to Fig. 4.21. The maximum signal level is not determined
34 The quality factor and gain could have been made programmable as well just by choosing different
coefficients in the state space equations.
35 Variables refer to Fig. 4.3.
Table 4.2 Design sizes

in μm and
capacitances
values
M1 4.5/2.5
M2 4.5/3
Ma3 2.5/14
Mc1 5/3
Mc3 7/2
Cin 66 fF
C 5 pF
Figure 4.15 Microphotograph of the biquad prototype
by the nonlinearities in the input devices but by output swing limitations. Still the
LP output exhibits less than 0.4 per cent distortion for 2V pp differential input signal
at 1.5 V voltage supply and as predicted it is mainly caused by the second-order
harmonic. Figure 4.22 shows experimental results for a two-tone test that measures
the IM3 in the BP filter. Figure 4.23 illustrates the performance of the CMFB showing
how the common mode level remains at 0.75 V. The transient response of the common
mode at the output is shown in Figure 4.24 when a 1V pp differential input signal is
applied. It can be seen how the CMRR remains below −40 dB. A summary of the
filter performance is given in Table 4.3.
The FGMOS transistor biased in the strong inversion ohmic region can be used to
perform LV linear voltage to current conversion. The advantages of using FGMOS
devices in this type of topologies are: a) The effective threshold voltage can be reduced
and hence the valid input signal range can be increased since the transistor remains in
the strong inversion region even for low/high (depending on whether it is an n- or a
p-channel device) values of the input signal; b) The effective input signal at the FG is
20
dB
LogMag
10
dB
/div
–20
dB 100 1k 10k 100k
Start: 75 Hz Stop: 100 000 Hz
Figure 4.16 Experimental LP filter frequency response (Vb (V ) = 0.02, 0.07,

0.1, 0.15)
–5
dBm
LogMag
5
dB
/div
–20
dBm 10 100 1k 10k 100k
Figure 4.17 Experimental BP filter frequency response (Vb (V ) = 0.02, 0.07,

0.1, 0.15)
1.0
0.8
0.6
THD(%)
0.4
0.2
0
0 0.5 1.0 1.5 2.0
A(Vpp)
Figure 4.18 THD for a LP function with fo = 1.4 kHz and a 500 Hz sinusoidal input
1.0
0.8
0.6
THD(%)
0.4
0.2
0
0 0.5 1.0 1.5 2.0
A(Vpp)
Figure 4.19 THD for a BP function with fo = 2 kHz and a 2 kHz sinusoidal input
scaled down by the input weight, which translates into a lower distortion and hence
a larger input range; c) The gate to drain coupling in the input transistor contributes
to the reduction of the THD since it appears subtracted from the coefficient of the
quadratic term in the transistor current law; d) The output resistance can be increased
by using two-input FGMOS cascode transistors, which can also simultaneously be
used in the common mode feedback strategy and e) Multiple input FGMOS can be
Format
Format
Single
Single Up/Dn
Up/Dn
Marker
On Off Trk
Marker width
Norm Wide
Spot
Marker seeks
Max Min
Mean
Grid div/scrn
Off 8 10
Graph style
Stop –18 dbU No Aug. Line Fill
Figure 4.20 THD for a sinusoidal input of 15 mVpp in a LP filter with f0 = 2 kHz
Frequency
899.994 Hz
Level
1000 mU
Return:
RUN –18 dbU No Aug.
Figure 4.21 THD for a sinusoidal input of 1 V in a LP filter with f0 = 2 kHz
used to perform signal averaging within a single transistor in a CMFB. This results in
a very compact, lower power and a more linear topology (when compared with other
existing ones).
But also, the transistor presents several drawbacks that the designer should bear in
mind. These are as follows: a) As the number of inputs, N , increases, the equivalent
Off
Sine
2-Tone
ON
Noise
Chirp
Configure
RUN –18 dbU No Aug. Source:
Figure 4.22 Spectrum at the output for a two harmonics input in a BP filter with
fo = 2 kHz
Tek T Trig’d M Pos: 49.95 ms CH1
Acoplamiento
CC
Limitar
Ancho Banda
20MHz
Ganancia
variable
Gruesa
Sonda
10X
2+ Invertir
NO
CH1 200 mVBw CH2 200 mVBw M 1.00 ms CH1
Figure 4.23 Single outputs transient response
transconductance for each input decreases as approximately 1/N for the same level
of power; b) The DR does not improve much since the equivalent input noise is scaled
up by the inverse of the input weight; c) Using the voltage supplies to shift effective
threshold voltages can seriously degrade the PSRR; d) The mismatch between the
Tek T Trig’d M Pos: 5.300 ms CH1
Acoplamiento
CC
Limitar
Ancho Banda
NO 20MHz
Ganancia
variable
Gruesa
Sonda
M+ 10X
Invertir
NO
CH1 200 mVBwCH2 20.0 mVBwM 250μs CH1 /0.00V
Figure 4.24 Common mode signal sensed at the output for a differential input
of 1 V pp
Table 4.3 Summary of filter performance parameters
Technology 0.8 μm, AMS(CXQ)

VTn , VTp 0.82 V, −0.8 V
Voltage supply 1.5 V
Area 0.13 mm2
fo , min 380 Hz
fo , max 2 kHz
Vb Range 20 mV, 150 mV
THD (Vpp < 2 V, (0.38–2) kHz) <−40 dB
IM3 <−40 dB
DR (THD < 1%) >50 dB
PSRR >40 dB
CMRR >40 dB
Maximum power (Vb = 0.15 V) 13 μW
input capacitances is going to add extra components to the common mode gain and
the gain from the power supplies, which will degrade the CMRR and PSRR and e)
There is a limitation for the maximum transconductance that can be achieved with
these topologies, imposed by the maximum drain voltage that keeps the transistor
in the ohmic region, and the maximum width of the device. The reason why there
is a maximum value for the width above which there is no benefit on increasing it
further is because, a larger width also means larger parasitic capacitances. This would
reduce the input weights (because CT increases) which in its turn would cause the
transconductance to decrease, apart from having other undesired effects in the fre-
quency response. An initial solution for this could be to increase the value of the input
capacitances, but this would add extra area and would also have a loading effect when
the transconductor is used in higher order configurations that would limit the maxi-
mum speed. Therefore, this kind of topology is only recommended for low-medium
speed applications.
Notation
β1i In analysis of mismatch ideal β1 in absence of mismatch
βi β parameter for transistor Mi
βn β parameter for an n-channel transistor (eq. (4.2))
ε See eqs. (4.26) and (4.27)
θ Mobility parameter
ωo Filter cut-off frequency (rad/s)
A Amplitude of a sinusoidal input signal
Ao N2 DC gain
ao Gain from vout(N2) to vR (eq. (4.38))
a0n Parameter equivalent to a0p in the n side of the integrator
a0p See eq. (4.82)
A1 See eq. (4.34)
a1n Equivalent to a1p but in the n side of the transconductor
a1p See eq. (4.88)
ACM ACMp and ACMn in the absence of mismatch
ACMn Equivalent to ACMp but in the n side of the transconductor
ACMp See eq. (4.50)
ACM1n Equivalent to ACM1p but in the n side of the transconductor
ACM1p See eq. (4.50)
AV N2 small signal gain
B See eq. (4.9)
B See eq. (4.57)
B1 See eq. (4.35)
B4 See eq. (4.68)
C Integrating capacitance
c0 See eq. (4.60)
C02 See eq. (4.30)
C22 See eq. (4.31)
C42 See eq. (4.32)
c1 See eq. (4.61)
C1 See eq. (4.18)

(C1 /CT )n (C1 /CT ) in the n branch of the fully differential
integrator (Fig. 4.3) when mismatch is assumed
(C1 /CT )p (C1 /CT ) in the p branch of the fully differential
c2 See eq. (4.62)
C2 See eq. (4.19)
C3 See eq. (4.20)
CCMn Capacitance connected to Voutcm in n-channel
cascode transistors (Fig. 4.3)
(CCMn /CT )n (CCMn /CT ) in the n branch of the fully differential
(CCMp /CT + Cgndf /CT )n (CCMp /CT + Cgndf /CT ) in the n branch of the fully
differential integrator (Fig. 4.3) when mismatch is
assumed
CCMp Capacitance connected to Voutcm in p-channel
(CCMn /CT )p (CCMn /CT ) in the p branch of the fully differential
(CCMp /CT + Cgndf /CT )p (CCMp /CT + Cgndf /CT ) in the p branch of the fully
differential integrator (Fig. 4.3) when mismatch is
assumed
CDBi Drain to bulk coupling capacitance in transistor Mi
Ceq See eq. (4.25)
CGDi Gate to drain coupling capacitance in transistor Mi
Cgnd Value of the input capacitance connected to ground
in the CMFB
Cgndf Capacitance connected to ground in p-channel
Ci for i = [1, N ] M1 input capacitances (N1; Fig. 4.1)
Cin Effective input capacitance when only one effective
input is considered in N1
Cinn Effective input capacitance in the n branch when
mismatch is assumed (Fig. 4.10)
(Cin /CT )n Effective input weight in the n branch when
mismatch is assumed (Fig. 4.3, Fig. 4.10)
Cinp Effective input capacitance in the p branch when
mismatch is assumed (Fig. 4.10)
(Cin /CT )p Effective input weight in the p branch when
mismatch is assumed (Fig. 4.3, Fig. 4.10)
CL Capacitive load at the output of N2
CLN2 Capacitive load in N1 due to the connection with the
tuning amplifier N2
CR Capacitance connected to VR in FGMOS transistors
Mn in N2 (Fig. 4.2)
CT Total capacitance for all the FGMOS transistors unless the

opposite is said
CVDD Capacitance connected to VDD in the transconductor FGMOS
input transistors and also in Mn in N2 and in the CMFB
(Fig. 4.1, Fig. 4.2)
(CVDD /CT )fi (CVDD /CT ) when mismatch is considered in the CMFB
for i = [1, 2]
CVDDf Capacitance connected to VDD in n-channel cascode
transistors (Fig. 4.3)
(CVDDf /CT )n (CVDDf /CT ) in the n branch of the fully differential
integrator when mismatch is taken into account
(CVDDf /CT )p (CVDDf /CT ) in the p branch of the fully differential
integrator when mismatch is taken into account
D See eq. (4.8)
d0 See eq. (4.63)
D1 See eq. (4.36)
d1 See eq. (4.64)
d2 See eq. (4.66)
d3 See eq. (4.69)
fo Filter cut-off frequency (Hz)
fomax Maximum value of the cut-off frequency in the tuning range
fomin Minimum value of the cut-off frequency in the tuning range
gdsbias Output conductance of the bias current source in N2
gdsbiasc Output conductance of the bias current source in the CMFB
gdsi Transistor Mi output conductance
gdsFi Drain conductance for FGMOS transistor Mi
gmbFi Bulk transconductance for FGMOS transistor Mi
CM
gmc1 See eq. (4.78)

gmc1 See eq. (4.78)
CM
gmc3 See eq. (4.78)

gmc3 See eq. (4.78)
gmgi See eq. (4.78)
gmf N2 transconductance (eq. (4.6))
(gmf Routf )n Gain of the N2 block in the n branch when mismatch is
assumed
(gmf Routf )p Gain of the N2 block in the p branch when mismatch is
assumed
Gmi Transconductance associated to effective input i (eq. (4.4))
gmi Transistor Mi transconductance

gmi See eq. (4.78)
gmsfi for i = [1, 2] Fig. 4.13
Goutn Output conductance for the n branch in the fully differential
transconductor
Goutp Output conductance for the p branch in the fully differential
transconductor (Fig. 4.3) when mismatch is assumed
HBP (ωo ) Bandpass filter gain at the cut-off frequency

HLP (0) Lowpass filter DC gain
IB2 Bias current in the tuning amplifier (Fig. 4.2)
ID M1 drain current (N1. Fig. 4.1)
Iout Fully differential integrator output current (Fig. 4.3)
iouti Ideal fully differential integrator small signal output current
when N2 gain is infinity (eq. (4.26))
ioutr Fully differential integrator small signal output current when
N2 gain is not ideal (eq. (4.26))
k Scaling factor for current mirrors in Fig. 4.3
ko Coefficient of the state variables and inputs in eq. (4.90).
Given in eq. (4.94)
k4 See eq. (4.58)
k5 See eq. (4.65)
k6 See eq. (4.65)
k7 See eq. (4.67)
N Number of inputs in an FGMOS transistor (N1 block; Fig. 4.1)
p N2 pole
p1 Transmission pole from vout(N2) to vR (eq. (4.39))
Q Filter quality factor
R1 R1n and R1p when no mismatch is assumed
R1n Parameter equivalent to R1p for the n side of the transconductor
R1p See eq. (4.49)
R2 R2n and R2p in the absence of mismatch
R2n Parameter equivalent to R2p for the n side of the transconductor
R2p See eq. (4.49)
Rout Output resistance in the fully differential transconductor
(Fig. 4.3) when no mismatch is assumed
Routf N2 output resistance (eq. (4.6))
Routn Output resistance for the n branch in the fully differential
Routp Output resistance for the p branch in the fully differential

si = si2
si Power spectral density of noise
u Filter input (eq. (4.90). (u1 -u2 ))
U1 Single voltage related to u
U2 Single voltage related to u
V1 Single voltage related to state variable x1
Vb Reference voltage in N2 (Fig. 4.1)
VD M1 drain voltage (Fig. 4.2)
VD3 M3 drain voltage (Fig. 4.3)
VD3r M3r drain voltage (Fig. 4.3)

VFG M1 FG voltage
VFGcn FG voltage for n-channel cascode transistor in the integrator in
Fig. 4.3 (eq. (4.12))
VFGcp FG voltage for p-channel cascode transistor in the integrator in
Fig. 4.3 (eq. (4.13))
Vi for i = [1, N ] FGMOS transistor voltage inputs in N1 (Fig. 4.1)
Vin Voltage at N1 effective input when only one input is considered
Vind Differential input in the transconductor in Fig. 4.3 when only
one effective input is considered in each input FGMOS
transistor
VinCM Common mode at the input (when a single input is considered)
Vinr M1 drain voltage (N1. Fig. 4.1)
Vjn for j = [1, N ] M1r effective inputs – branch n (Fig. 4.3)
Vjp for j = [1, N ] M1 effective inputs – branch p (Fig. 4.3)
Vout Differential output for the transconductor in Fig. 4.3
Vout1 Mf1 drain voltage (CMFB)
Vout(N2) N2 output (Fig. 4.1)
Vout(N2)Q Value of Vout(N2) when Vinr = Vb (eq. (4.6))
VoutCM Common mode at the output
Voutcm Output of the CMFB (Fig. 4.4)
Voutn Single output in the fully differential multiple input integrator
(Fig. 4.3)
Voutp Single output in the fully differential multiple input integrator
(Fig. 4.3)
VR Drain voltage of the input transistor in the tuning amplifier N2
(Fig. 4.2)
VS Variable used to refer to the source voltage in the differential
pairs (N2 and CMFB)
Vs1n Mc1r source voltage (Fig. 4.3)
Vs1p Mc1 source voltage (Fig. 4.3)
Vs3n Mc3r source voltage (Fig. 4.3)
Vs3p Mc3 source voltage (Fig. 4.3)
VTn n-channel transistors threshold voltage. The same is assumed
for all the devices unless the opposite is said

VTn Effective threshold voltages for n-channel cascode transistor
(eq. (4.16))

VTp Effective threshold voltages for p-channel cascode transistor
(eq. (4.17))
|VTP | p-channel transistors threshold voltage (the same for all the
devices unless the opposite is said)
x1 State variable related to bandpass output in eq. (4.90)
x2 State variable related to lowpass output in eq. (4.90)
z N2 zero
z1 Zero in the transfer function between vout(N2) and vR
(eq. (4.40))
Chapter 5
Low power analog continuous-time filtering
based on the FGMOS in the strong inversion
saturation region
5.1 Introduction
This chapter presents filter designs that use the characteristic features of the FGMOS
transistor operating in the strong inversion saturation region to be able to operate at
a very low power supply voltage with also a very low power consumption. Again, the
implementation of the mathematical functions is greatly improved thanks to the extra
degrees of freedom that the device provides, which results in a fewer number of power
consuming elements. Besides, the new effective threshold voltages make possible to
operate with signal levels that would have switched normal MOS transistors off,
hence enabling the reduction of the power supply voltage, while still keeping the
signal ranges.
The chapter also analyses the drawbacks of using FGMOS transistors and based
on this analysis advises the reader on when not to use them. Thus, for example it
shows how in certain cases, the PSRR can be seriously degraded, and if it is a design
constraint the transistor should be used in a different way, or even avoided.
Two different integrator designs illustrate the transistor performance. The first
one is an audio range linearised Gm -C block in which the linearisation is obtained
by combining the quadratic law in different FGMOS devices working in the strong
inversion saturation region. The second integrator is again a Gm -C structure but
operating in an intermediate frequency (IF) range under a very restrictive power
supply voltage constraint.
Finally, the chapter also describes what happens when the input capacitances in
the FGMOS transistors are realised in metal/poly instead of poly2/poly1, using as an
example the second integrator/filter block. The purpose is to illustrate what would
occur in those cases in which only purely digital technologies are available.
VDD VDD
M9 M10
IB VS7 Vcasp VS8 IB
M7 M8
Vout1 Vout2 CMFB

C Voutcm
Iout
I1 M5 M6 I2
VD1 Vcasn VD3
VDD VDD VDD VDD
Vb2 Vb1 Vb1 Vb2
Vin1 Vin2 Vin1 Vin2 Vin1
M1 M2 M3 M4 Gm C
Vin2
(a) (b)
Figure 5.1 (a) Schematic of the proposed FGMOS-based integrator. (b) Symbol
5.2 A Gm -C integrator based on the FGMOS in the strong inversion

saturation region for audio applications
This section describes an easy way to perform voltage to current linear conversion
based on the operation of the FGMOS transistor in the strong inversion saturation
region. It also explains how to use the transistor to maximise the input signal range
under low voltage constraints.
Figure 5.1 shows the schematic of a linear Gm -C integrator that uses FGMOS
transistors at the input to achieve the linear operation. The circuit works as follows:
the input stage is a pseudodifferential amplifier comprised of four equally sized tran-
sistors, M1–M4, working as two cross-coupled pairs. Transistors M5–M10 act as
active load36 . A different combination of one input signal and one bias signal is
connected to each input transistor in a way that both inputs and both bias are con-
nected to both branches. The aspect ratios and input capacitors are the same for all of
them. The bias voltages Vb1 and Vb2 are different. A third input is used as part of the
common mode feedback mechanism. The fourth input is connected to the maximum
available voltage, VDD , in order to reduce the effective threshold voltage according to
36 The block labelled CMFB represents the circuit for the output common mode control, which will be
described later on.
Filtering with FGMOS in the strong inversion saturation region 131
eq. (3.3). When the transistors are in the saturation region, the output current (Iout ) is
given by
I1 − I2
Iout = I1 − IB = IB − I2 = (5.1)
2
I1 + I2
IB = (5.2)
2
Assuming that transistors M1–M4 are biased in the strong inversion saturation region,
and using eq. (2.9), eq. (5.1) can be rewritten as a function of the input voltages:
I1 − I2 β1 Cin Cc (Vb2 − Vb1 )
Iout = = · (Vin1 − Vin2 ) (5.3)
2 2CT2
where β1 is the input transistors β parameter37 , Cc is the capacitance connected to

either Vb1 or Vb2 (depending on the transistor) and Cin is the capacitance connected
to either Vin1 or Vin2 . Equation (5.3) is the function of a linear V /I converter whose
output is the current Iout , and the input is the differential signal (Vin1 − Vin2 ). The
transconductance is given by
β1 Cin Cc (Vb2 − Vb1 )
Gm = (5.4)
2CT2
Gm can be tuned with voltages Vb2 and Vb1 . Equation (5.4) shows how values of
negative transconductances can also be obtained if necessary. Input capacitances and
bias voltages values have to be selected according to eq. (2.9) in order to maintain
the MOS transistors working in the strong inversion saturation region.
One of the advantages of the topology in Fig. 5.1 is that the transconductance is
completely independent of either the common mode signal or the threshold voltage
or the current IB . Equation (5.4) is valid whenever the input transistors are in the
strong inversion saturation region and the second-order effects can be neglected in
the quadratic current law. Hence, the effective threshold voltage can be shifted to
whichever value is convenient to keep the input transistors in strong inversion by
means of the extra input connected to VDD . Besides, the bias current does not need to
be constant, which means that it can be adapted to control the common mode at the
output as it will be explained later on. In the schematic in Fig. 5.1, IB is generated by
a voltage signal coming from the CMFB. This signal is also connected to one of the
inputs in the FGMOS transistors. In this way, it appears as a common mode signal
that does not have any effect on the transconductance but however helps to keep the
transistors in the right operating region. Hence, if, for example, the bias voltages,
Vb1 and Vb2 increase during the tuning phase, the constraint for the strong inversion
saturation becomes more restrictive. The common mode at the output would tend to
decrease in order to balance out the currents between the top and bottom part of the
circuit, but the CMFB will correct this by reducing Voutcm and hence increasing IB .
37 From now on β corresponds to the β parameter for transistor Mi.

i
In this way, the positive variation of Vb1 /Vb2 will be compensated with a negative
variation of Voutcm .
5.2.1 Design trade-offs. Power consumption, voltage supply and limits for
the transconductance
This section analyses the design trade-offs between power consumption, voltage
supply and speed in the transconductor in Fig. 5.1. The speed is proportional to the
value of the transconductance. Therefore, the trade-offs for speed are equivalent to
the trade-offs for the latter. The maximum value for the transconductance is given by
β1 Cin Cc (Vb2max − Vb1min )
Gm(max) = (5.5)
2CT2
And the minimum for a given aspect ratio depends on how small the difference
(Vb2 − Vb1 ) can be. Nevertheless, the maximum will be the most limiting value.
The maximum and minimum values for the power consumption are a function of
the current IB flowing through transistors M7–M10. IB is given by38
β10
IB = (VDD − Voutcm − |VTp |)2 (5.6)
2
Its maximum and minimum values can be written as functions of the maximum and
minimum values of the voltages at the inputs of the FGMOS transistors M1–M4, and
the maximum and minimum value of the voltage at the output of the CMFB:
2
IBmax β1 CCM Cin Cc
= Voutcm(min) + (VDD − 0.2) + Vb − VTFG
2 2 CT CT CT
β10
= (VDD − Voutcm(min) − |VTp |)2 (5.7)
4
2
IBmin β1 CCM 0.2Cin Cc
= Voutcm(max) + + Vb − VTFG
2 2 CT CT CT
β10
= (VDD − Voutcm(max) − |VTp |)2 (5.8)
4
where
CVDD
VTFG = VTn − VDD (5.9)
CT
Vb1 = Vb2 = Vb (5.10)
CVDD and CCM are the value of the input capacitances which are connected to VDD and
Voutcm , respectively. Equation (5.7) has been obtained assuming that the maximum
38 From now on |V | will refer to the threshold voltage of p-channel devices. The same value will be
Tp
assumed for all the devices unless the opposite is said. The same applies for VTn and n-channel transistors.
signal range is limited by the output swing (which is justified, as in general the
transconductor is not going to be used on its own), and the minimum VDS required
by the p-channel devices is 0.1 V. Also, transistors M9–M10 determine the IB current;
therefore, its maximum value will be sourced when Voutcm is minimum (Voutcm(min) )
and the minimum when Voutcm is maximum (Voutcm(max) ). From eq. (5.7) and (5.8),
the limits for Voutcm are given by
1
Voutcm(min) = √
CCM /CT + β10 /2β1
!
β10 Cin Cc
× (VDD − |VTp |) − (VDD − 0.2) − Vb + VTFG
2β1 CT CT
(5.11)
1
Voutcm(max) = √
CCM /CT + β10 /2β1
!
β10 0.2Cin Cc
× (VDD − |VTp |) − − Vb + VTFG (5.12)
2β1 CT CT
And the maximum and minimum power:

⎧
⎨Pmin = β10 VDD (VDD − Voutcm(max) − |VTp |)2
Plimits = (5.13)
⎩P
max = β10 VDD (VDD − Voutcm(min) − |VTp |)
2
The limits for the power are represented in Fig. 5.2. The vertical plane represents
the maximum available tuning voltage (VDD ). Hence, for example, for Vb = 1.25 V
and VDD = 2 V, the maximum power consumption required for the circuit to operate
Pmin (W) Pmax (W)
–5
10
–5
10
2.5 2.5
2 2.5 2 2.5
2 2
V b (V) 1.5 1.5 V b (V) 1.5 1.5 V DD (V)
V DD (V)
Figure 5.2 Limits for the power established by the tuning voltage and the voltage
supply for an input range from 0.2 V to (VDD − 0.2) V
3.5
3
Vbmax
2.5
2
Vb (V)
Vbmax=VDD
1.5
Vbmin
1
0.5
0
1 1.5 2 2.5 3
V DD (V)
Figure 5.3 Vb limits versus the voltage supply for the maximum and minimum bias
current
within the whole tuning range would be 7 μW, whereas the minimum power it would
need for the minimum tuning signal value would be 1.8 μW.
Once the limits of power required by the circuit in order to bias the devices
properly have been established, the maximum value of the transconductance can
also be obtained as follows: equations (5.11) and (5.12) are only valid when tran-
sistors M9–M10 are in the strong inversion saturation region and also Voutcm ≥ 0.
The maximum and minimum values of Vb that satisfy this are
" ! #
CT Cin β10
Vbmax ≤ − (VDD − 0.2) + VTFG + (VDD − |VTp |) (5.14)
Cc CT 2β1
CT CCM 0.2Cin
Vbmin ≥ − (VDD −|VTp |) − + VTFG (5.15)
Cc CT CT
Besides Vbmax ≤ VDD . These trade–offs are represented in Fig. 5.3. The dashed
line corresponding to Vbmax ≤ VDD shows how this condition is more restrictive
than eq. (5.20) for this particular design. Hence, the maximum possible difference
between the tuning voltages Vb1 and Vb2 is the difference between the two bottom
curves in Fig. 5.3. Combining these equations with eq. (5.5) the maximum limit for
the transconductance can be obtained. This result is represented in Fig. 5.4.
The previous derivations have assumed certain values for the sizes and technolog-
ical parameters39 . However, different ones could have been chosen if the important
trade-offs or the technology were different.
39 Those corresponding to the design example at the end of section.

Gm(max) (μA/V)
2
0
1 1.5 2 2.5 3
V DD (V)
Figure 5.4 Gm(max) versus power supply
VDD
M14 M15 Voutcm
VDD VDD
Vout1 VDD
M12 M13
Vout2 gnd
Vbias3 M11
Figure 5.5 Common mode feedback circuit (CMFB)
5.2.2 The Common-Mode Feedback Circuit (CMFB): effects on the OTA

performance
The functionality of the integrator in Fig. 5.1 is based on its fully balanced configu-
ration and as such will require a CMFB to stabilise the common mode. A CMFB that
performs well with this kind of structure is shown in Fig. 5.5. It is based on the same
principle as the CMFB described in the previous chapter. The difference between the
two of them is the diode connected p-channel transistors used in Fig. 5.5 and therefore
the lower gain of the latter.
The output of the CMFB (Voutcm ) is simultaneously fed back to transistors
M1–M4 and M9–M10. Hence, the DC gain from voutcm to one of the single
gds10 –gm10voutcm
vS8
gds8 (gm8+gmb8)vS8
vout2
gds6 –(gm6+g mb6 )vD3
vD3
(gdsF3 +gdsF4) (gm3+gm4)voutcm(CCM/CT)
Figure 5.6 Small signal equivalent circuit for a single branch in the transconductor
of Fig. 5.1. The input is the CMFB output, voutcm
outputs is
vout2 [(gm3 + gm4 )(CCM /CT ) + g m10 ]
= (5.16)
voutcm Goutn + Goutp
with
(g dsF3 + g dsF4 ) · g ds6 g ds8 · g ds10
Goutn = , Goutp = (5.17)
(g m6 + g mb6 ) (g m8 + g mb8 )
The small signal equivalent circuit for one of the circuit branches in Fig. 5.1 is shown
in Fig. 5.6. The subscripts refer to the name of the corresponding device.
The previous equations have been obtained under the assumptions
(g m6 + g mb6 )
(gds6 + g dsF3 + g dsF4 ) (5.18)
(g m8 + g mb8 )
(gds8 + gds10 ) (5.19)
Equation (5.16) shows how the gain is increased by applying feedback in the p-
and n-channel transistors. In this way, it is possible to have a low gain CMFB and the
variations at the output of the latter can be kept small enough to not exceed the output
swing constraint, which becomes very restrictive as the value of the voltage supply
is reduced.
The rationale behind this feedback mechanism can also be explained from a large
signal point of view: when the tuning voltages (Vb2 , Vb1 ), or the common mode
at the input (VinCM ) change, the common mode current does it as well. When this
happens the output of the CMFB changes in such a way that the variation of the input
550.0 m
500.0 m VoutCM = 0.48V
(a)
450.0 m
400.0 m
(b) VoutCM = 0.375 V
350.0 m
87.50 m
87.0 m
(c)
86.50 m VoutCM = 0.0867 V
86.0 m
4.0 m 5.0 m 6.0 m 7.0 m 8.0 m 9.0 m 1

time (lin)
Figure 5.7 (a) Evolution of the output when a sinusoidal input is applied to the
transconductor. Feedback at the load and input transistors. (b) Evolution
of the output when a sinusoidal input is applied to the transconductor.
Feedback at the load transistors. (c) Evolution of the output when a
sinusoidal input is applied to the transconductor. No feedback
transistors FG common mode operating point is up to a certain extent compensated.

Simultaneously, the IB current changes to correct for the still remaining variation.
Figure 5.7 illustrates this showing how a single output changes when a sinusoidal
differential input is applied. The DC level of the output signal represents common
mode at the output, VoutCM . The reference voltage it is being compared with is 0.5 V.
It can be observed how when only the gate of the p-channel transistors is connected to
the output of the common mode circuit, the common mode at the output changes more
abruptly than when the feedback mechanism is applied to both, the p and the n parts.
Another interesting aspect worth analysing is how this feedback might affect
the frequency response of the integrator. If the frequency response of the CMFB is
modelled as a single pole function and included into the small signal equations of the
integrator, for every single branch
a02
sCvout2 = −gm3 vinl − gm4 vin2 − (vout1 + vout2 )
s+p

− + vout2 (5.20)
(g m6 + g mb6 ) (g m8 + g mb8 )
a01
sCvout1 = −gm1 vinl − gm2 vin2 − (vout1 + vout2 )
s+p

− + vout1 (5.21)
(g m5 + g mb5 ) (g m7 + g mb7 )
with a02 and a01 being

[(CCM /CT )(gm3 +gm4 )+gm10 ]
a02 = (Ao · p)
((gdsF3 +gdsF4 ) · gds6 )/(gm6 +gmb6 )+(gds8 · gds10 )/(gm8 +gmb8 )
(5.22)
[(CCM /CT )(gm1 +gm2 )+gm9 ]
a01 = (Ao · p)
((gdsF1 +gdsF2 ) · gds5 )/(gm5 +gmb5 )+(gds7 · gds9 )/(gm7 +gmb7 )
(5.23)
where Ao is the DC gain of the CMFB circuit, and p its most significant root (pole):
gm12 (Cin /2CT )
Ao ≈ (5.24)
gm15
gm15
p≈ (5.25)
CLF
Cin is the value of the input capacitances connected to Vout1 and Vout2 in the CMFB40 .
CLF is the load capacitance at the output of the CMFB. Its main contributors are M15
gate to source capacitance, the gate to source capacitances of the bias transistors in
the transconductor, M9 and M10, and the input capacitances Cin of devices M1–M4.
Assuming that the devices are matched (a01 = a02 = a0 ), adding both equations,
and taking into consideration that gm3 = gm2 and gm4 = gm1 , the differential equation
for the common mode at the output is obtained:
2a0
sCvoutCM = −(gm3 + gm1 )vinCM − voutCM − (Goutn + Goutp )voutCM
s+p
(5.26)
where
(gdsF3 + gdsF4 ) · gds6 (gdsF1 + gdsF2 ) · gds5
Goutn = + (5.27)
(gm6 + gmb6 ) (gm5 + gmb5 )
gds8 · gds10 gds7 · gds9
Goutp = + (5.28)
(gm8 + gmb8 ) (gm7 + gmb7 )
The denominator of the common mode transfer function has the form
D(s) = s2 C + s[Cp + (Goutn + Goutp )] + [2a0 + (Goutn + Goutp )p] (5.29)
The common mode transfer function has then two poles whose location in the real-
imaginary plane will depend on the relationship between the different parameters. If
the radicand in eq. (5.29) is positive then the system will have two negative poles,
being dominant the lowest frequency one. Otherwise, the poles will be complex con-
jugate and the transient response will depend on the values of the real and imaginary
40 The input capacitance ratio has been assumed to be C /C , although a different value could have
in T
been chosen (it does not need to be related to the transconductor input weight).
parts. This will happen whenever
8a0 C > [(Goutn + Goutp ) − Cp]2 (5.30)
In this case, the output will show some ringing until it settles which will happen faster
as the absolute value of the real term in eq. (5.31) increases:
−[Cp + (Goutn + Goutp )]
(5.31)
C
The previous two equations show how the higher the bandwidth of the CMFB
is the smaller settling time. Also, in order to minimise the ringing the gain of the
CMFB should not be too high. That is one of the reasons why a low gain amplifier is
chosen for this topology. The gain and bandwidth can still be high enough to correct
for common mode variations and simultaneously avoid falling into oscillations.
Summarising, the advantages of using FGMOS transistors in the CMFB are the
following:
1. Design simplicity: the common mode at the output can be sensed in a single
device. The risk of increasing the THD with the common mode feedback is
reduced thanks to the linear summation.
2. The input transistors can be in the strong inversion saturation region for common
mode output levels that would not make it possible in normal MOS devices. This
has several advantages such as an increased bandwidth and lower offset.
3. The dominant pole can be shifted towards higher frequencies by increasing the
bias current in the differential pair. This is possible even for very low values of
VDD by reducing the effective threshold voltages of the input transistors with an
extra input connected to VDD . In this way, it is possible to minimise the risk of
ringing as predicted by eq. (5.30).
5.2.2.1 Common Mode Rejection Ratio (CMRR)

The topology in Fig. 5.1 is completely symmetrical, and because of this ideally the
common mode gain at the output should be zero. However, in reality, non-ideal
effects such as mismatch will increase the gain [160]. Yet, due to the symmetry of the
topology the common mode gain will still be smaller than in a single-ended version.
A small signal equivalent circuit for the analysis of the common mode DC gain is
shown in Fig. 5.8. The common-mode DC gain is given by

vout2 Cin Cin

≈− gm3 + gm4
vinCM vind =0 CT 3 CT 4

(gdsF3 + gdsF4 + gm6 + gmb6 )
× 2Ao gm10
gm6 + gmb6

−1
CCM CCM
+ gm3 + gm4 (5.32)
CT 3 CT 4
gds10 –2gm10 AovoutCM
vS8
( g m8 + g mb8) v S8
gds8
voutCM
– (g m6 + g mb6) v D3
gds6
vD3
C C C C
in in CM CM
(gdsF3+gdsF4) ------- g + --------- g v +2 -------------- g + -------------- g A v
C m3 C m4 in CM C m3 C m4 o out CM
T 3 T 4 T 3 T 4
Figure 5.8 Small signal equivalent circuit for the calculation of the common mode
output gain
Equation (5.32) represents the gain of one of the single outputs. If both symmetri-
cal branches were identical the common mode gain resulting from subtracting two
identical equations from both single branches would be zero. In practice, the mis-
match between devices will give rise to two different equations. The expression for
the other branch can be obtained from eq. (5.32) just by changing the subscripts to
those corresponding to the devices in it. New subscripts have been added to account
for the mismatch between ideally identical FGMOS. They refer again to the name
of the device they correspond to. From now on, for the sake of simplicity the mis-
match between capacitances will be modelled in the form of a parameter in the
following way:

Cin Cin Cin Cin
= = (1 + 1i ) (5.33)
CT 1 CT CT i=[2,4] CT

CCM CCM CCM CCM
= = (1 + 2i ) (5.34)
The differential output gain when the input is a common mode signal, assuming that
gdsF3 + gdsF4 gm6 + gmb6 , gdsF1 + gdsF2 gm5 + gmb5 and only mismatch between
the capacitances is41

(vout1 − vout2 )
vinCM
vind =0

CCM
≈ [gm3 (13 − 12 ) + gm4 14 ] gm10 + (gm3 + gm4 )
CT

CCM
+ (gm3 + gm4 )[gm3 (22 − 23 ) − gm4 24 ]
CT

CT CCM
× 2Ao gm10 + (gm3 + gm4 )
Cin CT

−1
CCM
× gm10 + [(gm3 + gm4 ) + (23 + 22 )gm3 + 24 gm4 ]
CT
(5.35)
Taking into account that

Cin
Gm = (gm4 − gm3 ) (5.36)
CT
and assuming that the transconductor is designed in such a way that
gm10 ≈ (gm3 + gm4 ) (5.37)
eq. (5.35) can be rewritten as

(vout1 −vout2 )
vinCM
vind =0

CCM CCM
≈ [(13 − 12 )+k1 14 ] 1+ + [22 − 23 − k1 24 ]
CT CT

CT CCM CCM
× 2A 1+ (1+k1 ) 1+
Cin CT CT

−1
CCM
+ (23 +22 + k1 24 ) (5.38)
CT

Gm CT
k1 = 1+ (5.39)
gm3 Cin
41 This assumption is made for the sake of clarity to give the reader an idea of how much the input
capacitances in the FGMOS devices can degrade the performance in terms of CMRR. The derivation of a
more accurate expression for the CMRR from eq. (5.32) is however trivial.
This gain represents the effect of the common mode signals on the OTA differ-
ential output. Ideally it should be zero or at least negligible when compared with
the differential gain. The latter can be obtained with the same equivalent circuit if
the common mode contributions are cancelled and a differential input is considered
instead. The result is in eq. (5.40)42 . The Common Mode Rejection Ratio (CMRR)
is given by the ratio between the differential gain when the input signal is differential
(eq. (5.40)), and eq. (5.38):
vout Gm
≈ (5.40)
vind Goutn + Goutp

Gm CT CCM
CMRR ≈ − × Ao 1+
Goutn +Goutp Cin CT

CCM CCM
× (1 − k1 ) 1+ +(23 −22 +k1 24 )
CT CT
−1
CCM CCM
× [(13 −12 )+k1 14 ] 1+ + [22 −23 −k1 24 ]
CT CT
(5.41)
Equation (5.41) shows that:
1. The rejection of the common mode signals will increase as the CMFB gain
increases.
2. Increasing the ratio CCM /CT will also increase the CMRR.
3. The exact value of the CMRR depends on the value of the transconductance,
more specifically on the ratio between the latter and the individual transistors’
transconductances, which will in their turn be a function of the bias current
IB . Hence, for example, for extreme transconductance values such as Gm =
(−Cin/CT )gm3 , eq. (5.41) can be written as
Gm
CMRR ≈
Goutn + Goutp
Ao (CT /Cin )(1+CCM /CT )[(1+CCM /CT )+(22 +23 )(CCM /CT )]
×
(13 −12 )(1+CCM /CT )+(CCM /CT )(22 −23 )
(5.42)
In order to illustrate this result, let us assume a typical differential gain of
40 dB, Ao = 0.5, capacitances ratios of 0.2, and a maximum deviation of
42 This equation has not considered mismatch between transistors because the latter would not affect
the CMRR significantly.
1 per cent. The minimum value of the CMRR predicted by eq. (5.42) would
be 82 dB.
Let us now consider the case in which Gm ≈ gm3 instead. Equation (5.41) would
then be

Gm CT CCM CT CCM
CMRR ≈ × Ao 1+ − 1+
Goutn + Goutp Cin CT Cin CT

CCM CT CCM
+ (23 + 22 ) + +1 24
CT Cin CT

CT CCM
× (13 − 12 ) + + 1 14 1 +
Cin CT

−1
CCM CT
+ 22 − 23 − + 1 24 (5.43)
CT Cin
And the value of the CMRR for the same hypothetical specifications would be 53 dB.
Although the CMRR value appears to be much smaller, one of the assumptions is not
quite accurate for this particular case: the fact that in both situations the differential
gain would be identical. This is only true if the output resistance changes in the same
way as the transconductance, which does not happen because the bias current does
not follow the transconductance variation trend. Hence, in reality, the differential
gain will be higher for larger Gm .
5.2.3 Second-order effects

5.2.3.1 Total harmonic distortion
The functionality of the transconductor, described by eq. (5.3), was derived assum-
ing ideal square functions of VGS for the currents in the input transistors operating
in the strong inversion saturation region. This is an approximation though which
does not take into account second-order effects such as mobility reduction and chan-
nel length modulation that cause deviations from this ideal behaviour. A better
fitting function for the mobility would be given by eq. (4.55). The channel-
length modulation as a source of nonlinear distortion is negligible in this structure
because of its high output resistance. It will be thus not considered in the following
analysis.
Replacing eq. (4.55) for the mobility in each of the expressions for the currents in
Fig. 5.1 input transistors, and using Maclaurin series expansion, eq. (5.44) is obtained,
where VCM = VinCM + (CCM /Cin )Voutcm + (Cc /Cin )Vb2 , and VinCM is the common
mode at the effective input/output:

β1 Cin CT θ (VCM −(CT /Cin )VTFG )+2(CT /Cin )
Iout ≈ VCM − VTFG × Vind
2 CT Cin [θ (VCM −(CT /Cin )VTFG )+CT /Cin ]2

β1 Cin CT Cc
− VCM − VTFG + (Vb1 −Vb2 )
2 CT Cin Cin
θ (VCM −(CT /Cin )VTFG +(Cc /Cin )(Vb1 −Vb2 )) + 2(CT /Cin )
× Vind
[θ (VCM −(CT /Cin )VTFG +(Cc /Cin )(Vb1 −Vb2 ))+CT /Cin ]2

(θβ1 /8)(CT /Cin )
+

(θβ1 /8)(CT /Cin )

− V3 (5.44)
[θ (VCM −(CT /Cin )VTFG )+CT /Cin ]4 ind
The THD is hence given by [171]43

θ (CT /Cin )2
THD =
16 [θ (VCM −(CT /Cin )VTFG +(Cc /Cin )(Vb1 − Vb2 ))+CT /Cin ]4

(CT /Cin )2 2 CT
− A VCM − VTFG
[θ (VCM −(CT /Cin )VTFG )+CT /Cin ]4 Cin

θ (VCM −(CT /Cin)VTFG)+2(CT /Cin) CT Cc
× − V CM − V TFG + (V b1 −V b2 )
[θ (VCM −(CT /Cin)VTFG)+CT /Cin]2 Cin Cin

θ (VCM −(CT /Cin )VTFG +(Cc /Cin )(Vb1 −Vb2 ))+2(CT /Cin ) −1
× (5.45)
where A is the amplitude of a sinusoidal input signal. Theoretical values of THD
for different values of common mode input voltages and (Vb1 − Vb2 ) are shown in
Fig. 5.9. The values of the parameters used to obtained these graphs are the same as
for the design example in Section 5.2.5.1 and θ is 0.234 V−1 44 . Curves of simulated
THD for this design are given in Fig. 5.15. They show how the THD remains low
in a wide range of Gm 45 . Therefore, as long as the noise floor is also low in the
same range, the circuit will have a good performance in terms of DR in the whole
programmability range.
The main conclusion that can be obtained from the theoretical expressions is that
having FGMOS transistors at the input reduces distortion because of the compression
of the effective input voltage at the gate by a factor of Cin /CT .
43 Ignoring mismatch between the two symmetrical branches.

44 This value can be derived from parameters of the BSIM49 model used in the design example in
section 5.2.5.1.
45 This is equivalent to a wide range of the difference (V − V ).
b1 b2
1.4 1.4
1.2 1.2
(Vb1−V b2) = 4 mV (Vb1−Vb2 ) = 0.4 V
1.0 1.0
THD (%)
0.8 0.8
0.6 0.6
V⬘CM,min = 0. 45 V V⬘CM,min = 0. 45 V
0.4 0.4
0.2 V⬘CM,max = 1.25 V 0.2 V⬘CM,max = 1.25 V
0 0
0.5 1.0 1.5 2.0 0.45 0.95 1.45 1.95
A(Vpp) A(Vpp )
Figure 5.9 Theoretical Total Harmonic Distortion for two different (Vb1 − Vb2 )
values
Table 5.1 Equivalent input noise contribution for every transistor in one of the
branches in Fig. 5.1
vin /Si M3 = M4 Si = S 2i

[gdsF3 + gdsF4 ](CT /Cin ) ρn gm6

M6 4KTgm6 + γn
(gds6 + gdsF3 + gdsF4 + gm6 + gmb6 )(gm3 − gm4 ) W6 L6 f

gds10 (CT /Cin ) ρp gm8

M8 4KTgm8 + γp
(gds8 + gds10 + gm8 + gmb8 )(gm3 − gm4 ) W8 L8 f

(CT /Cin ) ρn gm(3,4)

Mi (i = [3, 4]) 4KTgm(3,4) + γn
(gm3 − gm4 ) W3 L3 f

(CT /Cin ) ρp gm10

M10 4KTgm10 + γp
(gm3 − gm4 ) W10 L10 f
5.2.3.2 Noise
Noise of the low voltage transconductor
Approximate expressions for the equivalent noise contribution of each transistor at
the effective input are collected in Table 5.1. Similar equations can be obtained for
the symmetrical circuit branch just by changing the subscripts. This model does
not take into account parasitic capacitance frequency effects, and therefore will not
be accurate for high frequency calculations. However, it is good enough to give
an idea of the effects caused by the FGMOS transistors. The table does not show
the contribution of the CMFB. The latter will be analysed separately since there
is no need of having a CMFB for every transconductor, but only for every circuit
output, and the noise would be overestimated if it was considered as part of the
main cell.
From Table 5.1, if the circuit is designed in such a way that gdsF3 + gdsF4
gm6 + gmb6 then the main noise contributors will be transistors M3, M4 and M10.
As both branches are designed to be identical parameters with subscripts 1 and 9
will have the same value as parameters with subscripts 3 and 10, respectively. Also,
M3 = M4, which means W3 L3 = W4 L4 . Under these assumptions the equivalent
power spectral density of noise at the input is46
" " 2 +g 2
# #
2
vin CT 2 ρn gm3 m4 gm3 +gm4
= 8KT + γn
f Cin f (gm3 −gm4 )2 W3 L3 (gm3 −gm4 )2
2 " " # #
CT ρp 2
gm10 gm10
+ 8KT + γp
Cin f (gm3 −gm4 )2 W10 L10 (gm3 −gm4 )2
(5.46)
This expression can be simplified using the parameter k1 previously defined:

2
vin CT 2 1 (1 + k1 )2 ρn ρp
= 8KT +
f Cin f (1 − k1 )2 W3 L3 W10 L10

2 (k1 ) ρn (1 + k1 ) (γn + γp )
− + (5.47)
f (1 − k1 )2 W3 L3 (1 − k1 )2 gm3
Let us now consider as an example the limit condition, Gm = (−Cin /CT )gm3 .
Equation (5.47) can then be rewritten as

2
vin CT 2 1 ρn ρp γ n + γp
= 8KT + + (5.48)
f Cin f W3 L3 W10 L10 gm3
If the same derivation was carried out for normal MOS devices (Cin → CT ) the
final expression would be the same, except for the, larger than 1, multiplicative factor
(CT /Cin )2 . In simple terms, this means that the power spectral density of noise at the
input increases. But, going back to the linearity analysis performed in the previous
section, the power spectral density of the signal for a certain THD requirement also
increases. So, the final dynamic range will not degrade. In any case, having the
same topology with normal MOS transistors would not be possible since multiple
inputs at the gate are needed in order to implement the right functionality. A possible
modification would be to have the bias signals connected at the sources of the MOS
devices [164], but this would make the final structure much more complex since
46 In the case where the FGMOS devices’ gate to drain parasitic capacitances are not small enough for
the assumption gdsF3 + gdsF4 gm6 + gmb6 to be valid, there would be an extra term in eq. (5.46), that
would also have to be considered.
extra circuitry would be needed to generate those voltages. Also, the latter would
be connected to low impedance nodes and this would require a considerable amount
of extra power for the biasing circuits to be able to cope with the low impedance
loads.
The power spectral density of noise at the output can be obtained from eq. (5.46)
just by multiplying the input noise by the square of the gain. In this case, the scaling
factor given by the effective input weight would disappear and the effect of the
FGMOS would be as a reduction of the noise due to the lower output resistance and
gain because of the parasitic capacitances.
The general conclusions that can be drawn from eq. (5.46) are:
1. In order to reduce the flicker noise, transistors M1, M2, M3, M4, M9 and
M10 should be designed as big as possible. The problem with this is that as the
area of the transistors increases the frequency response is poorer. The same happens
to the distortion and the DC gain of the transconductor. This could be sorted out
by increasing the input capacitances accordingly, but in this case, the area would
increase, and also the loading effects when the transconductor is used as part of a
bigger system, which would degrade its frequency response.
2. The main advantage of using FGMOS is the small number of devices that are
required to linearise the structure which, in terms of noise, is equivalent to having
less noise contributors.
Common mode feedback circuit noise

As it was mentioned before the noise analysis for the CMFB deserves a separate
section since only one of these blocks is required for every couple of output nodes.
Table 5.2 Equivalent input noise

contribution for every
transistor in the CMFB
vin /S i

CT 1
M12
Cin gm12

∼ CT 1
M13 =
Cin gm12

CT 1
M14
Cin gm12

∼ CT 1
M15 =
Cin gm12
Table 5.2 summarises the approximate equivalent noise at the input of the CMFB for
every transistor. The subscripts refer to Fig. 5.5. The input capacitance ratio has been
assumed to be Cin /CT , although a different value could have been chosen. (It does
not need to be related to the transconductor input weight.)
And the equivalent power spectral noise density at the input for the CMFB is

vout1
2
CT 2
≈ 4KT
f Cin
" #
1 2ρn gm14 2 ρp γn gm14 γp
× +2 +2 + 2
f W12 L12 gm12 W14 L14 gm12 gm12
(5.49)
Deviations in transconductances and sizes between similar transistors have been

neglected.
The analysis of this block draws similar conclusions to those obtained for the
transconductor circuit. Again, the flicker noise can be reduced by increasing the
transistors areas, whereas the thermal noise is inversely proportional to the transcon-
ductance which on the other hand has to have a minimum value to keep an adequate
bandwidth.
The noise in eq. (5.50) will not have a significant effect in the performance of the
whole transconductor block. The reason for this is that it is a common mode signal
and thus its effect will be almost the same for both symmetrical branches. Its exact
contribution will depend on the common mode rejection ratio but, as a first approach,
and for the values of CMRR previously obtained it can be considered negligible.
Ideally, in absence of mismatch, it would be zero:

voutcm
2 1 2ρn gm12 2 ρp
≈ KT +2
f f W12 L12 gm14 W14 L14

gm12 γp
+ 2 γn 2 + (5.50)
gm14 gm14
5.2.4 Power supply rejection ratio

5.2.4.1 Rejection of VDD noise
The VDD noise in the transconductor in Fig. 5.1 is coupled to the output through two
different paths: on one hand, M10 and M9 sources and bulks as well as M8 and M7
bulks are connected to VDD ; on the other hand, also one of the inputs in transistors, M1,
M2, M3, M4, is connected to VDD aiming to reduce the effective threshold voltage.
Figure 5.10 shows a low/mid frequencies equivalent circuit for PSRR+ calculations.
v DD
gds10 v outcm
g m10 1 – ----------------
- v DD
v DD v
outCM = 0
vs8
gds8 ( g m8 + g mb8) v s8 – g mb8v DD
vout2
gds6 −( gm6+gmb6 )vD3
vD3
C
C
VDD CM v outcm
g
m3
----------------- + -------------- ----------------
- v DD
C C
(gdsF3 +gdsF4) T 3 T 3 v DD v
outCM = 0
C
C
VDD CM v outcm
+g
m4
----------------- + -------------- ----------------
- v DD
C C
T 4 T 4 v DD v
outCM = 0
Figure 5.10 Small signal equivalent circuit for the PSRR+
The voltage gain at the output is given by eq. (5.51)47 :

vout
v
DD vind =0
[gm3 (CCM /CT )3 + gm4 (CCM /CT )4 − gm1 (CCM /CT )1 − gm2 (CCM /CT )2 ]
≈−
Goutn + Goutp

voutcm
×
vDD voutCM =0
[gm3 (CVDD /CT )3 +gm4 (CVDD /CT )4 −gm1 (CVDD /CT )1 −gm2 (CVDD /CT )2 ]
−
Goutn +Goutp
(gm10 − gm9 )[1 − voutcm /vDD |voutCM =0 ]
+ (5.51)
Goutn + Goutp
A first conclusion that can be extracted from the last term in eq. (5.51) is that
if the gain (voutcm /vDD )|voutCM =0 is positive (as is the case) then the common mode
feedback at the gate of the top p-channel transistors is going to contribute to reduce
this term and therefore to increase the PSRR+. On the contrary, the feedback at the
bottom n-channel FGMOS transistors will have the opposite effect. For the sake of
47 The term corresponding to M8 bulk will be much smaller.

v DD
g m14 (vDD – v)D12 gds14 gds15 g m15 (vDD – voutcm )
vD12 voutcm
C C
VDD VDD
g
m12
-----------------
C
vDD gdsF12 gdsF13 g
m13
-----------------
C
vDD
T 12 T 13
C
VDD 2 Cin C
VDD Cgnd
– ----------------- + -------------- g v – ----------------- + -------------- g v
C C m12 s C C m13 s
T 12 T vs T 13 T
gds11
Figure 5.11 Small signal equivalent circuit for the calculation of the PSRR+ in
the CMFB
simplicity and to help to interpret the results, let us define again a parameter 48
to account for the mismatch and let us consider as a first approach only mismatch
between the capacitances:

CVDD CVDD CVDD CVDD
= = (1 + 3i ) (5.52)
Then, the final expression for eq. (5.51) is49

vout CCM voutcm
≈ − gm3 [23 − 22 + k1 24 ]
vDD vind =0 CT vDD voutCM =0

CVDD
− gm3 [33 − 32 + k1 34 ] (Goutn + Goutp )−1
CT
(5.53)
The contribution of the CMFB (Fig. 5.11) under the assumptions of no body effect in
the input transistors, only mismatch between the input capacitances (gm13 = gm12 ),
a much smaller parasitic capacitance between the FG and source than the total
48 As in the CMRR section.

49 If a similar percentage of mismatch is considered for the transconductances as well, the final result
for eq. (5.53) would approximately double.
capacitance seen by the FG and gdsF gm , is

voutcm CVDD CVDD 2Cin
≈1− +
vDD voutCM =0 CT 13 CT 12 CT

CVDD CVDD Cgnd
− + gm13
CT 12 CT 13 CT

CVDD 2Cin CVDD Cgnd −1
× gm15 + + +
CT 12 CT CT 13 CT
(5.54)
If the power supplies have been used to generate the reference voltage, the second
term in eq. (5.54) will always be positive which reduces the gain. Otherwise it will
be zero, unless there is mismatch. The final PSRR+ for the transconductor block is
given by

CCM voutcm
PSRR+ ≈ (k1 − 1) − [23 − 22 − k1 24 ]
Cin vDD voutCM =0
−1
CVDD
− [33 − 32 − k1 34 ] (5.55)
Cin
Equation (5.55) shows how the PSRR+ improves as the contribution of the input
capacitance to the total capacitance increases and the values of the capacitances
connected to the CMFB and VDD decrease.
Let us now analyse as an example the limit case k1 → 0, or what is the same
Gm = (−Cin /CT )gm3 . The equation for the PSRR+ in this case is
1
PSRR+ ≈
(CCM /Cin )[23 −22 ](voutcm /vDD |voutCM =0 )+(CVDD /Cin )[33 −32 ]
(5.56)
Hence, for example, in the hypothetical case of a design with a maximum of 1 per cent
mismatch between capacitances and CCM = Cin = 0.2CT , CVDD = 0.4CT , the worst
PSRR+ predicted by eq. (5.56) would be 25 dB.
Summarising the main sources of degradation of the PSRR+ are: (a) The VDD
connection at the source of the load transistors in the CMFB, which is coupled to
the main circuit block through the capacitance CCM in the input FGMOS devices.
(b) The connection to VDD used to bias the FGMOS input devices in the right operating
region. Hence, even when by doing this the topology is greatly simplified and the
power reduced the price to pay in terms of PSRR+ should be kept in mind when
establishing the design trade-offs.
5.2.4.2 Rejection of VSS noise

The rejection of the negative power supply noise can be estimated with a similar
circuit to the one in Fig. 5.10 but with VSS instead of VDD as the input, and also
a different input coming from the CMFB (voutcm /vSS |voutCM=0 ):

Cin
PSRR− ≈ (gm4 − gm3 )
CT
4
CCM voutcm
× 1− gmi
CT i vSS voutCM =0
i=3
2
−1
CCM voutcm
− 1− gmi (5.57)
CT i vSS voutCM =0
i=1
Considering now, as before, only mismatch between capacitances:

(k1 − 1)
PSRR− ≈ − (5.58)
(CCM /Cin )[23 − 22 + k1 24 ]((voutcm /vSS )|voutCM =0 )
where the value of (voutcm /vSS )|voutCM =0 , is the same as for a normal differential pair
with MOS transistors, unless that the negative power supply (VSS ) is also used to
generate the reference in the CMFB in which case it would be

voutcm Cgnd gm12
≈− (5.59)
vSS voutCM =0 2CT gm15
Hence, for example, for the same limit condition as before, Gm = (−Cin /CT )gm3 ,
and the same hypothetical assumptions together with Cgnd = 0.25 CT , the PSRR−
for a 1 per cent maximum mismatch would be 52 dB. Equation (5.58) predicts a
degradation of the PSRR− as the absolute value of the transconductance decreases
(due to the lower differential gain).
Summarising, using n-channel FGMOS devices is specially critical for the
PSRR+, if the positive power supply voltage is used to bias them in the strong inver-
sion saturation region. However, if p-channel input devices biased by the negative
power supply voltage are employed instead, the PSRR+ would not be so critical but
the PSRR− would be worse.
5.2.5 A filter example

This section illustrates the design of a lowpass/bandpass biquad filter that uses the
integrator in Fig. 5.1 as the main building block. The filter is described in the time
domain by the following state-space equations:
ẋ1 = −ωo2 x1 − ωo1 x2 + ωo4 u (5.60)

ẋ2 = ωo3 x1 (5.61)
where x1 , x2 and u are the bandpass output, lowpass output and input, respectively.
Four coefficients, ωo1 , ωo2 , ωo3 and ωo4 , define the filter programmability. The cut-
off frequency ωo , quality factor Q, lowpass DC gain HLP (0) and bandpass gain at the
Vinp V1p
C
Gm4 Gm2
Vinn V1n
Gm3
V2p
C
Gm1
V2n
Figure 5.12 Block diagram for a second order Gm –C filter circuit
cut-off frequency, HBP (ωo ), depend on these coefficients as follows:

√
ωo = ωo3 ωo1 (5.62)
√
ωo1 ωo3
Q= (5.63)
ωo2
ωo4
HLP (0) = (5.64)
ωo1
ωo4
HBP (ωo ) = (5.65)
ωo2
A differential form of state space equations (5.60) and (5.61) is given by
ẋ1p − ẋ1n = −ωo2 (x1p − x1n ) − ωo1 (x2p − x2n ) + ωo4 (up − un ) (5.66)
ẋ2p − ẋ2n = ωo3 (x1p − x1n ) (5.67)
The circuit that implements these equations is shown in Fig. 5.12, where
x1p − x1n V1p − V1n VBP (s) s(Gm4 /C)

= = = 2 (5.68)
up − un Vinp − Vinn Vind (s) s + sGm2 /C + (Gm3 Gm1 )/C 2
x2p − x2n V2p − V2n VLP (s) (Gm3 Gm4 )/C 2

= = = 2 (5.69)
up − un Vinp − Vinn Vind (s) s + sGm2 /C + (Gm3 Gm1 )/C 2
And the filter parameters are

.
Gm3 Gm1
ωo = (5.70)
C2
√
Gm3 Gm1
Q= (5.71)
Gm2
Gm4
HLP (0) = (5.72)
Gm1
Gm4
HBP (ωo ) = (5.73)
Gm2
5.2.5.1 Design performance

The integrator and filter are designed in a 0.8 μm double poly CMOS technology with
threshold voltages in the order of 0.8 V [214]. The design is intended for audio appli-
cations and again the main goals are to reduce the values of power consumption and
supply voltage. The time constants (Gm /2πC) required for a frequency range between
100 Hz to 10 kHz are in the interval [100 μs, 10 ms]. As the time constant is inversely
proportional to the integrating capacitance and proportional to the transconductance,
its lowest required value will determine either the minimum transconductance or the
maximum capacitance values that will be needed in the design. Sometimes, it is very
difficult to realise very low Gm s values which forces the designer to choose a high
capacitance value instead, hence compromising the area. If the area is a constraint,
the design will have to be programmable to achieve the low m values, which is much
more challenging. For our design, a relatively low capacitance value of 5 pF is chosen.
This means the Gm needs to be programmable in the range [3 ns, 0.3 μs].
300 n
0.3 V
V b2−Vb1 = 0.4 V
2 nA/div
0.2 V
I out
–300 n
–1 V Vind 1V
Figure 5.13 Gm -tuning for (Vb2 − Vb1 ) ∼ 0.2, 0.3 and 0.4 V
Transistors M1–M4 are designed with an aspect ratio W /L = 2.5 μm/10 μm and
input capacitances of 150 fF and 250 fF (for CVDD ). M12 and M13 aspect ratio is
2.5 μm/17 μm and their input and total capacitances are 133 fF and 533 fF, respec-
tively. Figure 5.13 and Fig. 5.14 show simulation results for the output current (Iout )
versus the differential input voltage for several values of the programming voltage
(Vb2 − Vb1 ). As predicted by eq. (5.3) the transconductor is very linear in the whole
operating range. The exact characterisation in terms of linearity is given by the THD
test. The results of this test for Gm1 = 3 ns and Gm2 = 300 ns are shown in Fig. 5.15(a)
and (b), respectively. Both have been obtained at the edges of the input common mode
20 n
V b2−Vb1 = 24 mV
14 mV
2 nA/div
4 mV
I out
–20 n
–1 V 1V
Vind
Figure 5.14 Gm -tuning for (Vb2 − Vb1 ) ∼ 4, 14 and 24 mV
(a) 1.0 (b) 1.0
0.8 (V −V ) = 4 mV 0.8 (V −V ) = 4 mV
b1 b2 b1 b2
0.6 0.6
THD(%)
V⬘CM,min = 0.45 V V⬘CM,min = 0.45 V

0.4 0.4
0.2 0.2
V⬘ =1 .25 V V⬘ =1 .25 V
CM,max CM,max
0 0
0 0.5 1.0 1.5 2.0 0 0.5 1.0 1.5 2.0
A (Vpp) A (Vpp)
Figure 5.15 THD versus peak-to-peak input amplitude: (a) for Gm = 3 ns, and
(b) for Gm = 300 ns
range (the difference between both common mode voltages is 0.8 V). The total THD
is always below 1 per cent, and again the results are in agreement with the theoretical
ones in eq. (5.45). The lossless integrator transfer functions for two transconductance
values is shown in Fig. 5.16. In both cases, the integrator losses are below 20 Hz, far
enough from the audio frequency range. Table 5.3 summarises the performance of
the integrator for these two transconductance values and Fig. 5.17 illustrates different
bandpass and lowpass filter transfer functions within the programmability range.
60
50 Gm=300 ns
40
30
20
Gm = 3 ns
(5 dB/div)
10
Iout
0
C = 5 pF
–10
–20
–30
–40
–50
1 10 100 1k 10 k 100 k
Fr equency (Hz)
Figure 5.16 Gm -C integrator transfer-function [120]
Table 5.3 Summary of Hspice simulations
Gm1 = 3 ns Gm2 = 300 ns
VDD 2V
Technology AMS0.8 μm (CXQ)
Vind 2 Vpp 2 Vpp
CMR 0.8 V 0.8 V
CMRR (1% mismatch) >60 dB
THD @ 2 Vpp <1%
Inband noise 210 μV
Power 10 μW
(a) 20
10
Volts (dB)
0
V LP (f)
–10
–20
–30
–40
20
10
0
Volts (dB)
VBP (f)
–10
–20
–30
–40
100 1k 10 k 100 k
Frequency (log) (Hz)
(b) 0
–20
Volts (dB)
VLP (f)
–40
–60
–80
0
–10
Volts (dB)
VBP (f)
–20
–30
–40
100 1k 10 k 100 k
Frequency (log) (Hz)
Figure 5.17 Programming of the filter. (a) Q-programming. (b) f0 -programming
5.3 An intermediate frequency FGMOS-based filter
Monolithic filters in the frequency range of 100 kHz to 10 MHz have many
applications in communication receivers and video processing as well as in data
communications and local area networks [147,150–155,157–159,166]. In general,
design techniques that are applicable to voice band applications do not perform prop-
erly in this frequency range. They become even more difficult to extrapolate as the
voltage supply limits are pushed down, because the maximum values of the currents
that devices can provide is more restricted. The consequence of this is the degradation
of the high frequency range.
This section covers the design of an FGMOS filter operating in the range of mega-
hertz at 1.25 V power supply voltage. As it was explained in previous designs along
this book, the FGMOS can be used as an alternative to MOS with a reduced effective
threshold voltage. Hence, higher current levels will be possible which increases the
upper reachable frequency limit. Besides, the transistor facilitates the common mode
feedback mechanism without introducing extra distortion.
5.3.1 The FGMOS-based IF transconductor

Let us consider the circuit in Fig. 5.18 with two equal FGMOS transistors (M1a , M1b )
operating in the strong inversion saturation region [215]. The drain to source currents
of transistors M1a and M1b can be expressed as50
⎛ ⎞2
β1(a.b) ⎝
I1(a,b) = wi Vi(a,b) − VS1 − VTFG ⎠ (5.74)
2
i=1, ... , N
where subscripts (a,b) apply to transistors M1a and M1b , respectively, wi = Ci /CT ,
and VTFG is the effective threshold voltage:
CCM CVDD
VTFG = VTn − Voutcm − VDD (5.75)
CT CT
Assuming that the input transistors are identical and the input weights have the
same value, w, the output current delivered to a load connected between the drains
of M2a and M2b is
⎛ ⎞
√
I1b − I1a w β1 ISS ⎝
Iout = ≈ (Vib − Via )⎠ (5.76)
2 2
i=1, ... , N
where β1 = β1a = β1b . A lossless integrator can be implemented just by joining an

integrating capacitor between the output nodes (M2a and M2b drains).
This approximation has the same limitations as a conventional differential
pair
[160]. The difference is that in this case the voltage that the FGs see is
w· i=1 ... ,N Vi1(a,b) as opposed to i=1 ... ,N Vi(a,b) . This means that for a certain value
of distortion the input range increases by a factor 1/w. Equation (5.76) shows a linear
relationship between voltages and current. The gain factor is the transconductance:
√
w β1 ISS
Gm = (5.77)
2
50 This equation is valid whenever the parasitic capacitances are negligible compared with the FGMOS
inputs.
VDD VDD
Voutcm M4 a M4 b Voutcm
VS3a
gnd gnd (Cgndf )
Voutcm M3 a M3 b Voutcm (CCMp )
VD2a
VDD Vout VDD (CVDDf )
M2 a M2 b
Voutcm Voutcm (CCMn)
VD1a I1a I1b
Via for i = [1,N] Vi b for i = [1,N] (C )
i
Voutcm M1 a M1 b Voutcm (CCM)
VDD VDD (CVDD)
ISS VFG1a VFG1b
ISS + Vi1a for i = [1,N]
VS1 Gm Vout
Vi1b for i = [1,N]
M5 a M5b
-
(a) (b)
Figure 5.18 (a) Proposed FGMOS transconductor. (b) Symbol
5.3.2 The FGMOS cascode load and the CMFB

As Fig. 5.18 shows, the transconductor load consists of a cascode structure. The
reasons that justify this, in spite of the fact that stacking transistors increases the
value of the minimum voltage needed for the circuit to work, are:
1. The output resistance is improved and hence the losses reduced, which is
even more important now than when using normal MOS devices, since the FGMOS
conductance is higher due to the added (CGD /CT )gm term (see Chapter 2).
2. The cascode bias voltages can be generated within the same devices just by
choosing the appropriate weights. The latter is equivalent to shifting the effective
threshold voltages appropriately so that the transistors can be biased using already
available voltages within the system. This avoids having to have extra circuitry for
this aim (and hence, extra power) or very large transistors with also large parasitic
capacitances that would affect the frequency response.
3. As it was already explained before, because the topology is differential a
CMFB is required [216]. The cascode devices can help to improve the efficiency of
the feedback mechanism in the following way: if Voutcm is the output voltage in the
CMFB and this is fed back to the cascode transistors through the input capacitances
CCMn (in the n-channel cascode device) and CCMp (in the p-channel cascode device)
as well as to the input and load transistors (M1(a,b) and M4(a,b) , respectively) the
output gain with respect to the output of the CMFB is51

vD2 CCM CCMp
ACM = ≈ − gm4 + gm1 + gds4
voutcm CT CT

(CCMn /CT ) · gm2 gdsF1

+ Rout
[(CCMn /CT + CVDDf /CT )gm2 + gmbF2 + gdsF2 + gdsF1 ]
(5.78)

gdsF3 gds4
Rout ≈
[(CCMp /CT + Cgndf /CT )gm3 + gdsF3 ]

−1
gdsF2 gdsF1
+ (5.79)
[(CCMn /CT + CVDDf /CT )gm2 + gmbF2 + gdsF2 + gdsF1 ]
Hence, the third and fourth term in eq. (5.78) which are due to the feedback in the
cascode transistor increase slightly the efficiency of the feedback mechanism.
Also, having feedback in the cascode devices as well minimises the voltage vari-
ations in the drains of transistors M1(a,b) and M4(a,b) , which is important for the
following reasons: (a) in order for this block to operate properly all the devices have
to be biased in the strong inversion saturation region. However, changes in the values
of VD1(a,b) or VS3(a,b) may bring the devices out of this region which would degrade
the performance of the whole block. (b) Also, from the point of view of VD1(a,b) ,
keeping this voltage constant can help to minimise the distortion caused by the term
corresponding to the gate to drain coupling in the transistor current law52 .
Analytically, the variations that VD1(a,b) and VS3(a,b) experience when the common
mode feedback voltage changes are
!
∂VS3(a,b) CCMp β4
= − (5.80)
∂Voutcm CT β3
!
∂VD1(a,b) CCMn CCM β1
= − (5.81)
∂Voutcm CT CT β2
where the first positive term in the equations (which compensates for the second
negative term) is due to the aforementioned feedback in the cascode transistors.
The price to be paid for the previous advantages is related to the degradation of
the PSRR because of the use of the power supply voltages to bias the cascode devices.
This will be analysed in more detail in a different section.
The CMFB suggested for this block is the same as in the previous design. However,
for the design example at the end of this section p-channel MOS transistors were used
51 V
D2a stands for the single side output as appears in Fig. 5.18. Also, the subscripts (a,b) have been
eliminated for now in the parameters names for the sake of simplicity as ideally parameters of the same
type that only differ in the a, b subscripts are supposed to be identical. As usual subscript i alludes to
device Mi.
52 This term was neglected in eq. (5.74) for the sake of simplicity.
VDD
VBIAS M7
gnd gnd (Cgnd)

(Cin) V D2a M6a M6b VDD (CVDD6)
V D2b
(Cin)
Voutcm
M5 a M5 b
Figure 5.19 The CMFB
at the input53 . The schematic with the values of the corresponding input capacitances
is shown in Fig. 5.19.
5.3.2.1 The Common Mode Rejection Ratio (CMRR)

As it was previously explained for other transconductors blocks, the fully differential
characteristic of the topology should reject all common mode signals. In reality mis-
match will make the common mode gain different from zero though. In the case of
this particular transconductor if only one effective input (Vin(a,b) ) is assumed54 , the
DC common mode gain under the assumption that matching is not perfect is

vD2b − vD2a vout
=
vinCM vind=0 vinCM vind=0
−gm1b (Cin /CT )b Routb

≈
1 + |ACMb |(2(Cin /CT )gm6 )/(gdsF6 + gds5 )
gm1a (Cin /CT )a Routa
+ (5.82)
1 + |ACMa |(2(Cin /CT )gm6 )/(gdsF6 + gds5 )
Subscripts a and b have been added to distinguish between the now different param-
eters in both branches (because of mismatch). It has also been assumed that the
effective input weights in the CMFB are the same as in the main transconductor in
absence of mismatch (Cin /CT ). Also, mismatch in the CMFB has been neglected,
since it will not change significantly the final results but would make the equations
however much less intuitive.
53 It was more convenient in this case because of the characteristics of the technology used. It was
easier this way to obtain adequate voltage levels.
54 For the sake of clarity in the expressions. If the transconductor has more than one input the respective
contributions add up. The equivalent input weight is Cin /CT .
The DC differential gain can be expressed as

vD2b − vD2a vout Cin
Ad = = ≈− · gm1 · Rout (5.83)
(vinb − vina ) vind vinCM=0 CT
Equation (5.83) does not take into account mismatch. This approximation has been
made for the sake of clarity. Also, it is justified because the main effect of mismatch
in the CMRR is going to be through the common mode gain. The CMRR is given by
the ratio between eq. (5.83) and eq. (5.82):

Cin (Cin /CT )b
CMRR ≈
CT 1 + |ACMb |[2(Cin /CT )gm6 ]/[gdsF6 + gds5 ]
−1
(Cin /CT )a
− (5.84)
1 + |ACMa |[2(Cin /CT )gm6 ]/[gdsF6 + g ds5 ]
The value predicted by eq. (5.84) can be as high as 80 dB if the relationship
between the transconductances and conductances is larger than 10, the input weight
is 0.5 and the mismatch is below 10 per cent.
5.3.3 Dynamic range

The transconductor in Fig. 5.18 can be seen as a modified version of a conventional
differential pair. From the point of view of linearity the behaviour referred to the
FGs (FG1a and FG1b ) is pretty much the same as in the normal differential pair.
This is because of the cascode devices that keep M1a and M1b drain voltages almost
constant, and hence reduce the nonlinear effect of the term (CGD VD1(a,b) )/CT in the
current equations. Therefore, the input range referred to the FGs would be the same.
However, referred to the effective input, the input range for a certain level of distortion
is amplified by the inverse of the effective input weight, this is by CT /Cin .
Regarding the noise the total RMS spectral density of noise referred to one of the
inputs is given by55
!
CT g2
VinRMS ≈ S1 + S4 m42
(5.85)
Cin gm1
where S1 and S4 are the power spectral density of noise for transistors M1(a,b) and
M4(a,b) , which are the main noise contributors.
Hence, taking into account that the equivalent input range for a certain level of
distortion increases by CT /Cin but also the equivalent input noise does, the final
dynamic range would remain the same as in a MOS realisation. The advantage is
that it is possible to keep the dynamic range and simultaneously have a large input
range, which can be very important for certain applications specially under the design
constraints of low power and low voltage56 .
55 Other noise contributions are negligible.

56 There is, for example, no need of designing extra circuitry to attenuate incoming signals, which
would consume extra power.
5.3.4 Rejection to supply noise

5.3.4.1 Rejection to VDD noise (PSRR+)
In the transconductor in Fig. 5.18, the VDD noise is coupled to the output through
transistors M1(a,b) and M2(a,b) inputs, M4(a,b) source, as well as from the feedback
from the CMFB. Ideally, as it was already said before, considering that this structure is
fully differential, both branches should be equally affected, and hence the difference
between the two of them at the output should be zero. However, mismatch increases
this difference, which is then given by

vout CVDD CVDD
≈ − g R + g Routa
vDD vind=0
m1b outb m1a
CT b CT a

voutcm
+ gm4b Routb − gm4a Routa +(ACMb − ACMa )
vDD voutCM =0
(CVDDf /CT )b · gm2b gdsF1b · Routa
−
[((CCMn /CT )b + (CVDDf /CT )b )gm2b + gmbF2b + gdsF2b + gdsF1b ]
(CVDDf /CT )a · gm2a gdsF1a · Routa
+
[((CCMn /CT )a + (CVDDf /CT )a )gm2a + gmbF2a + gdsF2a + gdsF1a ]
(5.86)
where (voutcm /vDD )|voutCM=0 is the output to vDD gain for the CMFB:

voutcm −gm6 CVDD6
≈ (5.87)
vDD voutCM=0 gdsF6 + gds5 CT
In order to get a more clear understanding of the FGMOS devices contribution to
the PSRR+, let us again assume that the main deviation from the ideal values occurs
in the input capacitances. Using eq. (5.83) for the differential gain, the PSRR+ is
given by
PSRR+

−1
Cin CVDD CVDD CCM CCM voutcm
≈ − + −
CT CT b CT a CT b CT a vDD voutCM =0
(5.88)
Equation (5.88) shows that the most significant contributions of the FGMOS
devices to the PSRR+ come from the input transistors in both the main transcon-
ductor block and the CMFB. In particular, if the gain of the CMFB is too high its
corresponding PSRR+ term will become dominant and the value in eq. (5.88) will be
seriously degraded. Again, a possible way to avoid this if it becomes a problem would
be to generate the reference voltage in the CMFB separately, instead of directly using
VDD for it. As an example of the orders of magnitude for the PSRR+ in this structure,
if the worst case of mismatch is 1 per cent, the CMFB gain is 10, CVDD = Cin ,
CVDD6 = 0.5CT and CCM = 0.5Cin the PSRR would just be 29 dB. However, if VDD
is not used to generate the reference voltage in the CMFB, the PSRR+ would increase
to 40 dB. Besides, these calculations have been carried out neglecting the mismatch
between transconductances, which is not strictly right. Hence, for example, variations
with respect to their ideal values of M4(a,b) transconductances, would generate a term
of around the same value as the denominator in eq. (5.88). This would degrade the
PSRR+ even more. The conclusion to this analysis is that the designer should be very
cautious when designing a circuit like this if VDD is used within a high gain CMFB to
generate the voltage reference, if the PSRR+ is an issue. Whenever it is, alternative
techniques should be used.
5.3.4.2 PSRR−
A similar analysis can be carried out to derive the effect of the negative power supply
voltage noise57 on the transconductor performance. Again, as eq. (5.89) shows, the
PSRR will be highly dependent on the CMFB:

vout voutcm
≈ (ACMb − ACMa ) (5.89)
vSS vind =0 vSS voutCM=0
where (voutcm /vSS )|voutCM=0 is the output to vSS gain for the CMFB. Under the same
assumptions as before

voutcm gm6 Cgnd Cgnd

≈ − (5.90)
vSS voutCM=0 gds5 + gdsF6 CT a CT b
Considering only mismatch between capacitances and using the eq. (5.83) for the
differential gain
(Cin /CT )
PSRR− ≈ (5.91)
((CCM /CT )b − (CCM /CT )a )(voutcm /vSS )|voutCM=0
As expected, the PSRR− is mostly affected by the ground connected input in the
FGMOS transistors of the CMFB. As an example, the resulting PSRR− using the
same values for the parameters as for the PSRR+ as well as Cgnd = Cin = 0.5 CT
would be 66 dB. But this is assuming that both of the capacitances connected to ground
in the CMFB have identical nominal values. However, if they are different, and for
example Cgndb > Cgnda because Cgndb is also used to generate the reference voltage,
the PSRR− value will be much smaller. Let us assume that Cgnda = (2/3)Cgndb .
In this case, the PSRR− would be reduced to just 36 dB. Comparing these two
numbers, the conclusions and recommendations are the same as for the PSRR+: if
the negative power supply voltage noise is a concern, the use of VSS should be avoided
for the generation of the reference voltage in the CMFB and alternative techniques
should be employed. However, using VSS to bias the transistors within this block in
the strong inversion saturation region is not much of a problem being still possible to
achieve a reasonably good value for the PSRR−58 .
57 v , which in this case is ground.

SS
58 Again, this is assuming only mismatch between the input capacitances in the FGMOS devices. The
contribution of the other terms must be added to these results for an accurate estimation of the PSRR.
Vid Gm3 Gm2 Voutl Gm1 Voutb
ViCM
Figure 5.20 Second-order filter schematic
5.3.5 A second-order filter prototype

A second-order filter can be built using this transconductor as the core block as it
is illustrated in this section. The chosen power supply voltage for our example is
1.25 V and the technology is a 0.35 μm CMOS process [217] with threshold voltages
of VTn = 0.45 V and VTp = −0.62 V59 .
The starting point for the design is the state space equations:
ẋ1 = −ωo2 x1 − ωo1 x2 + ωo3 xi (5.92)
ẋ2 = 2ωo2 x1 (5.93)
Three ωoi determine the filter programmability. The cut-off frequency, quality factor
and gains depend on them as

ωo = 2ωo1 ωo2 (5.94)
.
ωo2
Q= 2 (5.95)
ωo1
ωo3
HLP (0) = (5.96)
ωo1
ωo3
HBP (ωo ) = (5.97)
ωo1
The block diagram and symbol of the second-order filter is shown in Fig. 5.20,
where
Gmi w
ωoi = = β1 ISSi (5.98)
C 2C
As eq. (5.98) shows the ωoi parameters can be modified by changing the tail current
ISSi in each transconductor. The signal Voutb is the output of the bandpass filter and
Voutl the corresponding to the lowpass section.
The filter is designed using 1.5 pF integrating capacitors. The aspect ratio for the
input transistors is 30 μm/0.35 μm. The input capacitances are 70 fF and the capac-
itance that controls the input threshold voltage shift is 140 fF. A microphotograph of
the fabricated prototype is shown in Fig. 5.21. The total area is 0.05 mm2 . Examples
of the circuit performance are shown from Fig. 5.22 to Fig. 5.24, but a more detailed
59 Similar results were obtained in a 0.8 μm process [214] with threshold voltages V
Tn = 0.82 V and
VTp = −0.8 V, with the exception of the area and the tuning range that moves towards lower frequencies.
Figure 5.21 Microphotograph of the filter prototype
–2
–10
V outb (dB)
–20
1 × 10e6 5
f (Hz)
Figure 5.22 Cut-off frequency tuning (Iss (μA) = [4.7, 100])
description can be found in Table 5.4. The filter can be programmed in 0.7 decades
being the maximum power consumption 0.5 mW for the maximum frequency. The
DR is larger than 40 dB, being the maximum signal, for a THD of 1 per cent, 1.2 Vpp .
In any case, this is not the optimum design but just an example of typical results that
can be achieved with this kind of topology.
10
[Iss2, Iss1] = [52 μA, 4.7 μA]
5 [Iss2, Iss1] = [28 μA, 9 μA]
Voutb (dB) 0 [Iss2, Iss1] = [15.6 μA,15.6 μA]
[Iss2, Iss1] = [4.85 μA,50.5 μA]

–5
–10
–15
1 × 10e6 3
f (Hz)
Figure 5.23 Quality factor tuning60
0.8
0.6
THD (%)
0.4
0.2
0.0
0 1
A (V)
Figure 5.24 THD versus 100 kHz differential input for the lowpass filter with 1 MHz
cut-off frequency and quality factor of 1
The results in Table 5.4 demonstrate the suitability of the FGMOS technique for
intermediate frequencies. Again, by combining the extra degrees of freedom that the
FGMOS transistor offers it is possible to push down the power supply voltage in
circuits for IF frequencies applications while keeping a good DR.
60 I
ss2 and Iss1 are the bias currents used to tune ωo2 and ωo1 , respectively.
Table 5.4 Summary of performance for the IF BP

second-order filter
Power supply voltage 1.25 V

Area 0.05 mm2
fo @ISS = [2.75 μA, 145 μA] 700 kHz–5.2 MHz
Q 0.7–5
Input range (THD < 1%) 1.2 Vpp
DR (THD = 1%) >40 dB
Noise floor <−60 dB
PSRR >36 dB
Maximum power@5 MHz 543 μW
5.4 A second-order FGMOS filter in a single poly technology
Until now, the only technological requirement for the design of circuits based on
FGMOS transistors has been to have a double poly technology. This is because
these circuits’ functionality is based on capacitive couplings and using double poly
is the easiest and more accurate form of realising capacitors. However, theoretically
speaking it could also be possible to implement capacitors using metal on top of
polysilicon (poly). The aim of this section is to analyse some of the consequences
and somehow the feasibility of using a purely digital technology, without a second
layer of poly and low quality metal/poly capacitors. The analysis uses as example the
transconductor block and filter described in the previous section.
5.4.1 Effects of the metal–poly capacitors

The main drawback of having low quality metal/poly capacitors in FGMOS devices is
the fact that a much larger area is required to implement a certain capacitance value.
An alternative to this is to scale the capacitance values down while still keeping
the effective inputs capacitive weights. However, both cases suffer from the same
problem: the parasitic capacitance to the well beneath the capacitor might be of the
same order of magnitude as the nominal values, which causes a reduction of the
transistor effective transconductance. Let us express the effective input weight as
Ci ki Ai Cpma + ki Pi Cpmp
≈ (5.99)
CT CTa + CTp
CTa ≈ (1 + ki )Ai (Cpma + Cpwa ) + kca Cox At (5.100)
CTp ≈ (1 + ki )Pi (Cpmp + Cpmp ) + kcp Cox Pt (5.101)
εox
Cox = (5.102)
tox
Table 5.5 Typical capacitance-related

technological parameters
Cpma (fF/μm2 ) 0.054

Cpmp (fF/μm) 0.05
Cpwa (fF/μm2 ) 0.12
Cpwp (fF/μm) 0.05
where Cpma is the poly1 to metal capacitance per unit of area, Cpmp per unit of
perimeter, and Cpwa and Cpwp are the poly1 to well capacitance per unit of area and
perimeter, respectively. Typical values of these parameters for a 0.35 μm technology
are shown in Table 5.5. At and Pt are the area and perimeter of the transistor, and (ki Ai )
and (ki Pi ) are the area and perimeter of the equivalent input capacitances. Ai is the
area of the capacitance used to get a certain effective threshold voltage. Parameters
kca and kcp depend upon the transistor operating region, and are related to the intrinsic
parasitic capacitances. The other parameters have their usual meanings. The values in
Table 5.5 show how in, for example this particular technology, the parasitic to the well
will affect, at least, as a reduction of around a factor of two in the transconductance
value.
A problem added to this is that the variations with respect to their nominal values
in metal/poly capacitances can also be much bigger than in poly2/poly1 capacitances.
An example of typical deviation values is 60 per cent for metal/poly1 capacitances
versus 10 per cent for poly2/poly1.
The following sections analyse some of the most important consequences
of reducing the values of the input capacitances in the transconductor in
Fig. 5.18.
5.4.1.1 Offset
One of the main sources of systematic offset in the transconductor in Fig. 5.18 is
the capacitive coupling between gate and drain in the input transistors. The voltage
across the drain of transistors M1(a,b) , M2(a,b) obtained by using the fact that their
currents are the same is
!
CGD2 β1
VD1 ≈ VcmN + VD2 + VTn1 − VTn2
CT2 β2

1
× √ (5.103)
1 − CGS2 /CT2 + (CGD1 /CT1 ) β1 /β2
This voltage will couple to the FG of the input transistor through CGD giving rise to
an equivalent input offset61 :
√
(CGD1 /Cin )(VcmN + β1 /β2 VTn1 − VTn2 )
Voff + ≈ (5.104)
[A − Ad (CGD2 /CT2 ) · (CGD1 /Cin )]
Ad being the transconductor output gain, and
! !
CVDDf CCMn β1 Cin β1 CCM
VcmN = VDD + Voutcm − ViCM − Voutcm
CT2 CT2 β2 CT1 β2 CT1
! !
β1 CVDD β1 CGS1
− VDD + 1− VS1 (5.105)
β2 CT1 β2 CT1
1
A = √ (5.106)
1 − CGS2 /CT2 + (CGD1 /CT1 ) β1 /β2
Subscripts have been used to refer to the name of the transistor. In this case, since the
absolute values of the capacitances can change a lot, different total capacitances have
been considered. The differential offset can be obtained by subtracting two equations
with the form of eq. (5.103) for both differential branches, a and b62 :

(CGD1a /Cina ) · ta
offset ≈
(A a − Ada (CGD2a /CT2 )(CGD1a /Cina ))

(CGD1b /Cinb ) · tb
− (5.107)
(A b − Adb (CGD2b /CT2 )(CGD1b /Cina ))
where
!
β1a
t(a,b) = VcmN(a,b) + VTn1(a,b) − VTn2(a,b) (5.108)
β2a
Ideally ta and tb should be equal but, in practice, any mismatch affecting any of the
three terms in eq. (5.108) generates differences between the two of them. A new
variable del1 will account for these variations as shown in eq. (5.109):
tb = (1 + del l)ta (5.109)
Equation (5.107) shows how as CGD becomes smaller in comparison to the input
capacitance the offset decreases as well. The parameters that affect the offset the
most are the common mode at the input, and the variations between parameters
related to the term tb , as for example threshold voltages or βs. But also, the mismatch
between the input capacitances can be crucial. These conclusions are illustrated in
Fig. 5.25. The nominal values used to obtain these graphs are the same as for the
design example at the end of the section: CGD = 3.3 fF, CT1 = 81 fF, Ci = 11.15 fF.
61 A single effective input with associated capacitance C has been considered for the sake of simplicity.
in
62 The subscripts a and b have been added to distinguish between both.
(a) offset (V)
0.08
0.06
0.04
0.02
0.1
1.4
0 1.2
1
del1 –0.1 0.8
V inCM (V)
(b)
offset (V)
0.04
0.00
0.1
0.1
0.0
0.0
del1 –0.1 –0.1 (Cina /Cinb )– 1
Figure 5.25 (a) Offset versus common mode at the input and mismatch affecting
threshold voltages and β. (b) Offset versus mismatch in the input
capacitances and also versus mismatch affecting the threshold voltages
and β
5.4.1.2 Frequency response of the OTA

An exact analysis of the frequency response is complex and does not provide so
much information, so this section just describes very briefly what are the differences
between the frequency response of the transconductor built with metal/poly capacitors
and the same transconductor with two layers of poly capacitors instead.
As before, the DC gain of the transconductor block is given by eq. (5.83). How-
ever, the output resistance given by eq. (5.79) is now much smaller, because of the
increased output conductance of the FGMOS devices since the ratio CGD /CT is larger.
Also, the effective transconductance is lower because of the large parasitic term added
to the value of CT . All this contributes to reduce the DC gain.
The dominant pole moves towards lower frequencies due to the increased output
conductance. Moreover, due to the significance of the parasitic capacitances (which
are now comparable to the input ones), the second and third pole, that normally
appear in the frequency response of a transconductor differential pair, move closer to
the dominant pole.
The frequency response of the CMFB changes mostly because of the reduction
in its output resistance. The term CGD6 gm6 /CT is more significant in the expression
of gdsF6 which translates to a reduction of the amplifier gain.
5.4.1.3 The integrating capacitance

This section analyses the effects of the integrating capacitor parasitics on the frequency
response of the integrator circuit built with the transconductor in Fig. 5.18. Let us
start with a single pole model of the integrator for both single branches:
Gm vid ao (vD2a + vD2b )

sCpa vD2a + sC(vD2a − vD2b ) = − − − Gouta vD2a
2 s+p 2
Gm vid ao (vD2a + vD2b )
sCpb vD2b + sC(vD2b − vD2a ) = − − Goutb vD2b
2 s+p 2
(5.110)
where Cpa and Cpb represent the parasitic capacitances in branches a and b, respec-
tively. Ideally they should be the same if the capacitor is designed in a differential
way (two equal value capacitors in parallel but with inverted plates), but because
the expected variations with respect to their ideal values are higher in a metal/poly
realisation, they will be considered different. The frequency response of the common
mode feedback is modelled as ao /(s + p) where
gdsF6b + gds5b
p≈ (5.111)
i=(a,b) (CCMi + CCMni + CCMpi + CGS4i )
Cin (gm1 (CCM /CT ) + gm4 )gm6
ao ≈ × (5.112)
CT (CCM + CCMn + CCMp + CGS4 )Gout
The common and differential gain can be obtained by solving the systems of
equations (5.110). Assuming Gouta = Goutb = Gout
voutCM Gm (Cpa − Cpb )(s + p)s

= (5.113)
vid 2D(s)
where
vd2a + vd2b
voutCM = (5.114)
2
Equation (5.113) shows how the larger the mismatch between both parasitic capac-
itances, Cpa and Cpb , is the bigger the common mode gain at the output. In the worst
case scenario, this could originate premature clipping.
The differential gain is given by

vout = vd2b − vd2a (5.115)
vout (s + p)(2Gout + s(Cpa + Cpb )) + 2ao
= Gm (5.116)
vid D(s)

Cpa + Cpb G2 Cpa Cpb
D(s) = 4ao s C + + out + 2s2 (s + p) Cpa + Cpb + C
4 2ao C
+ 2[Gout (Cpa + Cpb ) + 2Gout C](s + p)s + 2Gout (Gout p + ao ) (5.117)
Equation (5.116) shows two parasitic zeros in the integrator located at
!
p Gout 1 2 2
4Gout (2ao + Gout p)
z1,2 = − − ± p + −4
2 (Cpa + Cpb ) 2 (Cpa + Cpb ) (Cpa + Cpb )
(5.118)
Also, the dominant pole shifts towards a different frequency:
Gout (Gout p + ao )
wp1 ≈ (5.119)
2ao (C + (Cpa + Cpb )/4)
The previous analysis applies to the integrator irrespectively of whether the inte-
grating capacitance is implemented in either poly1/poly2 or metal/poly. The difference
is that in the metal/poly implementation ao is expected to be smaller because of the
larger contribution of the parasitics to the value of CT . Besides, the integrator will
have extra losses (as predicted in eq. (5.119)) that might affect in more or less degree63
the filter parameters when the integrator is used in a higher order topology.
5.4.2 Experimental results

This section illustrates the performance of a second-order biquad, identical to that
previously described in Section 5.3, fabricated with metal/poly capacitors whose
values are collected in Table 5.6. The area of the input transistor was 8 μm2 .
A microphotograph of the filter is shown in Fig. 5.26.
Figure 5.27 illustrates the THD versus the input amplitude measured in the lowpass
filter for a single input tone of 10 kHz. A peak amplitude of 0.4 V results in less than
1 per cent THD. The HD3 is under 0.5 per cent. Figure 5.28 and Fig. 5.29 illustrate
the tuning of different filter parameters. The cut-off frequency can only be changed in
the range from 75 kHz to 120 kHz. The maximum possible cut-off frequency is lower
than in the example in Section 5.3 due to both the smaller transconductance and the
effect of the parasitic capacitances in the filter state space equations. The maximum
achievable quality factor is around 2. This value is limited because of the degradation
of the output resistance. The gain can be programmed in a 15 dB range. The resulting
offset for the low pass filter is 75 mV, and additional frequency roots appear just half a
63 Depending on how the filter frequency band compares with the values in eq. (5.118) and (5.119).
Table 5.6 Input capacitances for the

FGMOS transistors in a
metal/poly realisation of
Fig. 5.18 (fF)
CVDDf & Cgndf 126.2

CCMn & CCMp 37.5
Ci 11.15
CVDD 47.62
CCM 11.15
Figure 5.26 Microphotograph of the second-order filter in Fig. 5.20, implemented

with metal/poly capacitors
decade from the nominal cut-off frequency. They can be observed in Fig. 5.28 where
the slope of the high frequency rejection band is steeper than the ideal 20 dB/dec for
the bandpass realisation. Also, the lowpass transfer function shows a zero at around
150 kHz for a filter with a cut-off frequency of 70 kHz. Regarding the low frequency
losses of the bandpass filter, they are just under 20 dB below the peak value. This
is related to the degradation of the gain in the transconductor due to mismatch. A
summary of performance can be found in Table 5.7.
The main conclusion that can be extracted from these experimental results is
that FGMOS circuits implemented in single poly technologies64 can work, but their
performance will be much poorer than the performance of the same circuit realised
in a double poly technology. However, the latter will be equivalent if: (a) the area is
not a constraint; (b) the technology is such that the parasitic capacitance between the
FG and the bulk/well underneath is of the same order of magnitude as in a double
poly technology; (c) the matching between capacitors implemented in the metal/poly
technology is comparable to the matching in a double poly technology.
64 With no high quality capacitors.

2.0
1.6
1.2
THD (%)
0.8
0.4
0.0
0.2 0.4 0.6 0.8 1
A (V)
Figure 5.27 Total Harmonic Distortion versus input amplitude for the filter in
Fig. 5.20, implemented with metal/poly capacitors
–25
Voutb(dB)
–30
–35
105
f (Hz)
Figure 5.28 Cut-off frequency tuning in the BP function for the filter in Fig. 5.20,
implemented with metal/poly capacitors
5.5 Comparison
The last two chapters have presented the design of three different OTA-C integrator/
filter topologies, based on the FGMOS transistor operating in the strong inversion
region. All the topologies have as common aims the reduction of the power con-
sumption and power supply voltage while keeping a good circuit performance. The
–15
–20
–25
dB
–30
–35
–40
104 105
f (Hz)
Figure 5.29 Gain tuning in the LP function for the filter in Fig. 5.20, implemented
with metal/poly capacitors
Table 5.7 Summary of performance parameters for the filter in

Fig. 5.20, implemented with metal/poly capacitors
Technology 0.35 μm
Power supply voltage 1.25 V
Area 0.05 mm2
fo 70 kHz–125 kHz
Q <2
THDmax (Vpp < 1 V@10 kHz, Q = 1, <42 dB
HLP (0) = 1, fo = 90 kHz)
Maximum power 56 μW
first one employs the features of the transistor in the strong inversion ohmic region to
linearise the voltage to current conversion; the second one does the same but taking
advantage of the strong inversion saturation operation instead; the last one is not a
linearised structure but still achieves a large linear input range thanks to the signal
compression at the transistor FG. This topology also proves the feasibility of the
device to operate at intermediate frequencies consuming very little power even with
a low power supply voltage.
Table 5.8 compares the performance of the three main transconductor blocks in
terms of the main design target parameters.
As the table shows the first two transconductors are suitable for low frequency
applications. The second transconductor (transconductor based on FGMOS transis-
tors biased in the strong inversion saturation region) requires a higher voltage supply to
operate, offering however a larger programmability range than the first one (transcon-
ductor based on FGMOS transistors biased in the strong inversion ohmic region).
Table 5.8 Comparison between the three OTA realisations
OTA Chapter 4 OTA Chapter 5 OTA Chapter 5

Section 5.2 Section 5.3
Power supply 1.5 V 2V 1.25 V

Power consumption 6.5 μW (1 kHz) 8 μW (1 kHz) 181 μW (5 MHz)
Programming rate [Gm ] 5.3 [11.9 ns, 62.8 ns] 100 [3 ns, 300 ns] 7.1 [6.6 μs, 46.7 μs]
Input range (THD < 1%) 2 Vpp 2 Vpp 1.2 Vpp
The third transconductor is more suitable for intermediate frequencies applications.

Nevertheless, as it is the only non-linearised topology, the input range is smaller.
When comparing the values in the table, it should be kept in mind that the compar-
ison between the two first and the third topologies is not totally fair, since the results
shown for the last one correspond to a faster technology with lower threshold voltage
values. This is also one of the reasons why, even when it operates at a frequency that it
is almost three orders of magnitude higher than the maximum operating frequency of
the other two, the power consumption is only around 20 times larger. Another reason
is that in order to have a wide input range in the first two designs, the difference
VFG − VTn has to be large enough in the middle of the range (to keep the transistors
properly biased for low input signal values as well). This unavoidably increases the
power consumption.
Notation
β1(a,b) β parameter of input transistors M1a and M1b , respectively,
in Fig. 5.18. Only different when there is mismatch
β1 β parameter of input transistors in the transconductor in
Fig. 5.1. Also, in Fig. 5.18 when β1a = β1b
βi Mi β parameter
γn Parameter associated to the Flicker noise of n-channel
transistors (Table 5.1)
γp Parameter associated to the Flicker noise of p-channel
1i for i = [2, 4] Percentage of mismatch in capacitance Cin for transistor Mi
(eq. (5.33). Transconductor in Fig. 5.1)
2i for i = [2, 4] Percentage of mismatch in capacitance CCM for transistor
Mi (eq. (5.34). Transconductor in Fig. 5.1)
3i Percentage of mismatch in capacitance CVDD for transistor
Mi (eq. (5.52). Transconductor in Fig. 5.1)
ωo Cut-off frequency
ωo1 Coefficient of x2 in the state space eq. (5.60) and of both x1

and x2 in the state space eq. (5.92)
ωo2 Coefficient of x1 in the state space eq. (5.60) and also in
state space eq. (5.93)
ωo3 Coefficient of x1 in the state space eq. (5.61) and of the
input in state space eq. (5.92)
ωo4 Coefficient of the input (u) in the state space eq. (5.60)
ρn Parameter associated to the thermal noise of n-channel
ρp Parameter associated to the thermal noise of p-channel
(a,b) Subscripts used to simultaneously refer to transistors M1a
and M1b , respectively
Ao DC gain of the CMFB (Fig. 5.5, eq. (5.24))
a0 See eq. (5.26). Value of a01 and a02 when both of them are
the same
a0i for i = [1, 2] See eqs. (5.22) and (5.23)
A See eq. (5.106)
ACM See eq. (5.78)
ACMa ACM parameter for the transconductor branch named with
subscripts a in Fig 5.18
ACMb ACM parameter for the transconductor branch named with
subscripts b in Fig. 5.18
Ad Transconductor differential output gain (Fig. 5.18,
eq. (5.83))
Ai Area of the capacitance used to get a certain effective
threshold voltage (eq. (5.99))
At Area and perimeter of the transistor (eq. (5.99))
Cc Capacitance connected to either Vb1 or Vb2 (depending on
the transistor) in the transconductor in Fig. 5.1
CCM Capacitance connected to the output of the CMFB in the
transconductor input FGMOS transistors (Fig. 5.1 and
Fig. 5.18)
CCMn Value of the capacitance connected to the output of the
CMFB in the n-channel cascode transistor (Fig. 5.18)
CCMp Value of the capacitance connected to the output of the
CMFB in the p-channel cascode transistor (Fig. 5.18)
(CCM /CT )a Input weight associated to capacitance CCM in transistor
M1a when there is mismatch (eq. (5.88). Transconductor in
Fig. 5.18)
(CCM /CT )b Input weight associated to capacitance CCM in transistor
M1b when there is mismatch (eq. (5.88). Transconductor in
Fig. 5.18)
(CCM /CT )i Input weight associated to capacitance CCM in transistor Mi

when there is mismatch (eq. (5.34). Transconductor in
Fig. 5.1)
(CCMn /CT )a Input weight associated to capacitance CCMn in transistor
Fig. 5.18)
(CCMn /CT )b Input weight associated to capacitance CCMn in transistor
Fig. 5.18)
(CCMp /CT )a Input weight associated to capacitance CCMp in transistor
Fig. 5.18)
(CCMp /CT )b Input weight associated to capacitance CCMp in transistor
Fig. 5.18)
CGDi Gate to drain coupling for transistor Mi(a,b) in Fig. 5.18
CGSi Gate to source coupling for transistor Mi(a,b) in Fig. 5.18
Cgnd Value of the capacitance connected to ground in the CMFB
in Fig. 5.5 and in Fig. 5.19
Cgndf Value of the capacitance connected to ground in the
p-channel cascode transistor (Fig. 5.18)
(Cgnd /CT )a Input weight associated to capacitance Cgnd in transistor M6a
when there is mismatch (eq. (5.90). CMFB in Fig. 5.19)
(Cgnd /CT )b Input weight associated to capacitance Cgnd in transistor M6b
when there is mismatch (eq. (5.90). CMFB in Fig. 5.19)
Ci for i = [1, N ] Value of the capacitances connected to the effective inputs in
the transconductor in Fig. 5.18 (when there is more than one)
Cin Value of the effective inputs capacitances in the CMFBs
(each one of them connected to a single output. Fig. 5.5 and
Fig. 5.19) and in the input transistors of the transconductors
in Fig. 5.1 (connected to Vin1 and Vin2 ) and Fig. 5.18 (when
the transconductor only has one effective input)
(Cin /CT )i Input weight associated to capacitance Cin in transistor Mi
when there is mismatch (eq. (5.33). Transconductor in
Fig. 5.1)
(Cin /CT )a Effective input weight in the transconductor branch named
with subscripts a (Fig. 5.18) assuming mismatch
(Cin /CT )b Effective input weight in the transconductor branch named
with subscripts b (Fig. 5.18) assuming mismatch
CLF Load at the output of the CMFB (eq. (5.25))
Cox Capacitance per unit of area for a MOS transistor
Cpa Parasitic capacitance at the output in branch a for the
transconductor in Fig. 5.18
Cpb Parasitic capacitance at the output in branch b for the
Cpma Poly1 to metal capacitance per unit of area

Cpmp Poly1 to metal capacitance per unit of perimeter
Cpwa Poly1 to well capacitance per unit of area
Cpwp Poly1 to well capacitance per unit perimeter
CT Total capacitance seen by the FGs. It has been assumed
the same in all the FGMOS transistors unless the opposite is
said
CTa Total area capacitance (eq. (5.100))
CTi Total capacitance for transistor Mi(a,b) in Fig. 5.18
considering mismatch (metal-poly realisation)
CTp Total perimeter capacitance (eq. (5.100))
CVDD Value of the capacitance connected to VDD in the FGMOS
input transistors (Transconductors in Fig. 5.1 and Fig. 5.18)
(CVDD /CT )a Input weight associated to capacitance CVDD in transistor
Fig. 5.18)
(CVDD /CT )b Input weight associated to capacitance CVDD in transistor
Fig. 5.18)
(CVDD /CT )i Input weight associated to capacitance CVDD in transistor Mi
when there is mismatch (Transconductor in Fig. 5.1 and
CMFB in Fig. 5.5)
CVDD6 Value of the capacitance connected to VDD in the CMFB
(Fig. 5.19) for the transconductor in Fig. 5.18
CVDDf Value of the capacitance connected to VDD in the n-channel
cascode transistor (Fig. 5.18)
(CVDDf /CT )a Input weight associated to capacitance CVDDf in transistor
Fig. 5.18)
(CVDDf /CT )b Input weight associated to capacitance CVDDf in transistor
Fig. 5.18)
del1 Percentage of mismatch between ta and tb (eq. (5.109))
FG1a FG in transistor M1a (Transconductor in Fig. 5.18)
FG1b FG in transistor M1b (Transconductor in Fig. 5.18)
fo Cut-off frequency in Hz
gdsi MOS transistor Mi output conductance
gdsFi Effective conductance for FGMOS device Mi
gmbi Bulk transconductance for transistor Mi
Gm Transconductance of the transconductor in Fig. 5.1 (eq. (5.4))
Gmi for i = [1, 4] Value of the transconductance that implements the
coefficient ωoi
gmi Transistor Mi transconductance
Gm(max) Maximum value of the transconductance for the
transconductor in Fig. 5.1 (eq. (5.5))
Gouta Output conductance of the a branch in the transconductor in

Fig. 5.18
Goutb Output conductance of the b branch in the transconductor in
Fig. 5.18
Goutn Output conductance for the bottom (n-type) part of the
integrator in Fig. 5.1 (eq. (5.17))
Goutp Output conductance for the top (p-type) part of the
integrator in Fig. 5.1 (eq. (5.18))
Goutn See eq. (5.27)
Goutp See eq. (5.28)
HBP (ωo ) Bandpass filter gain at the cut-off frequency
HLP (0) Lowpass filter DC gain
I1 Sum of M1 and M2 drain currents in the transconductor in
Fig. 5.1
I2 Sum of M3 and M4 drain currents in the transconductor in
Fig. 5.1
I1a Drain current in transistor M1a (Transconductor in
Fig. 5.18)
I1b Drain current in transistor M1b (Transconductor in
Fig. 5.18)
IB Bias current provided by the load (eq. (5.2) and Fig. 5.1)
IBmax Maximum value of the bias current (eq. (5.7))
IBmin Minimum value of the bias current (eq. (5.8))
Iout Output current in the transconductors (Fig. 5.1 and
Fig. 5.18)
ISS Tail current in the transconductor in Fig. 5.18
k1 See eq. (5.39)
kca Parameter related to the intrinsic parasitic capacitances
(eq. (5.100))
kcp Parameter related to the intrinsic parasitic capacitances
(eq. (5.101))
ki Ai Area of the total designed capacitor used to process signals
ki Pi Perimeter of the total designed capacitor used to process
signals
offset Differential offset (eq. (5.107))
P Power consumption
p Pole of the CMFB (eq. (5.25) for the CMFB in Fig. 5.5 and
eq. (5.111) for the CMFB in Fig. 5.18)
Pi Perimeter of the capacitor used to get a certain effective
threshold voltage (eq. (5.99))
Q Quality factor
Pmin Minimum value of the power consumed by the
Pmax Maximum value of the power consumed by the
PSRR+ Positive power supply rejection ratio

Pt Perimeter of the transistor (eq. (5.99))
Rout Output resistance
Routa Output resistance of the a branch in the transconductor in
Fig. 5.18
Routb Output resistance of the b branch in the transconductor in
Fig. 5.18
Si Power
√ spectral density of noise for transistor Mi
Si Si
u Input in the state space equation (5.60)
un Single input (eq. (5.66))
up Single input (eq. (5.66))
VcmN See eq. (5.105)
V1n Single voltage implementing x1n in eq. (5.68)
V1p Single voltage implementing x1p in eq. (5.68)
V2n Single voltage implementing x2n in eq. (5.69)
V2p Single voltage implementing x2p in eq. (5.69)
VBP Bandpass output (eq. (5.68))
Vb Bias voltage for zero transconductance in the
transconductor of Fig. 5.1 (eq. (5.8))
Vb1 Bias voltage for the integrator in Fig. 5.1
Vb1min Vb1 minimum value (eq. (5.5))
Vb2 Bias voltage for the integrator in Fig. 5.1
Vb2max Vb2 maximum value (eq. (5.5))
Vcasn Bias voltage for the n-channel cascode transistor in the
integrator in Fig. 5.1
Vcasp Bias voltage for the p-channel cascode transistor in the
integrator in Fig. 5.1
V CM VinCM + (CCM /Cin )Voutcm + (CC /Cin )Vb2
VCM,min Minimum value of V CM , in Fig. 5.9
VCM,max Maximum value of V CM , in Fig. 5.9
VD1 M1 drain voltage for the transconductor in Fig. 5.1. Also,
M1a and M1b drain voltages when common mode signals
are applied and both transistors are assumed to be identical
for the transconductor in Fig. 5.18
VD1min Minimum M1 drain voltage for the transconductor in
Fig. 5.1
VD1a M1a drain voltage in Fig. 5.18
VD2 M2a and M2b drain voltages when common mode signals
VD2a M2a drain voltage in Fig. 5.18. It is also one of the single
output voltages
VD3 M3 drain voltage in the transconductor in Fig. 5.1
VD12 M12 drain voltage (Fig. 5.5)
Vi(a,b) for i = [1, N ] Effective inputs in the transconductor in Fig. 5.18
ViCM Common mode at the input of the transconductor in

Fig. 5.18
Vid Differential input voltage for the transconductor in Fig. 5.18
Vinn Single voltage implementing un in eq. (5.68)
Vin1 Single input voltage for the transconductor in Fig. 5.1
Vin2 Single input voltage for the transconductor in Fig. 5.1
Vin(a,b) Effective inputs in branches a and b, for the transconductor
in Fig. 5.18 when only one input is considered
Vind Differential input for the integrator in Fig. 5.1 and for the
transconductor in Fig. 5.18 when only one effective input is
considered
VinCM Common mode at the input of the integrator in Fig. 5.1.
Also, common mode at the effective input of the
transconductor in Fig. 5.18 when only one input is
considered
VinRMS RMS spectral density of noise at the effective input for the
2
vin
Equivalent power spectral density of noise at the effective
f input
Vinp Single voltage implementing up in eq. (5.68)
VLP Lowpass output (eq. (5.69))
Voff + Equivalent single side offset (eq. (5.104))
Vout Differential output voltage for the transconductors in
Fig. 5.1 and Fig. 5.18
VoutCM Common mode at the output of the transconductor in
Fig. 5.1 and in Fig. 5.18
Voutcm Output of the CMFB
Voutcm(max) Maximum Voutcm (eq. (5.7))
Voutcm(min) Minimum Voutcm (eq. (5.7))
Vout1 Single output voltage for the transconductor in Fig. 5.1
vout1
2
Equivalent power spectral density of noise at the effective
f input of the CMFB used for the transconductor in Fig. 5.1
(Fig. 5.5)
Vout2 Single output voltage for the transconductor in Fig. 5.1
Voutb Bandpass output in the filter in Fig. 5.20
Voutl Lowpass output in the filter in Fig. 5.20
VS M12 and M13 source voltage (Fig. 5.11)
VS1 Input transistors M1(a,b) source voltage (Transconductor in
Fig. 5.18)
VS3 M3a and M3b source voltages when common mode signals
(Fig. 5.18)
VS3a M3a source voltage in Fig. 5.18

VS8 M8 source voltage (Transconductor in Fig. 5.1)
VTFG Effective threshold voltage (eq. (5.9) and eq. (5.75))
(Different values for different transistors)
VTn n-channel MOS transistors threshold voltage (it has been
assumed the same in all the devices, unless the opposite is
said)
VTn1 M1 threshold voltage considering mismatch and body effect
(eq. (5.103), Fig. 5.18)
VTn2 M2 threshold voltage considering mismatch and body effect
(eq. (5.103), Fig. 5.18)
|VTP | p-channel MOS transistors threshold voltage (it has been
assumed the same in all the devices, unless the opposite is
said)
w wi when all the weights for the effective inputs have been
assumed to have the same value
wi Ci /CT . Effective inputs weight (Transconductor in
Fig. 5.18)
x1 Bandpass output
x1n Single state variable in eq. (5.66). (Related to the bandpass
output)
x1p Single state variable in eq. (5.66). (Related to the bandpass
output)
x2 Lowpass output
x2n Single state variable in eq. (5.66). (Related to the lowpass
output)
x2p Single state variable in eq. (5.66). (Related to the lowpass
output)
xi Input in the state space equation (5.92)
Chapter 6
Low power analog continuous-time Gm -C
filtering using the FGMOS in the weak
inversion region
The following two chapters are devoted to exploring the potential of the FGMOS tran-
sistor operating in weak inversion for the design of analog low-voltage micropower
continuous-time filters. It is demonstrated that the use of FGMOS in weak inversion
as an alternative to conventional MOS devices relaxes the constraints relating to volt-
age supply and frequency response. Moreover, the increased number of terminals
in the device permits a simplification in the topologies required to realise a certain
mathematical function. This brings with it a reduction in power consumption, as well
as other advantages such as a lower noise floor.
The term ‘weak inversion’ defines the operating mode in an MOS transistor
in which the bulk surface charge is inverted in relation to the rest of it, but still
it is mostly due to the charge in the depletion region. The corresponding inversion
layer charge can nevertheless cause non-negligible conduction. Weak inversion mode
is especially suited for the design of micropower circuits for several reasons. The
term ‘micropower’ defines a class of circuits with power consumption of just a few
microwatts. In order to obtain these power levels the maximum current the circuit
can handle has to be very small. The weak inversion region is very convenient when
these requirements have to be met for the following reasons:
1. Assuming that the required power consumption sets a very small, fixed value
for the current, the maximum operating frequency of a single transistor is determined
by the gate oxide capacitances, CGB , CGS and CGD . In order to maximise the device
bandwidth the latter needs to be kept as small as possible, which implies a small
device area. Assuming that the minimum transistor width is chosen, the designer
must choose a length of transistor and level of inversion in the channel that meets
the current constraint. If the designer reduces the level of inversion, the length of
the transistor may also be reduced. This reduces CGB , CGS and CGD , which in turn
increases the bandwidth.
2. The maximum voltage differences between the terminals of a device working in

the weak inversion region are smaller than the values demanded by strong inversion
operation. This allows for lower supply voltages, which also reduces the power
consumption.
3. The behaviour of the transistor is ruled by an exponential law whose derivative
is another exponential function. This makes it possible to use nonlinear processing
internally to implement linear state-space equations, in a simple form [182]. By doing
this, the internal nodal voltage swings can be reduced and less power is needed to
charge and discharge the nodes.
This chapter describes a linearised Gm -C topology in which the linearisation is
achieved by means of two ‘floating gate’ blocks: a nonlinear transconductor with
three output currents and a square-root block based on FGMOS transistors biased
in the weak inversion region that processes the three currents provided by the first
block, thus giving rise to a fully linear topology.
The chapter begins with a reformulation of the Translinear Principle (TP), which
is an ideal technique for implementing nonlinear functions using currents in FGMOS.
The advantages of an FGMOS implementation compared with a normal MOS imple-
mentation in a low-voltage context are then discussed. The linearised Gm -C integrator
is described afterwards followed by an analysis of second-order effects, such as PSRR
or THD, that can affect its performance. Finally the design of a second-order filter
based on these blocks is explained and experimental results are shown to illustrate
the conclusions.
6.1 FGMOS translinear principle
In 1975, B. Gilbert applied the name ‘translinear’ to an emerging class of circuits

whose large-signal behaviour was based on the exponential current–voltage charac-
teristic of the bipolar transistor and the linear voltage law of a voltage loop [218]. By
combining both principles, a nonlinear product of current densities within different
loops can be obtained that serves as a basis for many nonlinear circuits, includ-
ing wideband analog multipliers, RMS-DC converters and vector-magnitude circuits
[218–222]. The general circuit principle describing the specific topological arrange-
ment for the exponential devices which gives rise to the desired mathematical function
is called the Translinear Principle (TP). This section explains how the TP can be used
advantageously when FGMOS devices are available.
Consider two FGMOS transistors M0 and M1 with a grounded source and a
common input, V0i at M0 and V1j at M1, as illustrated in Fig. 6.1. M0 has N and M1
has M inputs. The Kirchoff’s voltage law in this trivial loop is
−V0i + V1j = 0 (6.1)
Under the assumptions that the parasitics to the FG are very small compared with
the input capacitances and that the charge trapped during the fabrication process is
negligible, the drain current in FGMOS devices working in weak inversion saturation
Gm -C filtering using FGMOS in the weak inversion region 187
I0 I1
V01 V11
M0 M1
V0i =V1j
V0N V1M
Figure 6.1 FGMOS translinear loop
mode with their source and bulk grounded is given by (see eq. (2.5))

ID = Is e i=1 wi Vi /nUT (6.2)
where wi refers to the input weights and Vi to the inputs. Expressing the voltages in
the translinear loop as a function of the currents flowing through the transistors gives
⎡ ⎤
1/w0i
N
⎢ I0 1 w0k V0k ⎥
V0i = nUT ⎣ln − × ⎦
IS w0i k=1
nUT
k=i
⎡ ⎤ (6.3)
1/w1j M
⎢ I1 1 w1k V1k ⎥
V1j = nUT ⎣ln − × ⎦
IS w1j k=1
nUT
k=j
where w0k , w1k are the equivalent weights for the input k in transistors M0 and M1,
respectively. Rewriting (6.1) using (6.3) and rearranging the terms gives
1/w0i −
N 1/w1j M
I0 k=1 (w0k V0k /w0i nUT ) I1 − k=1 (w1k V1k /w1j nUT )
e k=i = e k=j (6.4)
Is IS
This is the nonlinear relationship between currents, which is the objective of the TP.
The first advantage of using the FGMOS to implement the translinear loops is that
it is possible to realise the exponents in the current function with capacitance ratios,
and hence the source terminal can be fixed to a constant voltage. This fact is quite
significant for two different reasons. The first is that the voltage requirements are
reduced since only a VGS (gate to source voltage drop) is needed in the loop and
therefore there is no need to stack transistors. The second is that the lack of internal
nodes both simplifies and improves the frequency response and eliminates insta-
bilities in certain circuit topologies. This is because the CGS and CGB capacitors
are now tied to a constant voltage in contrast to translinear circuits designed

with MOS transistors, in which the frequency response sometimes exhibits an
undesired and even unstable behaviour at a few kilohertz due to the effects of these
parasitics.
6.2 A Gm -C FGMOS integrator using translinear circuits
This section shows how to use the FGMOS transistors as circuit primitives to design
linear and nonlinear (square-root) circuits. Once more lower voltage operation is
obtained by shifting the voltage levels wherever needed. However, for the first time
now in this book, the FGMOS is presented as a tool to realise nonlinear functions
through the TP. Combining the two saturated modes in the transistor (in strong inver-
sion and weak inversion), a design flow is developed that generates an FGMOS
transistor based circuit from a set of state space equations.
6.2.1 The state-space integrator equations

The starting point of the filter design process is a set of first-order differential equations
of the type
ẋ + αx = ηxi (6.5)
where x is a state variable; xi is the external excitation which can be an independent
source, other state variable or any combination of both; and parameters α and η
are related to the filter specifications. The fundamental idea behind the proposed
implementation relies on the versatile operation of the FGMOS technique which
permits the realisation of complex algebraic operations under low voltage conditions.
In order to make the filter less sensitive to noise and enlarge the linear range at
the input, a fully differential operation is chosen. Hence, equation (6.5) has to be
rewritten as
(ẋp − ẋn ) = −α(xp − xn ) + η(xip − xin ) (6.6)
Assuming that all these signals are voltages that have been obtained from current
signals, the following variable changes are introduced:

Vp − Vn = (Ip − In )/(K1 Ic ) (6.7)

Vip − Vin = (Iip − Iin )/(K1 Iic ) (6.8)
where Ic , Ip , In , Iic , Iip , Iin are the currents and K1 is a constant. These parameters are
discussed in detail later.
Using eq. (6.7) and eq. (6.8) and applying the variable change to the right-hand
side (RHS) of eq. (6.6), the final differential equation is obtained:

Ip − In Iip − Iin
(V̇p − V̇n ) = −α √ + η √ (6.9)
K1 Ic K1 Iic
Cα (In – I p ) Cη ( I in – I ip )
----------------------------- ---------------------------------
K1 Ic K 1 I ic
I
Vp C Vn
Cη ( I in – I ip)
--------------------------------- Cα ( I n – I p )
K 1 I ic -----------------------------
K1 I c
Figure 6.2 Schematic of an ideal circuit that performs the first-order state-space
eq. (6.10)
Then, multiplying both sides in eq. (6.9) by C (physically an integrating capacitor)

allows the RHS of eq. (6.9) to be identified with the current flowing through this
capacitor C:

Ip − In Iip − Iin
I = C(V̇p − V̇n ) = −Cα √ + Cη √ (6.10)
K1 Ic K1 Iic
Equation (6.10) is symbolised in Fig. 6.2 using ideal circuit elements. Referring
to Fig. 6.2, it can be seen that Kirchoff’s current law is evaluated in nodes Vp and Vn
according to eq. (6.10).
6.2.2 FGMOS blocks

6.2.2.1 Nonlinear transconductor
A circuit that performs the current to voltage conversion required by equations (6.7)
and (6.8) is shown in Fig. 6.3. It operates as follows:
If the three FGMOS transistors are equally sized, operate in the strong inversion
saturation region and have the same input weights, given by
wVDD = CVDD /CT = wDD , win = Cin /CT = w/2 (6.11)
the currents flowing through them will be
βn βn
Ip ≈ (wDD VDD + wVp − VTn )2 In ≈ (wDD VDD + wVn − VTn )2
2 2
(6.12)
βn w 2
Ic ≈ wDD VDD + (Vp + Vn ) − VTn (6.13)
2 2
where parameters βn and VTn have their usual meanings, and second-order effects such
as the parasitics to the FG, the possible residual charge trapped in the interface oxide-
silicon and the non-zero output conductance, have been neglected. If all the constant
VDD
Ic
wDD
VDD M3⬘ VDD
Ip w/2
In
wDD wDD
M1⬘ w/2 M2⬘
w/2 w/2
Vp Vn
w/2 (a) w/2
Vp In
NG Ip
Vn
Ic
(b)
Figure 6.3 (a) FGMOS circuit for variable change. (b) Symbol
terms in the previous equations are grouped together, a new effective threshold voltage
can be defined and is given by
VTFG = VTn − wDD VDD (6.14)
Equation (6.14) shows a reduced value of the threshold voltage obtained by con-
necting one of the inputs to the maximum voltage available in the circuit (VDD ).
This allows the operation of the transistor in the strong inversion region even under
a tight constraint of low supply voltage. Taking the difference between Ip and In , it
can be shown that the relationship between the differential voltage and current has
the desired general form of eq. (6.7):
$ %
Vp − Vn = (Ip − In )/(w 2βn Ic ) (6.15)
where Ic is defined as the common mode current derived from Vp and Vn . Ic senses the
common mode voltage, Vcm = (Vp + Vn )/2, of both voltage signals. Equation (6.15)
represents a nonlinear relationship between voltages and currents. From now on,
we will refer to the circuit that realises eq. (6.15) as the nonlinear transconduc-
tor (NG-circuit), and its symbol is shown in Fig. 6.3(b). Each NG-circuit has a
transconductance given by

Gm(NG) = w 2βn Ic (6.16)
√
6.2.2.2 The y/ x circuit
Now, the RHS of eq. (6.9) can be easily implemented adding/subtracting
√ currents in a
summing node. The circuit required by eq. (6.15) has to realise a y/ x-type operator,
which can be obtained using the TP previously depicted for the FGMOS in weak
inversion saturation. Each RHS term in eq. (6.9) can be rewritten in a more general
form as

αId /(K1 Ic ) (6.17)

ηIid /(K1 Iic ) (6.18)
√
where K1 = w 2βn , Id = Ip − In and Iid = Iip − Iin . Multiplying both sides in
eq. (6.9) by a capacitance C will give a dimensionally correct equality if
Cα Cα Cη Cη
= √ = Aj IAj = √ = Ai IAi (6.19)
K1 w 2βn K1 w 2βn
where Aj and Ai are non-dimensional factors and IAj and IAi are independent currents.
An alternative form of (6.10) is thus obtained:
Id Iid
C(V̇p − V̇n ) = −Aj IAj √ + Ai IAi √ (6.20)
Ic Iic
Two currents are required to implement eq. (6.20), one for the α coefficient
and another for η. A circuit whose output has the general form of one of the RHS
terms in eq. (6.20) is shown in Fig. 6.4(a). In eq. (6.20), Id is any differential cur-
rent (as for example (Ip − In )) associated with a state variable, Ic is its associated
common mode current and IAj is an independent current source. The independent
voltage sources are used for shifting the threshold voltage and controlling the cir-
cuit gain. All the FGMOS transistors have to be working in the weak inversion
saturation region. The corresponding single output currents for the circuit are now
given by65

Io(n,p) = Aj IAj /Ic × I(n,p) (6.21)
where Aj depends on the constant reference voltage sources at the inputs of the
FGMOS transistors. Considering the same total capacitance and sizes for all the
FGMOS, the input capacitors are chosen to give the weights in Fig. 6.4, and all
the extra inputs are set to a value Vc given by an independent voltage source
(except VAj ):
Aj = e(VAj −Vc )/4nUT (6.22)
65 Subscripts (n,p) in I
o(n,p) have been used to refer to either Ion or Iop , and the same applies for I(n,p) .
This will apply to other variables from now on.
In
0.25 I on
Vc VAj 0.25
Ic I Aj
M1 M4
0.5 0.5
0.25 0.25
Vc Vc
0.25 0.25
M2 0.5 M3
0.5 Vc
Ip Vc
0.25 I op
0.25 0.25
Vc 0.25
M6
M5 0.5
Vc
0.5
0.25
0.25
(a)
In I on
Ip y-
------
x I op
Ic IAj,VAj
(b)
√
Figure 6.4 (a) y/ x circuit with FGMOS-based TL loops. (b) Symbol
Vc is chosen to ensure weak inversion operation. VAj and IAj are used for tuning.
Hence

α = (w IAj 2βn )e(VAj −Vc )/4nUT /C
(6.23)
η = (w IAi 2βn ) · e(VAi −Vc )/4nUT /C
Negative values of Io(n,p) can be easily implemented with current mirrors. The
symbol for the circuit in Fig. 6.4(a) is shown in Fig. 6.4(b).
6.2.2.3 The transconductor

The circuit that implements each term in the space state equation is shown in Fig. 6.5.
From now on it will be called FG-block. The transconductance of this cell is given by

Gm = C/τ = wAj 2ke βn IAj (6.24)
where ke is a constant that scales the magnitude of the currents coming from the
NG-blocks. It is set by the ratio between the sizes of the output and input transistors
Vp Ion
ke I n Vp
1 Ion
keIp ------ FG
NG x (Aj)
Iop
keIc IAj ,VAj Iop Vn
Vn
Figure 6.5 The FGMOS transconductor (FG)
in a pMOS current mirror that provides the input to the square-root block. The fact
that this is a large signal transconductor makes it very useful in the context of LV
designs, since having a linearised structure increases the input range for a given value
of distortion, which is one of the critical factors as the power supply voltage is scaled
down.
6.2.2.4 The state space circuit

The final realisation of the circuit that implements eq. (6.6) is shown in Fig. 6.6(c).
Two equal G-blocks have been used, one for each voltage-node: the state variable x
and the input voltage xi . The G-block drawn in Fig. 6.6(a) and (b) is just an FG-block
with an added cascode output stage that provides the differential currents. It can be
shown that the use of cascode FGMOS transistors improves the output resistance
(with all the consequences that this fact has on the frequency response of a higher-
order filter designed with it). Also, the output swing is enlarged and the common
mode feedback circuitry (required due to the fully differential structure) is simplified.
One of the inputs to each cascode transistor is connected to the output of the common
mode feedback block (CMFB) shown in Fig. 6.7. This CMFB is identical to the one
described in Chapter 4, so the reader is referred to this chapter for more information
about it.
6.2.3 Second-order effects

The currents injected into the integrating capacitors could deviate from their ideal
behaviour mainly due to the following reasons: (a) mismatch between the input
capacitances, and the effects of the parasitics (CGD ); (b) mismatch between the
transistors (βn and VTn ); (c) output resistances (degradation of the square and expo-
nential laws) and (d) dependence of the β parameter on gate voltage (θ). This
section focuses on the analysis of how these effects can affect the performance of the
transconductor.
6.2.3.1 Offset and nonlinearities

1. Mismatch between the input capacitances and effect of the gate to drain cou-
pling in diode-connected transistors. If mismatch between the input capacitances is
taken into account the currents given by eqs. (6.12) and (6.13) have to be rewritten as

2
βn Vd
I(n,p) = w(1 + (2,1) ) Vcm ± − VTFG(n,p) (6.25)
2 2
Iop Ion
Voutcm Iop-Ion Voutcm

gnd gnd
Ion-Iop
VDD VDD
Ion Iop
FG
Vp Vn
G
(a)
Vp
Vp Vip
-Id = Ion-Iop Ion - Iop
C
G G
(Aj) G
(Ai) I (Aj)
Id = I op-Ion Iop - Ion
Vn Vin
Vn
(b) (c)
Figure 6.6 Differential transconductor. (a) G-block. (b) Symbol. (c) Implementation
of eq. (6.6)
βn & w w '2
Ic = (1 + 31 )Vp + (1 + 32 )Vn − VTFGc (6.26)
2 2 2
where the i parameters represent the percentage of mismatch, and Vd is the differ-
ence between Vp and Vn . These expressions also take into account the deviations of
the effective threshold voltages with respect to their nominal values. Now
VTFG(n,p) = VTn(n,p) − wDD(2,1) VDD VTFGc = VTn − wDD VDD (6.27)
In the square-root block, the influence of mismatch is demonstrated through the
exponents of the currents in the nonlinear functions since they differ from the expected
ideal values. However, there is another source of variation in the exponents with
respect to their ideal values: the gate-to-drain parasitic coupling in those transistors
VDD
Mf 3 Mf 4
Voutcm
Vp VDD
Vn Mf1 Mf 2
VDD gnd
Vbias3
which have one of their inputs connected to drain. Both sources of error can be
studied together since the effect they have in the final response is qualitatively the
same. The general expression for the weight w, when all these deviations are taken into
account is

CGDi
wRi = w(1 + ci ) = w 1 + i + (6.28)
wCTi
√
And the real function implemented by the 1/ x circuit is given by
1/(1+c1(n,p) ) 0.5/(1+c1(n,p) )(1+c2 ) 0.5/(1+c3 )

I(n,p) Is IAj
Io(n,p) ≈ Aj(n,p) Is
Is Ic Is
1+x
In,p 1(n,p) x
2(n,p)
= Aj(n,p) √ · Ic (6.29)
Ic
2. Mismatch between the transistors. Mismatch between the transistors will

cause a different multiplicative constant for the currents which can be included in
the value of Aj(n,p) .
3. Output resistance. The main contribution to degradations in the output resis-
tance in the FGMOS is the capacitive coupling between the drain and the gate. For
the diode connected transistors, it has been analysed previously in eq. (6.28). In the
case of output transistors, their output resistances have been improved by cascoding,
hence their influence will be negligible.
4. Dependence of the β parameter on the gate voltage. The dependence of the β
parameter on the voltage at the FG of the transistors generating In and Ip can be taken
into account by using the following modelling equation:
" #
βn(n,p) βn(n,p) (1 + (2,1) )
βn(n,p) ≈ ≈ 1 ∓ θw Vd
I + θ (VGS(n,p) − VTn ) (1 + k2(n,p) ) 2(1 + k2(n,p) )
(6.30)
k2(n,p) = θ[w(1 + (2,1) )Vcm − VTFG(n,p) ] (6.31)
Ic can be considered constant since the dependence with the differential input is
weak (it is scaled by the percentage of mismatch, and also x2 is very small compared
x x x
with 0.5). Therefore, as a first approximation it will be assumed that Ic 2 = Ic 2n = Ic 2p .
Also, the β of the transistor generating Ic will be referred as βnc .
Taking all the previous considerations into account, and after a few laborious
derivations, the following conclusions can be drawn:
1. Mismatch causes an offset, whose value can be calculated using the previously
derived expressions, and is approximately
x
ke βnn ke βnp Ic 2
offset ≈ a0n − a0p ε−1 + HD2 (6.32)
2(1 + k2n ) 2(1 + k2p ) ke βnc /2
The magnitude of this offset depends upon the common mode of the signal
through ε. The variables that have not been defined previously are given in
Table 6.1.
2. There is variation in the gain of the block. The new gain is
x x
−1 AIc 2 ke a1n βnn − ke a1p βnp −1 AIc 2
HD1 ≈ ε −ε
ke βnc /2 2 ke βnc /2
x2
θwa0n ke βnn AIc θ wa0p ke βnp
× (1 + 2 ) − ε −1 (1 + 1 )
4(1 + k2n ) ke βnc /2 4(1 + k2p )
x
w(31 − 32 ) AIc 2 a0n ke βnn − a0p ke βnp
− + 3HD3
4ε2 ke βnc /2 2
(6.33)
where A represents the amplitude of a sinusoidal input signal.

Table 6.1 Definition of the parameters used for the calculation of second-order
effects

32 + 31
ε w 1+ Vcm − VTFGc
2
" #
x 21 x 3
A0(n,p) = ao(n,p) k1(n,p) 1 + x1 ln k1(n,p) + (ln k1(n,p) )2 + 1 (ln k1(n,p) )3 + · · ·
2 6
x 2
A1(n,p) 1 + x1 + x1 ln k1(n,p) + 1 ln k1(n,p)
2

x1 ln k1(n,p)
× 2 + ln k1(n,p) 1 + x1 +
3
a1(n,p) A1(n,p) b1(n,p)
x1 x 21 x 31
A2(n,p) + (ln k1(n,p) + 1) + ln k1(n,p) (ln k1(n,p) + 2)
2k1(n,p) 2k1(n,p) 4k1(n,p)
a2(n,p)
A1(n,p) B(n,p) + A2(n,p) b21(n,p)
⎛ ⎞
x1
2
x 21 x 31 ⎝1 −
ln k1(n,p)
⎠
A3(n,p) − 2
− 2
ln k1(n,p) + 2
6k1(n,p) 6k1(n,p) 6k1(n,p) 2
A2(n,p)
a3(n,p) b1(n,p) B(n,p) + A3(n,p) b31(n,p)
2
[((c1(n,p) (1+c2 )+0.5)/(I +c1(n,p) ))−0.5/(1+c3 )]
Aj(n,p) Aj(n,p) Is

w2 [1 + (32 + 31 )/2]2 2
B(n,p) Aj(n,p) E(n,p)
4
b1(n,p)
2C(n,p) ε + 2D(n,p)
& '
ci(n,p) i(n,p) + CGDi(n,p) / wCTi(n,p)

C(n,p) (∓Aj(n,p) )(E(n,p)
2 · w[1 + (32 + 31 )/2]/2)

D(n,p) (±[Aj(n,p) E(n,p)
2 (VTFG(n,p) ) · w(1 + (32 + 31 )/2)])/2
E(n,p) (1 + (2,1) )/[1 + (31 + 32 )/2]

= A(n,p) /Aj(n,p)

F(n,p) Aj(n,p) E(n,p)
2 (VTFG(n,p) )2

G(n,p) −Aj(n,p) E(n,p)
2 (VTFG(n,p) )
k1(n,p) A(n,p) ε 2 + 2G(n,p)
ε + F(n,p)

x1(n,p) (−c1(n,p) /(1 + c1(n,p) )

x2(n,p) 0.5(c1(n,p) + c2 + c1(n,p) c2 )/[(1 + c1(n,p) )(1 + c2 )]
VTFG(n,p)
VTFG(n,p) − VTFGc
E(n,p)
3. The second-order harmonic should be zero, since the structure is fully differential.
However, due to the mismatch, it is now
x
A2 Ic 2 w(31 − 32 )
HD2 ≈ −
2 ke βnc /2 4ε2
x
a1n ke βnn − a1p ke βnp A2 θ wIc 2 w(31 − 32 )
× +
2 2 ke βnc /2 4ε2

a0n ke βnn (1 + 2 ) a0p ke βnp (1 + 1 )
× +
4(1 + k2n ) 4(1 + k2p )
x
A2 Ic 2 a2n ke βnn − a2p ke βnp
+
2ε ke βnc /2 2

a1n ke βnn (1 + 2 ) a1p ke βnp (1 + 1 )
− θw + (6.34)
4(1 + k2n ) 4(1 + k2p )
4. The third-order harmonic, which is a source of intermodulation distortion, can
be expressed as follows:
x
A 3 Ic 2 −w(31 − 32 )
HD3 ≈
4 (ke βnc /2) 4ε 2
x
a2n ke βnn − a2p ke βnp A3 Ic 2 w(31 − 32 )
× +
2 4 ke βnc /2 4ε 2

ke βnn ke βnp
× θwa1n (1 + 2 ) + θ wa1p (1+1 )
4(1 + k2n ) 4(1 + k2p )
x x
ε −1 A3 Ic 2 a3n ke βnn −a3p ke βnp ε −1 A3 Ic 2
+ −
4 ke βnc /2 2 4 ke βnc /2

θ wa2n ke βnn θwa2p ke βnp
× (1 + 2 ) + (1 + 1 ) (6.35)
4(1 + k2n ) 4(1 + k2p )
All these analytical results are illustrated from Fig. 6.8 to Fig. 6.10 using the
design values of a filter design example given in Section 6.2.4. The total harmonic
distortion (THD) due to the variations of the exponents (but in absence of mismatch)
is shown in Fig. 6.8. The deviations in the exponent of the linear term in the NG-block
will be the main source of distortion. This is smaller than 1 per cent and it can be
THD (%)
0.8
0.4
0
0.08
1.2
0.04 0.8
0.4
x⬘1n = x⬘1p 0 0 Vcm (V)
Figure 6.8 Total Harmonic Distortion versus the common mode signal and the
variation in the exponents of In and Ip (eq. (6.29), A=0.25 V)
(a) HD2/HD1(%) (b) HD3/HD1(%)
4 0.04
3
2 0.02
1
0 0.00
0.08
0.10 0.08 0.10
0.04 0.08 0.08
0.04 0.04
x⬘2n 0 0 x⬘1n x⬘2n
0.04
0 0 x⬘1n
x⬘1p = x⬘2p = 0 x⬘1p = x⬘2p = 0.05
Figure 6.9 Second- and third-order harmonics when the exponents x2 (eq. (6.29))
and x1 (eq. (6.29)) change in a different form in the negative and positive
part of the NG-block
as low as 0.2 per cent, provided that the common-mode signal is under 1V (which,
on the other hand, has to be satisfied in order to prevent the output from saturation).
Figure 6.9 to Fig. 6.11 show the different harmonics that result when not only one of
the exponents differs from the ideal value, but also the deviation is different in the
positive and the negative parts of the circuit (subscripts p and n, respectively). The
second harmonic will dominate if mismatch exists between both parts. The figures
show that the main source of distortion comes from the changes in the x1 . Hence, the
filter should be designed in such a way that the parasitics are minimised, especially
in transistor M1, and also transistors M1 and M4 are well matched.
offset (A)
× 10 –10
4
0.08 0.10
0.04 0.08
0.04
x’2n 0 0 x⬘1n
Figure 6.10 Offset versus different values of the exponents x2 and x1 for Gm =
8e-10 A/V
(a) H D2/H D1(%) (b) H D3/H D1(%)
0.18
6
0.12
4
2 0.06
0 0
0.8 0.8
0.10 0.10
0.4 0.08 0.4 0.08
A(V) 0.04 x⬘1n A(V) 0.04
0 0 0 0 x⬘1n
Figure 6.11 Second- and third-order harmonics versus input amplitude and devia-
tion in the exponent of In when a 2 per cent mismatch is considered in
the other parameters
The offset when ke = 1 is shown in Fig. 6.10. Again, it is important to keep

under control the mismatch between the capacitors that affect the exponents that set
the state-space currents. Otherwise the offset could be a significant term. In any case,
ke will always be smaller than 1 and this will reduce its effect. All these figures have
been obtained for an input amplitude A = 0.25, a common mode current of 225 nA
and a common mode voltage of 0.625 V.
An example of the nonlinearities that arise when different input amplitudes are
applied is shown in Fig. 6.11. Again, the matching in the exponents of In and Ip is
the important factor to take into account. Hence, in general, the THD will be dom-
inated by the second harmonic, even when the structure is fully differential. This
term is mainly the consequence of the mismatch between the exponent of the sin-
gle output currents when they are processed by the NG-block, so the capacitances
involved in this ratio have to be designed carefully. Furthermore, the parasitics
should be minimised, or at least they should be as equal as possible in both single
sides.
6.2.3.2 PSRR+
This section analyses how the variations of the positive voltage supply can affect the
performance of the integrator block. This study is important since VDD is connected
to some FGMOS inputs (these connections were required either to scale up the values
of the currents (equation (6.22)) or to scale down the effective threshold voltage
(equation (6.14)). Taking mismatch into account, eq. (6.6) gives rise to two new
different equations for the positive and negative side:
V̇p − V̇n + α11 Vp − α12 Vn = η11 xip − η12 xin (6.36)
V̇p − V̇n + α21 Vp − α22 Vn = η21 xip − η22 xin (6.37)
Considering the case when the differential input signal is zero (this is valid as VDD
is taken as the input) the system of equations given by eq. (6.36) and eq. (6.37) can
be solved for the differential output as a function of the common mode at the input
to give66
(α21 − α22 )(η11 − η12 ) − (α11 − α12 )(η21 − η22 )

Vd = Vicm (6.38)
(α12 α21 − α11 α22 )
This equation can be rewritten in a different form considering how the sources of
mismatch are going to affect the different parameters. Hence, the difference between
some of them is going to appear as just a multiplicative factor, whereas for others
their difference is going to affect as an exponential function. Taking this into account,
new variables can be defined in the following way:
α21 = k1 α11 , α22 = k2 α12 , η21 = k3 η11 , η22 = k4 η12 (6.39)
Also, the α and η parameters depend on VDD as

/nU )V
(wDDj
(wDDj /nUT )VDD

α1j = α1j e T DD
, η1j = η1j e for j = [1, 2] (6.40)
66 The following derivations are only valid under the stated assumptions, and as long as all the devices
in the blocks are operating in the right operating region. If the mismatch is such that the CMFB stops
working or the cascode transistors in one of the branches leave the saturation region, these expressions
would not be valid anymore. Also, usual contributions to the PSRR+ have been neglected and only those
due to the FGMOS devices are considered. This is done for the sake of simplicity and also to give the
reader an idea of how much the FGMOS devices on their own would contribute to the degradation of the
performance.
(k1 α11 − k2 α12 )(η11 − η12 ) − (α11 − α12 )(k3 η11 − k4 η12 )
Vd = Vicm (6.41)
α12 α11 (k1 − k2 )
Assuming that the variations of VDD are much smaller than its nominal value
eq. (6.41) can be approximated as
Vd
VDD

(k1 − k3 )α11 η11 (wDD1
) + (k − k )α η (w
− wDD2 2

4 12 12 DD2 − wDD1 )
≈ Vicm
nUT α11 α12 (k1 − k2 ) VDD

(k4 − k1 )α12 η12 (wDD2
) + (k − k )α η (w
− wDD2 3

2 12 11 DD1 − wDD1 )
+ Vicm
nUT α11 α12 (k1 − k2 ) VDD
(6.42)
If the differential output/input gain is high enough in comparison with the percentage
of mismatch and so the latter can be neglected in its expression, the PSRR+ is then
given by
(nUT /Vicm )
PSRR+ = (6.43)
(wDD2 − wDD2 ) + ((k4− k3 )/(k1 − k2 ))(wDD1
)
− wDD2
The main conclusion that can be extracted from this expression is that the PSRR+
can get seriously degraded because of the exponential relationship between the
currents and the VDD connected inputs in the square root block. Assuming the hypo-
thetical case of the same percentage of mismatch of 1 per cent for all the devices,
and taking that all the wDD weights have the ideal value of 0.2 and Vicm = 0.5 V, the
PSRR+ would be only 25 dB. A way to improve this value is to reduce the capac-
itive weight connected to VDD or to connect that input to a low noise bias voltage
instead.
6.2.4 A second-order filter example

This section describes the implementation of a second-order filter prototype for
audio applications designed following the design procedure explained above. The
power supply voltage was aimed to remain below 1.25 V, for a circuit implemented
in a 0.8 μm CMOS technology with nominal values for the threshold voltages of
VTn = 0.8 V and |VTp | = 0.82 V. The starting point for the design is equations of
the form
ẋ1 = −ωo1 x1 − ωo2 x2 + ωo3 xi (6.44)
ẋ2 = ωo2 x1 (6.45)

In this case, three parameters, ωoi for i = [1, 3], determine the filter programma-
bility. The LP and BP filter transfer-functions are as follows:
ωo3 ωo2
X2 (s) = Xi (s) (6.46)
s2 + ωo1 s + ωo2 ωo2
sωo3
X1 (s) = Xi (s) (6.47)
s2 + ωo1 s + ωo2 ωo2
which are related to the filters specifications as
ωo2
ωo = ωo2 , Q= ,
ωo1
(6.48)
ωo3 ωo3
HLP (0) = , HBP (ωo ) =
ωo2 ωo1
A block diagram of a circuit that implements eqs. (6.44–6.45) is shown in Fig. 6.12,
where
x1 = (V1p − V1n ) (6.49)
x2 = (V2p − V2n ) (6.50)
Vip V1p
C
G G
(A3) Ic1 (A1)
Vin V1n
G
(A2)
V2p
C
G Ic2
(A2)
V2n
Figure 6.12 Second-order filter circuit with the proposed FGMOS transconductor
Figure 6.13 Photograph of the second-order filter
√
wA2 2βn IA2
ωo = (6.51)
C
!
IA2 A2
Q= × (6.52)
IA1 A1
!
IA3 A3
HLP (0) = × (6.53)
IA2 A2
!
IA3 A3
HBP (ωo ) = × (6.54)
IA1 A1
Tuning of either ωo or Q can be performed through either the reference voltages

or the independent current sources. The ωo frequency will remain constant if VA2 and
IA2 are constant. Q can be programmed either through IA1 or VA1 . The circuit outputs
are either the currents Ic1 and Ic2 through both capacitors or the differential voltage
across them.
The circuit in Fig. 6.12 is formed by two equal C capacitors and four transcon-
ductors that are designed according to the filter specifications. Because of the
low-frequency range chosen high output resistances are required in order to preserve
the low time constants of the filter. A value of 3.4 pF for the integrating capaci-
tances is chosen. For the NG-circuits βn = 30.7 μA/V2 ; for the M1–M3 FGMOS
transistors in Fig. 6.3. 50 fF and 100 √ fF input capacitors implement the 0.25 and
0.5 weights, respectively. For the y/ x circuits in Fig. 6.4 transistors were sized to
give β1 = 972 μA/V2 , β2 = β3 = 322 μA/V2 and β4 = 41 μA/V2 . These different

aspects add a multiplicative factor to the τ coefficients, which has been omitted in
the equations for the sake of simplicity. A photograph of the final circuit layout is
shown in Fig. 6.13. The total area is 0.23 mm2 .
Some experimental results that illustrate the performance of the filter are given
below. Figure 6.14(a) and (b) represent the LP and BP transfer functions.
The distortion for the LP function can be observed in Fig. 6.15 for a sinewave
input signal of 200 Hz in a filter with a cut-off frequency of 1 kHz. The maximum
value of distortion is −41 dB for an input amplitude of 0.5 VDD .
(a) 10
dB
Logmag
10
dB
/div
X2(f ) /Xi(f )
–90
dB 100 1k 10 k 100 k
(b)
10
dB
Logmag
10
dB
/div
X1(f ) /Xi(f )
–90 100 1k 10 k
dB 100 k
Figure 6.14 Experimental LP (a) and BP (b) filter

–40.0
–42.0 VA2= 1 V
VA1 = VA3 = 1.25 V
–44.0 f = 200 Hz
THD(dB)
fo = 1 kHz
–46.0
–48.0
–50.0
0.0 200.0 400.0 600.0
Vid (mV)
Figure 6.15 THD versus differential input amplitude in the LP filter
Frequency
199.9969 Hz
Level
100.0 mU
X2(f) 20 dB/div
Return:
RUN –12 dbU No Aug.
Figure 6.16 Distortion measured with the spectrum analyser HPSR770 for the filter
and input signal corresponding to Fig. 6.15
Figure 6.16 illustrates an example of THD measurement obtained with a spectrum

analyser. Figure 6.17 clearly shows a second-order harmonic that appears due to the
mismatch between both symmetrical branches.
Frequency
849.9908 Hz
Level
400 mU
X2( f ) 20 dB/div
RUN –12 dbU No Aug. Return:
Figure 6.17 Second-order harmonic in the low-pass filter with Q = 1, gain = 1

and fo = 2 kHz
The performance of the filter is also verified when the input common mode
changes. For this design the experimental CMR was 0.6 V–0.8 V. Figures 6.18 and
6.19 show the Q and gain programming for both transfer functions, respectively.
The noise floor is shown in Fig. 6.20. The cursor marks the output for a single
input amplitude of 0.1 mV, which is used as the reference.
Results of the performance are summarised in Table 6.2. They are in agreement
with the theoretical and simulated ones.
The design procedure for a low power continuous-time linearised Gm -C filter based
on the operation of the FGMOS transistor in the weak inversion region has been
presented in this chapter. The technique is specially suitable for very low frequency
applications in which dynamic range and power consumption are the most important
issues. The advantages the FGMOS devices add to this kind of topology are as
follows:
1. The operating point at the gate of the transistors can be shifted, this making
possible their operation in the desired region for effective input levels that on its
own would not meet the minimum requirements for that.
(a) 20
dB
LogMag
10
dB
/div
X2 (f )/Xi (f )
–80
–dB 100 1k 10 k 100 k
(b) 15
dB
LogMag
10
dB
/div
X1 (f )/Xi (f )
–80 100 1k 10 k 100 k

–dB
Figure 6.18 Q-programming for LP (a) and BP (b) functions for fo = 0.75 kHz
2. Translinear loops can be implemented without using the source terminal. Because
of this only two transistors need to be stacked. This is very important when low
voltage is a design constraint.
3. Less devices are needed to get a nonlinear ratio between currents, thanks to the
different possible weights in the transistor.
4. The functions can be programmed by using different inputs in the transistors.
5. It facilitates the common mode sensing and feedback mechanism.
(a) 20
dB
LogMag
10
X2 (f )/Xi(f ) dB
/div
–80
dB 100 1k 10 k 100 k
(b) 15
dB
LogMag
10
dB
X1 (f )/Xi(f )
/div
-85 100 1k
dB 10 k 100 k
Figure 6.19 Gain-programming for LP (a) and BP (b) functions for fo = 0.75 kHz
However, these advantages do not come for free. There are two main disadvan-
tages of using FGMOS instead of MOS for this type of circuit:
1. A larger area is needed because of the input capacitors.

2. Mismatch between them can cause undesired effects such as an increase in the
PSRR+ as well as THD.
Therefore, it is ultimately up to the designer to decide what weighs more, the

advantages or the disadvantages, in their specific design.
Frequency
74.9969 Hz
Level
0.1 mU
X2(f ) 20 dB/div
RUN –12 dbU No Aug. Return:
Figure 6.20 Noise floor. Cursor mark in an input amplitude of 0.1 mV
Table 6.2 Summary of the second-order filter

performance
Technology 0.8 μm-AMS CXQ

VTn , VTp 0.8 V, −0.82 V
VDD 1.25 V
Area 0.23 mm2
fo,min 100 Hz
fo,max 2 kHz
Qmin 0.7
Qmax 7
THDmax (VPP < 1 V@200 Hz, <−40 dB
Q = 1, HLP (0) = 1, fo = 900 Hz)
IM3 <−40 dB
(Vpp1 = Vpp2 < 0.5@200 Hz,
Q = 1, HLP (0) = 1, fo = 900 Hz)
Noise in band −75 dB
DR 78–62 dB
PSRR+ >40 dB
Power 2.5 μW
Notation
α Coefficient of the state variable in eq. (6.5)
αij α parameter with mismatch (eqs. (6.36) and (6.37))
αij Terms in αij that do not depend on VDD
βnc βn parameter for transistor M3 in NG-block (Fig. 6.3)
βnn βn parameter for transistor M2 in NG-block (Fig. 6.3)

βn(n,p) See eq. (6.30)
βnp βn parameter for transistor M1 in NG-block (Fig. 6.3)
3i for i = [1, 2] Percentage of deviation with respect to its ideal value of the
input capacitance connected to transistor Mi in transistor M3
(Fig. 6.3)
i Mismatch between diode connected capacitances in transistor
Mi (Fig. 6.4, eq. (6.28))
i for i = [1, 2] Mismatch between input capacitive weights, referring to
transistor Mi in NG-block (Fig. 6.3, eq. (6.25))
VTFG(n,p) See Table 6.1
ε See Table 6.1
η Coefficient of the input variable in the state-space
equation (6.5)
ηij η parameter with mismatch (eqs. (6.36) and (6.37))
ηij Terms in ηij that do not depend on VDD
τ Time constant Gm /C
ωo Cut-off frequency (rad/s. Eq. (6.48))
ωoi for i = [1, 3] Coefficients in the state-space equations of the second-order
filter prototype
A0(n,p) See Table 6.1
ao(n,p) See Table 6.1
a1(n,p) See Table 6.1
Ai Non-dimensional gain (eq. (6.19)) related to the input term in
the state-space equation
Aj Non-dimensional gain (eq. (6.19)) related to the state variable
term
Aj(n,p) Aj(n,p) parameter when there is mismatch that affects
multiplicative constants in the expression of the currents
(eq. (6.29) and Table 6.1)
b1(n,p) See Table 6.1

B(n,p) See Table 6.1
ci (i + CGDi /wCTi ). Contribution of mismatch as well as gate
to drain parasitic capacitance to the effective input weight in
transistor Mi (eq. (6.28))
ci(n,p) See Table 6.1

C(n,p) See Table 6.1

D(n,p) See Table 6.1
E(n,p) See Table 6.1
fo Cut-off frequency (Hz)
F(n,p) See Table 6.1

G(n,p) See Table 6.1
Gm(NG) Transconductance of NG-block. See Fig. 6.3
HBP(ωo ) Bandpass filter gain at the cut-off frequency (eq. (6.48))
HLP (0) Lowpass filter DC gain (eq. (6.48))
I0 Drain current in M0 (Fig. 6.1)
I1 Drain current in M1 (Fig. 6.1)
IAi Independent current in the input square root block (Fig. 6.4)
IAj Independent current in the state-space square root block
(Fig. 6.4)
Ic Common mode current (eq. (6.7))
Ic1 Current through one of the integrating capacitors in Fig. 6.12
Ic2 Current through one of the integrating capacitors in Fig. 6.12
Id Differential current Ip − In
Iic Common mode input current (eq. (6.8))
Iid Differential input current Iip − Iin
Iin Current related to Vin through eq. (6.8)
Iip Current related to Vip through eq. (6.8)
In Current related to Vn through eq. (6.7)
Ip Current related to Vp through eq. (6.7)
Ion Output current implementing the square root function for In
(Fig. 6.4)
Iop Output current implementing the square root function for Ip
(Fig. 6.4)
K1 Gain coefficient
√ in the change of variables in eq. (6.7) and
(6.8). w 2β n
k1(n,p) See Table 6.1
k2 θ [w(1 + (3,4) )Vcm − VTFG(n,p) ]
ke Scaling factor for the currents from the NG-block (from the
gain of a current mirror)
ki for i = [1, 4] Parameters to account for mismatch in αij and ηij that only
affects as a multiplicative constant (eq. (6.39))
M Transistor M1 number of inputs in Fig. 6.1
N Transistor M0 number of inputs in Fig. 6.1
(n, p) Subscript used to simultaneously refer to variables with

subscripts n and p
Q Quality factor (eq. (6.48))
V0i FGMOS inputs (transistor M0 in Fig. 6.1)
V1j FGMOS inputs (transistor M1 in Fig. 6.1)
V1n Single state variable (voltage) related to x1n in eq. (6.49)
V1p Single state variable (voltage) related to x1p in eq. (6.49)
V2n Single state variable (voltage) related to x2n in eq. (6.50)
V2p Single state variable (voltage) related to x2p in eq. (6.50)
VAj Tuning voltage for the square root block (Fig. 6.4)
Vc Constant voltage in one of the inputs of some FGMOS
transistors in the square root block shown in Fig. 6.4
Vcm Common mode voltage (Vp + Vn )/2 (Fig. 6.3)
vDD Variation of VDD
Vd Vp − Vn
Vi for i = [1, N ] Effective inputs in an FGMOS in weak inversion (eq. (6.2))
Vid Vip − Vin
Vicm Common mode at the input
Vin Single input in the state space-equation (voltage) (eq. (6.8))
Vip Single input in the state space-equation (voltage) (eq. (6.8))
Vn Single state variable (voltage) (eq. (6.7))
Vout Differential output voltage
Voutcm Output of the CMFB
Vp Single state variable (voltage) (eq. (6.7))
VTFG Effective threshold voltage of transistors in Fig. 6.3
(eq. (6.14))
VTFGc VTFG . Effective threshold voltage with mismatch for transistor
M 3 in the NG-block (Fig. 6.3). Taken as nominal value in the
calculations (eq. (6.27))
VTFG(n,p) Effective threshold voltage with mismatch for transistors M 2
and M 1, respectively, in the NG-block (Fig. 6.3, eq. (6.27))
VTn Threshold voltage for n-channel MOS devices (for the sake of
simplicity it is assumed the same for all of them unless the
opposite is said)
VTp Threshold voltage for p-channel MOS devices (for the sake of
simplicity it is assumed the same for all of them unless the
opposite is said)
w See win
w0k Effective input weight for the input k in transistor M0
(Fig. 6.1)
w1k Effective input weight for the input k in transistor M1
(Fig. 6.1)
wDD wVDD = CVDD /CT . Weight of input connected to VDD in
transistors in the NG-block (Fig. 6.3)
wDD(2,1) Weights of inputs connected to VDD in transistors M2 and

M1 in NG-block when mismatch is considered (Fig. 6.3)

wDDj for j = [1, 2] Sum of all the weights that contribute to the VDD term in αij

wDDj for j = [1, 2] Sum of all the weights that contribute to the VDD term in nij
wi Ci /CT . FGMOS input weights (eq. (6.2))
win Cin /CT = w/2. Weight of inputs connected to Vp and Vn in
transistors in NG-block (Fig. 6.3)
WRi w(1 + ci ). Weight w when mismatch and gate to drain
parasitic capacitance are taken into account in transistor Mi.
x State variable in the state eq. (6.5)
xi Input variable in the state eq. (6.5)
x1 Parameter related to the percentage of mismatch as shown in
eq. (6.29) and Table 6.1
x2 Parameter related to the percentage of mismatch as shown in
eq. (6.29) and Table 6.1
xj for j = [1, 2] State variables in eq. (6.44) and eq. (6.45)
xin Single input for the fully differential implementation in
eq. (6.6)
xip Single input for the fully differential implementation in
eq. (6.6)
xn Single state variable for a fully differential implementation in
eq. (6.6)
xp Single state variable for a fully differential implementation in
eq. (6.6)
Chapter 7
Low power log-domain filtering based on
the FGMOS transistor
The filtering operation performed by a linear, time invariant system (LTI) is mathe-
matically described by a set of LTI equations as it has been shown several times in
previous chapters. These equations relate two magnitudes: the input and the output
of the system. However, they do not impose any mathematical constraint in internal
variables, which implies that internal variables can be related in any manner and not
necessarily in a linear way. This opens up a wider set of options in the internal pro-
cessing of the signals. Among the nonlinear functions that can be chosen to process
the internal variables, logarithmic and exponential functions have proven to be very
interesting, mostly because they link directly to physical models of transistor devices.
The concept of log-domain signal processing was originally proposed by Adams
[179] and rigorously formalised by Frey [182–184]. Frey proposed a nonlinear (expo-
nential) mapping of the state variables in the state-space description of any linear
transfer function that results in a set of nonlinear equations which can be inter-
preted according to the Kirchoff’s current law. This class of circuits is known as
Exponential State Space (ESS) circuits. Later on, Tsividis came up with a more
generalised approach which included all the topologies that are Externally Linear
Internally Nonlinear (ELIN) [21].
Log-domain filters can be found in many different applications due to the fact that
an exponential function models the current of both a bipolar and a MOS transistor
in the weak inversion region. The main difference between the two of them from
the application point of view is that the current levels are different in both devices.
Hence, BJTs can be used for high-frequency applications that require high current
and transconductance values [189–193,223,224], whereas MOS are more suitable for
low-frequency and low-power design [26,194,195,225]. Nevertheless, regardless of
the application both of them offer an extended dynamic range under reduced power
supply voltage as well as easy programmability [226].
Previous chapters have shown how the FGMOS in weak inversion also behave
in an exponential way, but simultaneously many more parameters control the char-
acteristics of the exponential function. Besides, as it was described in Chapter 3 a
much more complex nonlinear processing than a simply exponential function can
be performed within a single FGMOS device. This can be used advantageously on
the implementation of log domain, low power, low frequency circuits. This chapter
will explain how to design a log-domain filter using FGMOS transistors. It starts
with a review of the basic log-domain principle, followed by the description of basic
FGMOS blocks that implement the required functions. The blocks are afterwards put
together on a basic lossy integrator circuit which serves as the basis of a higher order
filter. As usual, the advantages and disadvantages of using FGMOS devices in these
kinds of topologies will be discussed along the text. Also, as these topologies are inter-
nally nonlinear any deviation with respect to the ideal behaviour in the transistors,
caused by for example mismatch, will give rise to strongly nonlinear terms that can
degrade the performance in terms of distortion. A qualitative and quantitative analysis
of second-order effects focused on the latter will follow. The chapter ends with a real
design example that illustrates the performance of a log-domain second-order filter
designed following the previously developed theory.
7.1 Log-domain integration with FGMOS
7.1.1 Basic principle

Let us start again with the general equation of a linear integrator:
ẋ = αx + ηxin (7.1)
In the general formulation of logarithmic filters [182], both the state variable, x, and
the input, xin , are mapped into two new variables, y and u, through the exponential
relationships:
x = ky e y
(7.2)
xin = kin eu
Hence eq. (7.1) can be rewritten as a nonlinear function of u and y:
ẏ = α + Ke(u−y) (7.3)
ηkin
K= (7.4)
ky
If the terms in eq. (7.3) are associated to real physical variables, the left-hand side
could represent the current flowing through a capacitor, in which case y would be
proportional to the voltage difference between its terminals. The state variable, x,
could be the current flowing through a device with an ‘exponential behaviour’, as for
example a BJT transistor biased in the active region or an MOS transistor in the weak
inversion saturation region (these two would be the traditional choices). Another
Low power log-domain filtering based on the FGMOS transistor 217
Expander block
Ix
Vx
VB
C
Figure 7.1 State variable definition: expander block
option is to use the FGMOS transistor also biased in the weak inversion saturation
region, as it is shown in the following section.
7.1.2 Basic FGMOS circuits

This section describes basic circuit blocks that realise the mathematical transforma-
tions described before.
Figure 7.1 shows the schematic of a circuit block that generates the state variable,
which in this case is the current Ix . The latter is the current in the channel of an FGMOS
transistor with one of its inputs connected to a capacitor. A second input is connected
to a constant voltage VB . The mathematical function describing this current, assuming
that the transistor is operating in the weak inversion saturation region and also that
the ratio between the parasitic capacitances and the total capacitance observed by the
FG is negligible67 , is
Ix = Is ewB VB /nUT ewx Vx /nUT = ky ewx Vx /nUT = ky ey (7.5)
where wB and wx are the weights for the corresponding inputs. Hence, y is pro-
portional to the voltage across the capacitor C and ky depends exponentially on the
voltage VB . As in other logarithmic implementations the current ‘expands’ the input
voltage because of the exponential function relating the two of them. Therefore, this
transistor will be called expander from now on.
The input variable will also be a current; therefore, it will have to generate a
voltage that is related to it in a logarithmic way (as shown in eq. (7.2)). Consequently
if the argument of the logarithmic function is larger than 1 then big variations in
this current will translate into smaller variations of the voltage. This voltage can
be generated by an FGMOS transistor with one of its inputs connected to its drain.
Because of the aforementioned ‘compression’ function, this device will be referred
to as compressor.
67 This will be assumed from now on unless the opposite is said.

I in
Vin
VBin
Figure 7.2 Generation of the input voltage from the input current
The circuit block is represented in Fig. 7.2. The current Iin can be written as a
function of the voltages, yielding68
Iin = Is ewBin VBin /nUT ewin Vin /nUT = kin ewin Vin /nUT = kin eu (7.6)
And Vin will be related to the input current in a logarithmic way as it was explained
before:

nUT Iin wBin
Vin = ln − VBin (7.7)
win Is win
where win and wBin are the weights corresponding to the inputs Vin and VBin ,
respectively.
Equation (7.3) can now be rewritten considering the already defined physical
variables as shown in eq. (7.8). The RHS is the sum of two terms, a constant current
plus a term that is proportional to the ratio between the input current and the state
current. These currents are added into the integrating capacitance:
wx V̇x Iin
=α+η (7.8)
nUT Ix
Therefore, the following step would be to generate the ratio between the input and the
state current, which can be done with the help of the translinear principle explained
in the previous chapter. The schematic of the divider block is shown in Fig. 7.3. It
works as follows: both transistors are equally sized and have the same value input
capacitances. Assuming that all the previously defined weights are also the same, w,
the output current Iout1 is
Iin w(VB −VBin )/nUT
Iout1 = IA ew(Vin −Vx )/nUT = IA e (7.9)
Ix
which is proportional to the ratio between the input and the state-space current.
68 Both transistors have been assumed to be of the same size. This will be assumed from now on for
all the devices unless the opposite is said.
IA
Iout1
VA
M2 M3
Vx Vin
TL
Figure 7.3 Nonlinear block generating the ratio between the currents of the
compressor and expander (NLB)
Vx Ix=Iout
Compressor
Iin
Expander
NLB
Vin C
NLI VB
VBin
Figure 7.4 Logarithmic integrator block diagram
7.1.3 The integrator

All the previous blocks can be put together to form the lossy integrator described by
eq. (7.1). A block diagram that illustrates the interconnections between them is shown
in Fig. 7.4. The input block is the compressor. This circuit takes the input current and
converts it logarithmically to the voltage Vin . The second block performs a nonlinear
integration (NLI) as in (7.8). This block is formed by the nonlinear block (NLB),
whose output is the ratio between the state and input currents, and the integrating
capacitor C. The output signal is finally obtained taking the voltage at the integration
node and processing it through the expander. The schematic for the final circuit is
shown in Fig. 7.5.
In reality, this block is implemented by sourcing instead of sinking current to
the integrating capacitance (for the sake of simplicity). This is done just by using a
conventional current mirror to copy and invert the sign of the current delivered by M3
and connecting its output together with an inverted version of IB to the integrating
capacitance.
The expressions of the constants α and η can be obtained from Fig. 7.5 just by
writing the equation of the voltage Vx in the integrating capacitance, rearranging it
Compressor Iin
Nonlinear Integrator
IA (NLI) Vin
w11
w22 VA VBin M1
M2
w24
IB Vx Expander Ix = Iout
w43
M4
w31 VB
C
M3
w32
Figure 7.5 Integrator schematic69
to obtain the output current and comparing this expression with eq. (7.1):
IB IA Iin
V̇x = − ew(VB −VBin )/nUT × (7.10)
C C Ix
w IB w 1A w(VB −VBin )/nUT
İx = × Ix − × e Iin (7.11)
nU C nU C
/ T01 2 / T 01 2
α η
In order to improve the dynamic range of the integrator, a differential version of

it might be required. The differential version can be obtained taking as starting point
a differential form of the state-space equations
(ẋp − ẋn ) = α(xp − xn ) + η(xinp − xinn ) (7.12)
where the subscripts p and n have been used to differentiate between the positive and
negative parts of the differential signal. For a lossless integrator, eq. (7.12) becomes
(ẋp − ẋn ) = η(xinp − xinn ) (7.13)
In reality eq. (7.13) can be obtained by subtracting two single equations, one for the
positive and another one for the negative side, respectively. Each one of the single
equations can be realised using the procedure explained before. The subtraction can be
performed just by inverting the currents with current mirrors. A schematic of a circuit
that realises eq. (7.13) is shown in Fig. 7.6(a). Its block diagram is in Fig. 7.6(b).
The + and − signs in the NLB have been used to differentiate between positive and
negative currents. The differential version of the compressor is formed by transistors
M1p and M1n . The expander is constituted by transistors M4p and M4n .
69 Parameters w are related to mismatch and are explained in a further section.

ij
IA IA
VAp VAn
M2 p M2 n
C C
M3p M5p M5n M3n
IDC IDC
I inp Vinp Vinn I inn
Ixp = I outp Ixn = I outn
Vxp Vxn
M1p VBin M1n
VBin M4p M4n
VB VB
(a)
Compressor
I inp +– Exp ander Ixp = I outp

Vinp NLB Vxp
I inn Ixn = I outn
+ –
Vinn Vxn
VBin VB
(b)
Figure 7.6 (a) Schematic of the fully differential integrator. (b) Block diagram
7.1.4 Advantages and disadvantages of using FGMOS transistors

The schematic of the integrator in Fig. 7.6 outlines some of the advantages of using
FGMOS transistors:
1. First of all, neither the source nor the bulk terminal is required to implement
the translinear equation. Because of this, it is possible to fix them to ground
(or the negative power supply) as well. Therefore, there is no need for stacking
transistors to implement the translinear loop, which is important when dealing
with tight voltage and power constraints.
2. Besides, as the source and bulk terminals are connected to constant voltage values,
VSB turns out to be constant as well (in this case equal to zero) and separate wells
do not have to be used anymore to reduce distortion70 . It has been shown that
sometimes these wells can cause oscillations in the output that are not predictable
70 The gate to source voltage as a function of the current is also a function of V , which causes
SB
distortion unless these terms are cancelled when implementing the translinear loop.
by conventional simulations, and even if they were special circuit strategies are
required to get rid of them [29]. This instability is due to the leakage currents
in the wells. The problem does not exist in FGMOS topologies, which do not
require and therefore do not use wells to implement the translinear function.
Shortcomings of using FGMOS are as follows:
1. An FGMOS device requires larger area when compared with the normal MOS
transistor, although the difference is up to a certain extent compensated because
of the elimination of the wells mentioned before. The extra cost in terms of area
is difficult to estimate. It will depend on how big the transistors are since the
input capacitances have to be designed according to that as it was explained in
Chapter 2. Other factors that will determine the area are the number of inputs as
well as the matching between devices. A better matching requires larger transis-
tors and larger transistors need larger input capacitances to minimise the effects
of the parasitics.
2. The second shortcoming is related to the level of compression of the voltage
signals at the internal nodes. The compression is not as powerful as in all MOS
loops, since now the logarithmic relationship between the voltage and the current
signals is divided by the value of the weight which is always smaller than 1.
7.2 The second-order filter
This section explains how to design a higher-order filter using the previous integrator.
The filter chosen for the example is a second-order lowpass/bandpass filter, which in
the time domain is described by the state-space eqs. (7.14) and (7.15):
(İ1p − İ1n ) = −ωo1 (I1p − I1n ) − ωo2 (I3p − I3n ) + ωo3 (Iinp − Iinn ) (7.14)
(İ3p − İ3n ) = ωo4 (I1p − I1n ) (7.15)
Equations (7.14) and (7.15) can be transformed applying changes of variables as

in eq. (7.2). After doing so, the resulting expressions are71 as follows:
w −1 −1 −1 −1
(V̇1p − V̇1n ) = −ωo1 − ωo2 (I3p I1p − I3n I1n ) + ωo3 (Iinp I1p − Iinn I1n )
nUT
(7.16)
w −1 −1
(V̇3p − V̇3n ) = ωo4 (I1p I3p − I1n I3n ) (7.17)
nUT
A physical realisation of eqs (7.16) and (7.17) is shown in Fig. 7.7. This schematic
has been a bit simplified for the sake of clarity. Hence, for example the values of ωoi
71 The changes of variables have been carried out in the single equations first and subsequently these
single equations have been subtracted.
Compressor
Expander
I inp +
+ I1p
Vinp NLB V1p
I inn I1n
–
–
Vinn V1n
VBin VB1
+– +
–
NLB NLB Vtun1
+ +
– –
Expander
+
+ I3p
NLB V3p
I3n
––
V3n
VB2
Figure 7.7 Block diagram of the second-order filter
have been designed to be the same for i = [2, 3]:

w
ωo1 = × IA
nUT C
w
ωoi = × IA ew(VB1 −VBin )/nUT i = [2, 3] (7.18)
nUT C
w
ωo4 = × IA ew(VB2 −VBin )/nUT
nUT C
The way to make ωo2 different from ωo3 is either to change the current IA for each
NLB, or as it is done in the design example at the end of the chapter, to add an extra
input to the output transistor in the NLB (M3 in Fig. 7.6) and connect it to a tuning
voltage Vtuni . If this is the choice, but still all the FGMOS transistors are wanted to
be identical for the sake of matching, the extra input can be added to all of them. It
can be connected to a constant voltage Vcm in those devices in which this input is not
required. Under these circumstances, the general output of an NLB with an applied
tuning voltage Vtuni is
Iout1i = IA ew(Vin −Vx )/nUT ew(Vtuni −Vcm )/nUT
Iin w(VB −VBin )/nUT w(Vtuni −Vcm )/nUT
= IA e e (7.19)
Ix
Also, in the final design another input can be added in order to compensate for
common mode variations. This input will be connected to a voltage generated by a
CMFB. The output of the CMFB only has to be connected to those FGMOS transistors
whose drain currents are either sinked from the integrating capacitance or mirrored
through a pMOS current mirror and sourced to it. Depending on whether it is a
transistor providing current that is going to be sinked or sourced it will be connected to
p n
either the negative or the positive output of the CMFB, Voutcm or Voutcm , respectively.
Hence, taking all this into account Fig. 7.6 would be modified by adding two extra
inputs to the transistors. In M1(n,p) and M2(n,p) , both of them would be connected to
p
a constant voltage Vcm . In transistors M5(n,p) they would be connected to Voutcm and
n
Vtuni , and in transistors M3(n,p) they would be connected to Voutcm and Vtuni . The new
coefficients of the state-space equations considering this are in eq. (7.21). Superscripts
p n
p and n have been used to differentiate between the coefficients with Voutcm or Voutcm ,
respectively. Ideally, they would be the same but this will be explained later on. Gex
is the gain of the expander block as given by eq. (7.20):
Iout(n,p) = Is Gex e3wVcm /nUT ewVx(n,p) /nUT (7.20)
The term corresponding to the common mode feedback in coefficient ωo4 depends on
(n,p) (n,p)
V outcm instead of Voutcm . This is because this voltage corresponds to the output of a
different CMFB (a different CMFB is required for each couple of integrating nodes):
wGex (n,p)
ωo1 = IA ewVtun1 /nUT e−2wVcm /nUT ewVoutcm /nUT
nUT C
wGex (n,p)
nUT C
(7.21)
wGex (n,p)
nUT C
wGex (n,p)
ωo4 = IA ewVtun4 /nUT e−2wVcm /nUT ewV outcm /nUT
nUT C
A further modification to the filter in Fig. 7.7 aimed to reduce power consists of
eliminating transistor M2 in the NLB realising the coefficient ωo1 and connecting the
two inputs in transistor M3 to Vtun1 . The new ωo1 coefficient is then given by
w (n,p)
ωo1 = Is e3wVtun1 /nUt ewVoutcm /nUT (7.22)
nUT C
7.2.1 The expander

The expander block for the filter in Fig. 7.7 is a modification of the simple basic
expander circuit in Fig. 7.1. The schematic for this circuit is shown in Fig. 7.8. One of
the inputs in each FGMOS is connected to either a positive or a negative integrating
node. The other three inputs are connected to a constant voltage, Ve , which can also
be used to adjust the signals levels. The output current is hence given by
Ioutj = Ijp − Ijn = Is e3wVe /nUT (ewVjp /nUT − ewVjn /nUT )
= Is Gex e3wVcm /nUT (ewVjp /nUT − ewVjn /nUT ) j = 1, 3 (7.23)

I jp
Vjp Expander I jn
I outj –I outj
Vjp Vjn Vjp Vjn
Ve Ve Ve I jp Ve I outj
Vjn I jn Ve
Ve Ve Ve –I outj
Ve Ve Ve Ve
Figure 7.8 Fully differential expander
Iinp Iinn
Vinp Vinn
Vcm Vcm
2IA
Figure 7.9 Input stage
7.2.2 The input stage

Although all the previously described blocks are operating in the current mode some-
times the input signal might be a voltage and a voltage to current conversion would be
required. A simple circuit that can be used with this aim and proved to work with this
filter (obviously under certain limitations) is shown in Fig. 7.9. Despite the transis-
tors being operating in the weak inversion region, which makes the differential pair
strongly nonlinear, the compression of the input voltage carried out at the FG by the
ratio between the effective input capacitance and the total capacitance enables these
to have relatively large signals at the input. However, it should be kept in mind that
the dynamic range remains the same as the equivalent input noise is amplified by the
inverse of the input weight.
An example of the performance of this block designed in a 0.35 μm technology
is shown in Fig. 7.10. The input transistors aspect ratio is 10 μm/10 μm and the
input capacitances have a value of 120 fF [217]. The effective input weight is 0.25.
Fig. 7.10(a) illustrates the compression at the FG. Fig. 7.10(b) shows the output
current. The THD is lower than 0.3 per cent for maximum signal levels of 200 mV.
7.3 The common mode control
It is a well-known fact that in order to avoid degradation in the frequency response

of a Gm -C filter the output resistance in parallel with the integrating capacitance has
(a) Transient response

700m
600m
VFG (V)
500m
400m
300m
(b)
0.00 –
–
–50.0n
Iin (A)
–100n
–
–150n
–500m –100m 300m 700m
(V)
Figure 7.10 (a) Compression at the FG in the input stage. (b) Output current
to remain high enough72 . But this is a fully differential structure with transistors
operating in the weak inversion region, which means that mismatch could seriously
affect the expected value of the output resistance. This is one of the reasons to have
a circuit to control the common mode at the output. A circuit that works well with
this filter is shown in Fig. 7.11. It is a p version of the CMFB described in Chapter 5.
The principle of operation is the same. The only two differences between the two
of them is that the bias current might be lower in this case and the transistors will
p
then operate in the weak inversion region and also both differential outputs Voutcm
n
and Voutcm are required in the feedback mechanism. These outputs are connected to
transistors that supply current to the integrating capacitance. Depending on whether
it is sinked or sourced current the connection should be to either one or the other. Any
difference between the common mode level and Vcm creates a difference between
the output voltages. This tends to generate different values for the currents provided
to the integrating capacitances by the n and p branches. The difference has to be
corrected in response to the Kirchoff’s current law. The output resistance controls
the change in the voltage which moves the operating point towards the reference
voltage.
7.4 Design considerations. Second-order effects
7.4.1 Multiple operating points in the filter

Having separate wells in log-domain CMOS filters can provoke instabilities and make
the output go into oscillation. This sometimes cannot be predicted by conventional
simulations during the design process. Fox detected the problem while testing some
circuits in the lab and proposed a method to detect this behaviour [203]. Log-domain
filters implemented with FGMOS do not need any extra well to realise the translinear
72 How high will depend on the characteristics of the filter.

2IA
V1p
Vcm
V1n Vcm
Vcm Vcm
V nocm V pocm
1
0.9
0.8
0.7
0.6
V3n(V)
0.5 Operating point

0.4
0.3
0.2
0.1
0.2 0.4 0.6 0.8 1

V1n(V)
Figure 7.12 Analysis of stability in the filter
equation, and therefore do not have this problem. This is illustrated in Fig. 7.12 for
the filter in Fig. 7.7 using Fox’s method. The integrating capacitances are substituted
by voltage sources which are swept in the range of interest. Subsequently, the currents
flowing through these sources are printed and processed with the CONTOUR function
in Matlab to find the crossing points in the integrating nodes when the currents are
equal to zero. The latter would determine the possible operating points. If more than
one exists (as it happens in normal MOS logarithmic filters) their stability would have
to be analysed. The graph in Fig. 7.12 shows a single operating point, thus proving
that the filter is inherently stable73 .
73 Although the graph represents only two of the single outputs the results for the others are equivalent,
showing also a single crossing point.
7.4.2 Effect of the parasitic capacitances: qualitative study

As it was explained in Chapter 2 one of the drawbacks of using FGMOS is the larger
area of the device when compared with a conventional MOS transistor, owing to
the input capacitors. The extra percentage of the total area can be made as small as
the topology and performance tradeoffs permit, since parameters such as THD, gain
and output resistance will degrade as the input capacitors are made smaller. This is
because of the effect of the parasitic capacitances on the circuit performance. The
latter is qualitatively analysed in this section.
Let us start with the gate to bulk capacitance CGB . CGB has two different compo-
nents, one of them depending on the transistor size (this is the same as in a normal
MOS transistor) and the other one due to the coupling between the poly1 bottom plate
(FG) and the bulk (or well, in case the outer part of the FG has a well underneath). The
first component can be controlled by changing the total area of the MOS transistor
used to build the FGMOS. As for the second one, the control the designer has on its
value is quite small since it is a fixed fraction of the total FG area. It is therefore a
technological constraint. Besides, the effect this capacitance has on the scaling factor
of the gate voltage depends upon the relationship defined by eq. (2.22), so there is
not much the designer can do to reduce it. The alternative is to take it into account.
Sometimes it can be a disadvantage since it increases the signal compression at the
gate but, in general as it is tied to a constant voltage (unless the back gate is used), it
will not be too problematic.
A different parasitic capacitance but with a similar effect is CGS . As the source
terminal does not need to be used in FGMOS logarithmic filters, this capacitance
contributes with another term added to the two previously discussed.
The most problematic parasitic capacitance is the gate to drain parasitic capac-
itance, CGD . Three different case scenarios can occur as shown in Fig. 7.13. The
schematic in Fig. 7.13(a) shows an FGMOS transistor with one of its inputs connected
to the drain. In this kind of topology, the equivalent weight to the diode connected
input is not Cin /CT , as it would be ideally assumed, but (Cin + CGD )/CT . Since the
translinear loop is designed to perform a nonlinear relationship between currents,
and the exponents in this function are obtained with ratios between capacitances, a
change either in the numerator or in the denominator of this ratio will alter the values
of the exponents and hence the function. This can be a source of distortion and will
be studied later on.
Figure 7.13(b) shows an FGMOS with a pMOS current mirror as a load. The output
resistance has now an additional component in parallel equal to (gm CGD /CT )−1 .
Keeping this value much smaller than the pMOS transistor transconductance its effect
will be negligible.
Figure 7.13(c) shows an NLB with an integrating capacitance at the output. If
CGD is assumed to be the same for all the FGMOS transistors, the voltage, VA , is
given by

nUT IA w
VA = ln − Vx (7.24)
wD + w Is wD + w
IA
Iout1
VA
Vx Vin
(a) (b) (c)
Figure 7.13 FGMOS topological arrangements in the log-domain integrator. Anal-

ysis of the output resistance effects
where wD = CGD /CT . Hence, the current Iout1 is

w/(w+wD )
IA
ewVin /nUT · e−w Vx /(w+wD )
2
Iout1 = Is (7.25)
Is
Equation (7.25) shows how both the gain of the function as well as the value
of the exponent are related to Ix change. This is going to translate on a variation
of the integrator time constant as well as distortion. Besides, the time constant will
depend on the technological parameter Is . The effects of CGD can be reduced either
by increasing the minimum input capacitance or by limiting the swing at the drain
node by using, for example, cascode transistors. The drawback of the latter is that
stacking transistors makes the minimum voltage supply constraint more demanding.
7.4.3 Effect of CGD and the mismatch between the inputs:

a quantitative study
The effects of mismatch between the input capacitances and CGD on the overall
transfer function can be analysed using the same general functional dependence. This
is because qualitatively they affect the performance of the block in the same way;
although, quantitatively they can be very different. What will happen in general (either
because of CGD or mismatch or both) is that the relative weights will change with
respect to their nominal values. This means that instead of getting the linear ratio
between the input current and the state variable current shown in eq. (7.10), the ratio
will be between non-integer powers of them. Taking this into account eq. (7.11) can
be rewritten as
I˙x = α Ix − η (Iin )1+x1 · Ix−x2 (7.26)
where α and η represent the modified expressions of α and η and if mismatch and
CGD are taken into account. Their exact equations are irrelevant at the moment, so
they will not be given for the sake of clarity.
As x1 1, a Taylor series expansion can be performed. Also if an input consisting
of a DC (IinDC ) plus a sinusoidal term (A sin ωt) is considered and the amplitude of
the varying term is as a first approximation much smaller than the DC term a second
series expansion can be carried out that yields for the exponential input:

x1 A sin ωt (A sin ωt)2 (A sin ωt)3

(Iin ) ≈ 1 + x1 (ln IinDc ) + − + + ···
IinDC 2(IinDC )2 3(IinDC )3
x12 x3
+ (ln IinDC )2 + 1 (ln IinDC )3 + · · · (7.27)
2 6
The same can be performed for Ix . In this case, as Ix is the integrated signal, as a
first approximation it can be written as a cosine function (Ix = IxDC + A1 cos(ωt)).
Then, according to eq. (7.6) the integrator equivalent input is now given by
(Iin )1+x1 · (Ix )−x2 (7.28)
It consists of different harmonics (apart from the fundamental one) which will generate
distortion. The expression for the resulting THD considering as dominant the third-
order harmonic is given by the following equation:

[(b0 b5 b6 + b1 b4 b6 − b1 b2 b8 )2 + (b0 b9 b2 − b4 b0 b7 − b1 b3 b7 )2 ]

× [16[(b b b + b b b + 3 (b b b + b b b − b b b )
0 3 6 1 2 6 0 5 6 1 4 6 1 2 8
THD ≈ 100
4
+ b1 b2 b8 )2 + (b0 b2 b7 + b4 b0 b7

3
+ (b0 b9 b2 − b4 b0 b7 − b1 b3 b7 ) + b1 b3 b7 )2 ]]−1
4
(7.29)
The values of the parameters are shown in Table 7.1.
Let us now analyse briefly which are the factors that could make x1 and x2 different
from zero. The names used for the different weights are shown in Fig. 7.5. Due to the
Table 7.1 Equations for the different parameters used

in the THD calculations
b0 IinDC
b1 A
x2 x13
b2 1 + x1 (ln IinDC ) + 1 (ln IinDC )2 + (ln IinDC )3 + · · ·
2 6
b3 (x1 A)/IinDC
b4 2
−x1 A2 /2Iin
DC
b5 3
(x1 A3 )/3Iin
DC
x2 x23
b6 1 − x2 (ln I1DC ) + 2 (ln I1DC )2 − (ln I1DC )3 + · · ·
2 6
b7 −x2 A1 /I1DC
b8 x2 A21 /(2I12 )
DC
b9 −x2 A31 /(3I13 )
DC
(M1)
gate to drain capacitance in transistor M1(CGD , where the superscript refers to the
name of the transistor) the input weight is
(M1)
(M1)

Cin + CGD Cin CGD
w11 = = 1+ (7.30)
CT CT Cin
Also, if the input capacitance is connected to Vin in M3, it differs from the ideal value
Cin /CT as:
Cin
w31 = (1 + 1 ) (7.31)
CT
Then the exponent of the input current is
* M1

CGD
(1 + 1 ) 1+ = 1 + x1 (7.32)
Cin
A similar analysis can be done for the state current, whereby

Cin
w32 = (1 + 2 ) (7.33)
CT
The general form of the voltage VA generated by the transistor M2 is

nUT IA w24
VA = ln − Vx (7.34)
w22 Is w22
where w24 and w22 are the relative weights from every input to the FG in transistor
M2 as shown in Fig. 7.5. w22 takes into account the contribution of the gate to drain
coupling capacitance. Hence, the current provided to the integrating capacitance is
then a function of Vx , and Ix :
(M2)
I ∝ e−(1+1 )/(1+CGD /Cin2 +3 )(Cin /CT )Vx
∝ (Ix )−((1+4 )/(1+CGD /Cin2 )(Cin /Cin4 ))
M2
(7.35)
where Cin2 and Cin4 are the unit capacitors in transistors M2 and M4, respectively, and
(M2)
w24 = Cin2 /CT , w22 = (Cin2 /CT )(1 + CGD /Cin2 + 3 ), w43 = Cin4 /CT and 4 =
2 − 1 . Therefore x2 would be
*" (M2)
#
Cin4 CGD
x2 = (1 + 4 ) 1+ (7.36)
Cin Cin2
Figure 7.14 and Fig. 7.15 show the THD versus different parameters74 , depending
upon the amplitude of the input signal for IinDC = IA = 50 nA. The main sources of
distortion are the gate to drain parasitic capacitance in transistor M2 and the mismatch
between the input capacitance in the transistor providing current to the integrating
74 Those that affected the most the value of the THD. It was studied for different ones, and also for
different variation ranges in xi , and these were found to be the ones that affected the THD the most.
THD(%)
0
0.1
0.08
0.06 2.5 3
2
0.04
1 1.5 ×10 –8
(M2) 0.02 0.5
C GD /CT 0 0 A (A)
Figure 7.14 THD versus the gate to drain capacitance in M2, and the input
amplitude
(a) (b)
THD(%) THD(%)
0.14 1.6
0.10 1.2
0.8
0.06
0.4
0.02
0 0
0.1 0.1
0.08 0.08 2
0.06 2 2.5 3 0.06 1.5
0.04 0.04
x2 0.02 1.5 ×10–8 x1 0.02
1 ×10–8
0 0.5 1 A(A) 0 0
0.5
0 A(A)
Figure 7.15 (a) THD versus the mismatch between the input capacitances affecting
Ix eq. (7.36), and the input amplitude. (b) THD versus the mismatch
between the input capacitances affecting Iin eq. (7.32), and the input
amplitude
capacitance, and the input transistor, Fig. 7.14 and Fig. 7.15(b), respectively. The first
source of distortion can be controlled by dimensioning the input capacitors in such
a way that the relative weight CGD /CT is small enough compared with Cin /CT . The
second source of THD is much more technology dependent, but for typical mismatch
values of 8 per cent and amplitudes below 20 nA it should be well below 1 per cent.
The previous analysis was performed under the assumption that both differential
branches were completely symmetrical. Should this be true the even order harmonics
would cancel out. However, in reality, mismatch is going to cause asymmetry and
this in its turn might generate even-order harmonics comparable in magnitude to
the odd-order ones. Mismatch can be included in the previous study by assuming
that x1 and x2 are different for the positive and negative branches: x1(n,p) , x2(n,p) .
Also, the mismatch affecting both blocks gains can be taken into account as different
coefficients normalising the amplitudes. Hence, all the bi parameters will be split into
two terms, bin and bip , and the expressions for the two dominant distortion harmonics,
second (HD2 ) and third one (HD3 ), are as follows:
HD2 = [(b1p b3p b6p − b0p b2p b8p + b0p b4p b6p )
− (b1n b3n b6n − b0n b2n b8n + b0n b4n b6n )]2

b1p b2p b9p + b1p b4p b7p + b0p b3p b9p + b0p b5p b7p
+ b2p b1p b7p +
2
b1n b2n b9n + b1n b4n b7n + b0n b3n b9n + b0n b5n b7n
− b2n b1n b7n −
2

2
+ b0p b3p b7p − b0n b3n b7n (7.37)
HD3 = ((b0p b5p b6p − b0n b5n b6n + b1p b4p b6p − b1n b4n b6n − b1p b2p b8p
+ b1p b2p b8p )/2)2 + ((b0p b9p b2p − b0n b9n b2n − b4p b0p b7p
+ b4n b0n b7n − b1p b3p b7p + b1n b3n b7n )/2)2 (7.38)
And the main one is as follows:
HD1 = 16(b0p b3p b6p − b0n b3n b6n + b1p b2p b6p − b1n b2n b6n
3
+ (b0 b5 b6 + b1 b4 b6 − b1 b2 b8 ) + b1p b2p b8p − b1n b2n b8n
4
3
+ (b0p b5p b6p − b0n b5n b6n + b1p b4p b6p − b1n b4n b6n − b1p b2p b8p
4
+ b1n b2n bSn ))2 + 16(b0p b2p b7p − b0n b2n b7n + b4p b0p b7p − b4n b0n b7n
3
+ (b0p b9p b8p − b0n b9n b8n − b4p b0p b7p + b4n b0n b7n − b1p b3p b7p
4
+ b1n b3n b7n ) + b1p b3p b7p − b1n b3n b7n )2 (7.39)
Hence, the second-order and third-order THD terms will be

!
HD2
HD2 = 100 (7.40)
HD1
!
HD3
HD3 = 100 (7.41)
HD1
Figure 7.16 and Fig. 7.17 represent these equations. Figure 7.16(a) shows the distor-
tion caused by the second-order harmonic when the ratios between the capacitances
that realise the exponents for the currents in the mathematical law are different
(a) (b)
HD2(%) HD3(%)
0.5
3
0.4
2
0.3
1
0 0.2
0.08 0.08
0.08 0.08
0.04 0.04 0.04
x2n 0 0 x 1n x2n 0.04 x
0 0 1n
Figure 7.16 HD2 and HD3 versus x1n and x2n , when x1p = x2p = 0.05 and the
input amplitude is 20e-9A
HD2(%) HD3(%)
3
0.4
2
0.3
1
0.2
0
0.08
0.08 0.08
0.08 0.04 0.04
0.04 x 1n
x 1n 0 0.04 A n (1–An /A p) 0 0 A n (1–An /A p)
0
Figure 7.17 HD2 and HD3 versus x1n and the mismatch affecting the gains (being
x1p = x2p = −0.05, and x2n = 0)
from one and also different from each other. The graph has been obtained for
x1p = x2p = 5 per cent and an input amplitude of 20e-9A. The ratio between the
third-order and first-order harmonic is represented in Fig. 7.16(b). It can be observed
how under certain conditions the second-order harmonic may be larger than the third-
order one. In Fig. 7.17 x2p = x1p = 0.05, x2n = 0 and the y axis represents any
mismatch affecting the positive and negative gains (Ap and An play the role of a nor-
malised A taking into account different gains for the positive and negative differential
sides in the integrator).
The main conclusion that can be extracted from the previous analysis is that
the fundamental source of distortion is the mismatch between the same parame-
ters for the two different single sides of the integrator. The second-order harmonic
originated because of this could even be larger than the third-order one. The most
Figure 7.18 Fabricated prototype layout
critical parameters are those related to the integrator time constant (included under
the form of different gains for the positive and negative parts, Ap and An ). They could
change mainly because of different Is for the transistors, which is very usual in weak
inversion. In any case, all these figures represent critical situations. In general, the
typical values of mismatch for, for example, capacitances are going to be well below
10 per cent75 .
7.5 A design example
This section illustrates the previously developed theory with some experimental
results obtained from a prototype design in a 0.35 μm technology ((AMS-CSX)
[217]). The filter parameters obtained from eq. (7.14) and eq. (7.15) are as follows:
√
ωo = ωo2 ωo4 (7.42)
√
ωo2 ωo4
Q= (7.43)
ωo1
ωo3
HLP (0) = (7.44)
ωo2
ωo3
HBP (ωo ) = (7.45)
ωo1
The filter operating at 1 V consumes less than 2 μW of power (this is without
taking into account the expander block, as its power will vary depending on the current
amplification required at the output). The filter performance and design parameters
are summarised in Table 7.2. The layout is shown in Fig. 7.18. Figures 7.19–7.24
75 And also the values of mismatch plus parasitic capacitances.

Table 7.2 Summary of the 2nd-order filter performance and design

parameters
Technology parameters VTn = 0.45 V VTp = −0.62 V

Voltage supply 1V
Maximum power 2 μW
Input capacitances 58 fF
Transistor sizes (μm/μm) 5/0.8
HD3@70 mVpp (Iin = 43 nApp ), fo = 500 Hz <0.4 per cent
DR >45 dB
PSRR+ >40 dB
fo programming range >60 dB, [40 Hz,40 kHz]
Q programming range >20 dB, [0.5,5]
Gain programming range >40 dB
Area 175 μm × 265 μm
5
dB
Log Mag
10
dB
/div
–95
dB 10 100 1k 10 k 100 k
Figure 7.19 Programming of the cut-off frequency in the bandpass output
show experimental examples of programming. The transfer function of the lowpass

realisation exhibits a peak which is caused by additional roots due to the parasitic
capacitances. Nevertheless, the peak is at least 30 dB below the gain in the passband
of the filter and so is not critical. It originates from the parasitic capacitances and is
not easy to deduce analytically, although it is normal in weak inversion realisations.
The measured Signal to Noise Ratio is shown in Fig. 7.25. Results of THD and the
HD3 (which is directly related to the IM3 as seen in previous chapters) are shown
in Fig. 7.26 and Fig. 7.27. Figure 7.27 stresses the importance of the second-order
5
dB
LogMag
10
dB
/div
–95
JB 10 100 1k 10 k 100 k
Figure 7.20 Programming of the cut-off frequency in the lowpass output
–10
dB
LogMag
5
dB
/div
–60 100
dB
Start: 30 Hz Stop: 300 Hz
Figure 7.21 Programming of the quality factor in the bandpass output
harmonic previously discussed in the theoretical study. In the figure the latter is only
32 dB beneath the fundamental tone.
This chapter described how to use FGMOS transistors to design internally nonlinear
and externally linear logarithmic filters. It has been shown how very compact
5
dB
LogMag
10
dB
/div
–95 10 100 1k 10 k 100 k

dB
Figure 7.22 Programming of the quality factor in the lowpass output
2.5
dB
LogMag
5
dB
/div
–47.5 1k 10 k
dB
Figure 7.23 Programming of the gain in the bandpass output
realisations can be designed by using the inputs in the transistor for different pur-
poses, such as tuning, signal processing, biasing and control of the common mode.
Because of this there is no need of stacking transistors to implement translinear
loops, which simplifies the design under the low-voltage constraint apart from hav-
ing other added advantages that have also been explained along the chapter. The main
5
dB
LogMag
10
dB
/div
–95 10 100 1k 10 k 100 k

dB
Start: 10 Hz stop: 200 000 Hz
Figure 7.24 Gain programming in the lowpass filter
–35
dB
LogMag
10
dB
/div
–135 10 100 1k 10 k 100 k

dB
Start: 10 Hz Stop: 200 nno Hz
Figure 7.25 1/Signal to noise ratio
disadvantage of these realisations is related to the parasitic couplings to the FGs in

the FGMOS devices that together with the mismatch increase the distortion levels.
In general though this kind of topology performs very well in low frequency, low
accuracy applications in which the main design constraints are power and low supply
voltage.
0.9
HD3 (%) 0.8
0.7
0.6
0.5
0.4
0.07 0.08 0.09 0.1 0.11 0.12
(Vinp –Vinn ) (Vpp )
Figure 7.26 HD3 versus the amplitude of a 500 Hz input signal for a bandpass filter
with unity gain and quality factor, and a cut-off frequency of 500 Hz
10–2
10–3
10–4
Gain
10–5
10–6
10–7
10000 20000 30000
fo (Hz)
Figure 7.27 Experimental second order harmonic for a low pass filter with a cut-off
frequency of 40 kHz, quality factor of 1, and an input signal of 200 mVpp
Notation
α Coefficient of the state variable in the state-space equation
of a linear lossy integrator (eq. (7.1))
α Modified expression of α if mismatch and CGD are taken
into account (eq. (7.26))
i for i = [1, 4] Parameter accounting for the percentage of mismatch
between different capacitances
η Input coefficient in the state-space equation of a linear lossy
integrator (eq. (7.1))
η Modified expression of η if mismatch and CGD are taken
into account (eq. (7.26))
ωo Biquad cut-off frequency (rad/s. Eq. (7.42))
ωoi for i = [1, 4] Constants in the state-space equations of the second-order
filter prototype (eqs. (7.14) and (7.15))
A Iin amplitude
An Normalised A for the n side of the fully differential
integrator taking into account different gains
Ap Normalised A for the p side of the fully differential
integrator taking into account different gains
A1 Ix amplitude
bi for i = [1, 9] See Table 7.2
bin bi parameters for the n side of the fully differential
integrator when mismatch is considered
bip bi parameters for the p side of the fully differential
integrator when mismatch is considered
Mi
CGD Gate to drain capacitance for transistor Mi (Fig. 7.5)
Cin /CT Weight of the effective input w
Cin2 Unit capacitance in transistor M2 (For study of mismatch.
Fig. 7.5)
Cin4 Unit capacitance in transistor M4 (For study of mismatch.
Fig. 7.5)
fo Biquad cut-off frequency (Hz)
Q Biquad quality factor (eq. (7.43))
Gex Gain of the expander block as given by eq. (7.20)
HLP (0) Lowpass filter DC gain (eq. (7.44))
HBP (ωo ) Lowpass filter DC gain (eq. (7.44))
I1n Single state variable related to the bandpass output
(eq. (7.14))
I1p Single state variable related to the bandpass output
(eq. (7.14))
I3n Single state variable related to the lowpass output
(eq. (7.14))
I3p Single state variable related to the lowpass output
(eq. (7.14))
IA Bias current in the divider block (Fig. 7.3), in the input stage
(Fig. 7.9) and in the CMFB (Fig. 7.11)
IB Current related to the parameter α in the integrator (eq. (7.11),
Fig. 7.4)
Iin Input current (eq. (7.6))
IinDC Iin DC component
Iinn Single input current in one of the sides of the fully differential
Iinp Single input current in one of the sides of the fully differential
Ijn for j = 1, 3 Output voltage in the expander block
Ijp for j = 1, 3 Output voltage in the expander block
Iout Output current in the integrator in Fig. 7.4
Ix State current (eq. (7.5))
IxDC Ix DC component
Ixn See Ioutn and Ix
Ixp See Ioutp and Ix
Iout1 Output of the NLB (Fig. 7.3)
Ioutj Ijp − Ijn for j = [1, 3]. Output of the expander block (Fig. 7.8,
eq. (7.23))
Ioutn Ixn . Single output of the expander block in the fully differential
integrator (Fig. 7.4)
Ioutp Ixp . Single output of the expander block in the fully differential
integrator (Fig. 7.4)
Iout1i Output of a generic NLB with an extra input connected to Vtuni
K ηkin /ky (eq. (7.4))
kin Multiplicative coefficient in the equation relating xin and u
(eq. (7.2))
ky Multiplicative coefficient in the equation relating x and y
(eq. (7.2))
(n, p) Subscript used to simultaneously refer to devices in the
negative and positive side of the fully differential integrator
u Variable related to xin in a logarithmic way (eq. (7.2))
V1n Voltage across one of the terminals of the integrating
capacitance associated with the bandpass output (eq. (7.16))
V1p Voltage across one of the terminals of the integrating
capacitance associated with the bandpass output (eq. (7.16))
V3n Voltage across one of the terminals of the integrating
capacitance associated with the lowpass output (eq. (7.16))
V3p Voltage across one of the terminals of the integrating
capacitance associated with the lowpass output (eq. (7.16))
VA Voltage generated by IA in the NLB (Fig. 7.3)
VAn VA voltage in the n side of the fully differential integrator
VAp VA voltage in the p side of the fully differential integrator
VB Constant voltage input in the expander block (eq. (7.5))
VB1 Voltage input in one of the two expanders in Fig. 7.7

VB2 Voltage input in one of the two expanders in Fig. 7.7
VBin Constant input in the compressor block (eq. (7.7))
Vcm Constant voltage applied to certain FGMOS transistors as
explained in Section 7.2
Ve Constant voltage in the expander block (Fig. 7.8)
Vin Voltage generated by Iin in the compressor block
(eq. (7.7))
Vinn Voltage generated by Iinn in the compressor block (eq. (7.7))
Vinp Voltage generated by Iinp in the compressor block (eq. (7.7))
Vjn for j = [1, 3] Input voltage in the fully differential expander block
Vjp for j = [1, 3] Input voltage in the fully differential expander block
n
Voutcm Negative output of the CMFB associated with the bandpass
output
n
Voutcm Negative output of the CMFB associated with the lowpass
output
p
Voutcm Positive output of the CMFB associated with the bandpass
output
p
Voutcm Positive output of the CMFB associated with the lowpass
output
Vtuni Tuning voltage that controls the value of ωoi (for i = [1, 4];
eq. (7.21))
Vx Voltage that generates Ix in the expander block (eq. (7.5))
Vxn Voltage that generates Ixn in the expander block (Fig. 7.4)
Vxp Voltage that generates Ixp in the expander block (Fig. 7.4)
w Weights of the FGMOS transistors in the divider block
wB Weight of input VB in the expander block (eq. (7.5)). Equal to
w in subsequent sections
wBin Weight of input VBin in the compressor block (eq. (7.6)).
Equal to w in subsequent sections
wD CGD /CT
wij Input weight associated with capacitance in transistor Mi
connected to transistor Mj (Fig. 7.5)
win Weight of input Vin in the compressor block (eq. (7.6)). Equal
to w in subsequent sections
wx Weight of input Vx in the expander block (eq. (7.5)). Equal to
w in subsequent sections
x State variable in the equation of a linear lossy integrator
(eq. (7.1))
x1 Exponent of the input current in the state-space equation when
mismatch is considered within the same branch only
(eq. (7.26))
x1(n,p) Parameter used to refer simultaneously to x1 for the n and p
side of the integrator when mismatch makes them different
x2 Exponent of the input current in the state-space equation when

mismatch is considered within the same branch only
(eq. (7.26))
x2(n,p) Parameter used to refer simultaneously to x2 for the n and p
side of the integrator when mismatch makes them different
xin Input variable in the equation of a linear lossy integrator
(eq. (7.1))
xinn Single input variable for a fully differential implementation in
eq. (7.12)
xinp Single input variable for a fully differential implementation in
eq. (7.12)
xn Single state variable for a fully differential implementation in
eq. (7.12)
xp Single state variable for a fully differential implementation in
eq. (7.12)
y Variable related to x in a logarithmic way (eq. (7.2))
Chapter 8
Low power digital design based on the FGMOS
threshold gate
8.1 Introduction
This chapter introduces a completely different application for the FGMOS transistors,
in digital circuit design. The device can be used to implement digital functions in
a much more compact way, which results in the desired reduction of power. The
methodology followed for this consists of the following steps: 1. Design the digital
functions at the system level using threshold gates. 2. Design the corresponding
threshold gates at the circuit level using FGMOS transistors. 3. Design the whole
system at the circuit level using the threshold gates developed in 2.
Threshold Gates are digital circuits whose functionality is based on the so called
majority or threshold decision principle. The threshold decision principle states that
the value of the signal at the circuit output depends on whether the arithmetic sum of
values at the inputs exceeds or falls below a threshold value. This general definition
includes as special cases conventional logic gates, such as AND and OR.
For many years, logic circuit design based on threshold gates has been considered
as an alternative to the traditional logic gate design procedure. The threshold gate
intrinsically implements a complex function. If, instead of using standard logic gates,
the threshold gate is used as main block of the system, the number of gates as well
as the gate levels can be reduced. However, the growing interest in threshold logic is
mainly due to the recent theoretical results that show how bounded level networks of
threshold gates can implement functions that would require unbounded level networks
of standard logic gates. Important functions such as multiple-addition, multiplication,
division or sorting can be implemented by polynomial-size threshold circuits with a
small constant depth.
Threshold gates are very attractive for low power design as their use can simplify
the topology of digital systems which normally results in a power reduction. This
chapter shows how a threshold gate can be built in a very compact form using the
x1 ≥T x1 w1
x2 f x2 w2 f
xn xn wn T
(a) (b)
Figure 8.1 (a) Standard and (b) Non-standard symbol for the threshold gate
FGMOS (Neuron-νMOS76 ) transistor. This threshold gate is used to design several

important digital blocks that are commonly used in more complex digital systems.
The figure of merit power × delay for all presented blocks is shown and compared
with the figure of merit of conventional architectures, confirming the potential of the
νMOS device for low power digital design.
8.2 Threshold gates
An n-input threshold gate (TG) is defined as a logic gate with n inputs, xi , (i =

1, . . . , n), which can take values 0 and 1, and for which there is a set of (n + 1) real
numbers w1 , w2 , . . . , wn and T , called weights and threshold, respectively, such that
the output of the gate is
⎧
⎪ n
⎪
⎪ wi xi ≥ T
⎨1 for
i=1
f = (8.1)
⎪
⎪ n
⎪
⎩0 for wi xi < T
i=1
A function represented by the output of a TG, denoted by f (x1 , x2 , . . . , xn ), is called a

threshold function. The set of weights and threshold can be written in a more compact
vector notation as [w1 , w2 , . . . , wn ; T ].
Figure 8.1(a) shows the IEEE standard symbol for a TG with all the input weights
equal to 1 and threshold T . There is no standard symbol for a threshold gate with
weights different from 1, but it can be built by tying together several inputs (a weight
of wi for the input xi can be obtained by connecting xi to wi gate inputs). Instead of
that the non-standard symbol for TGs with generic weights is used (Fig. 8.1(b)).
A threshold gate can be defined with positive and negative weights as well
as with positive and negative thresholds. However, in order to facilitate the
physical implementation it is useful to have an alternative TG description only
with the positive parameters. This can easily be obtained by taking into account
the elementary relations of threshold functions shown in Fig. 8.2. The arrow
labelled with (1) means that if f (x1 , x2 , . . . , xn ) is a threshold function defined by
76 Both names will be used indistinctly from now on. The second one is commonly used in digital
design, whereas the first is a much more general nomenclature.
Low power digital design based on the FGMOS threshold gate 247
· ·
f ( x) f f (x )
[w; T ] (1) [–w; 1−T ]
x x
[–w; T− Σ wi ] f [w; 1−T + Σ wi ]

·
f (x ) f (x )
Figure 8.2 Elementary relations of threshold functions
[w1 , w2 , . . . , wn ; T ], then its complement f˙ (x1 , x2 , . . . , xn ) is also a threshold function

defined by [−w1 , −w2 , . . . , −wn ; 1 − T ]. If a function can be realised as a thresh-
old element, then by selectively complementing the inputs, it is possible to obtain a
realisation using a threshold element with only positive weights.
Once the architecture has been chosen, the effectiveness of the system implemen-
tation is determined by the availability, cost and capabilities of its basic building block.
Various interesting circuit concepts have been reported for developing standard-
CMOS compatible threshold gates. They can be classified as static and dynamic
realisations, depending upon the context in which they are used. The Neuron MOS
(νMOS) threshold gate constitutes one of the most notable contributions to the
implementation of static-CMOS threshold gates [227–235].
8.3 The νMOS threshold gate
The realisation of very large logic circuits using νMOS transistors has been found to
be a very promising alternative to conventional approaches [50–52,58]. This is mainly
due to the enhancement of the functional capability of a basic transistor element which
effectively reduces the complexity of the overall circuit.
Logic design techniques for implementation of νMOS circuits were introduced
by Shibata [50–52], but their usefulness was limited because, in general, they led to
complex circuit configurations which required handling 2n logic states for an n-input
logic function. This imposed stringent constraints on process tolerances which are
not realisable by present technologies, even in cases when n is small. A different
and more powerful approach for designing using νMOS takes advantage of the fact
that the functionality of νMOS circuits is closely related to the functionality of a TG,
Fig. 8.3 [235–237]. The rest of this section explains the implementation of the TG
with νMOS transistors, first at a theoretical level and then briefly discussing several
practical considerations.
8.3.1 Theory
The simplest νMOS-based threshold gate (νMOS-TG) is a complementary inverter
using both p- and n-channel FGMOS devices, followed by a normal CMOS inverter.
The schematic of this TG is shown in Fig. 8.3. The FG is usually shared by both the
VFG
x1
x2
x3 f
xn
VC
Figure 8.3 νMOS Threshold Gate (νMOS TG)
pMOS and the nMOS transistors, although this does not have to be always the case.
There are several input gates (x1 , x2 , . . . , xn ), corresponding to the TG inputs plus
an extra input (indicated by Vc in the figure) for logic threshold adjustment (it will
be explained later). Neglecting the parasitic capacitances, and ignoring for now the
extra control input, the voltage at the FG is
n *

VFG = Ci Vxi CT (8.2)
i=1
where Ci is the input capacitance corresponding to the input xi , CT is the total capac-
itance seen by the FG and Vxi is the voltage applied to the input gate xi . When
VFG becomes higher than the inverter switching threshold voltage, Uinv , the output
switches to logic 1.
A νMOS TG has digital entries, i.e. Vxi = xi VDD , where VDD is the power supply
voltage and xi ∈ {0, 1}. There is a relation between the above equation for VFG and
the definition
of the TG given by eq. (8.1). Namely, the weighted summation in the
TG, ni=1 wi xi , can be implemented by the capacitive network in the νMOS device:
n *

Ci xi VDD CT (8.3)
i=1
The weight for each input is then proportional to the ratio between the corresponding
input capacitance Ci and CT :
wi = Ci VDD /CT (8.4)
Thus, design involves mapping the logical inequalities

n
wi xi ≥ T (8.5)
i=1

n
wi xi < T (8.6)
i=1
to the electrical relationships given by eq. (8.7) and eq. (8.8) by means of capacitances
sizing and tuning of the switching threshold of the inverter:
n *

Ci xi VDD CT > Uinv (8.7)
i=1
n *

Ci xi VDD CT < Uinv (8.8)
i=1
When the logical threshold T is not centred77 , the threshold voltage of the inverter
also needs to be non-centred (near 0 or VDD ). This can be achieved using extra inputs.
For example, let us assume a single control input c with capacitance CControl . If Vc is
applied to this control input c, the new VFG is78
n *

VFG = Ci xi VDD Ctot + (CControl Vc )/CT (8.9)
i=1

n
CT = Ci + CControl (8.10)
i=1
From the point of view of the TG, a comparison is performed between (8.3) and
Uinv − (CControl Vc )/CT (8.11)
Thus, the effective threshold voltage of the inverter can be modified. In practical
digital design analog voltages are avoided and so the role of the analog extra input
Vc is realised by a set of digital inputs with the appropriate coupling capacitances.
In a real design of a νMOS TG second-order effects also need to be considered,
although they were not included in the previous equations for the sake of clarity. The
main issue that should be taken into account in the circuit in Fig. 8.3 is related to the
coupling effects when switching on the power supply. Capacitances CGD , CGS and
CGB are responsible for the effects, causing additional terms in VFG . However, these
can be minimised either by sizing adequately the input capacitors and transistors or
choosing the switching threshold, accordingly.
8.3.2 Practical design aspects

The previous section described the theory behind the implementation of TGs with
νMOS devices. However, practical design considerations also have to be taken into
77 i.e. T is far from (n w )/2.

i=1 i
78 Neglecting the parasitic couplings.
account and the following issues must be solved:

1. Signal regeneration. On one hand, these TGs exhibit reduced noise mar-
gins as a consequence of the circuit electrical operation. On the other hand, some
input combinations can produce output voltages different from VDD or ground.
These two considerations are specially critical when logic networks are built up of
interconnecting νMOS-TGs.
2. Electrical simulation. The electrical simulation problem has been already
addressed in Chapter 2; so, it will not be considered here again. The simulation
problems can be avoided by using the ITA method described in detail in Section 2.4.
8.4 νMOS Threshold gates applications
The implementation of digital circuits using νMOS TGs is illustrated in this section
with several examples. These focus on the design of low power-high speed blocks.
These examples can be classified into combinational and sequential circuits. The first
part of the section is divided into arithmetic combinational blocks, specifically TG
based adders using floating-gate circuits [238] and νMOS-based compressor designs
[239–240]. It also describes a sorting network [241] that is used as the main functional
cell in the implementation of arithmetic circuits afterwards [242–244]. The second
part describes, characterises and illustrates with experimental results the design of a
sequential block: the multi-input Muller C-element, as an example of the behaviour
of the νMOS TG in a configuration with feedback [124].
8.4.1 Threshold logic based adders using floating-gate circuits

Binary addition is an arithmetic operation widely used in many fields. It is also essen-
tial for any system containing circuitry for subtraction, multiplication and division
[245–248]. Often it is the limiting element in the speed operation of such systems,
and therefore implementing high-performance adders is of utmost importance. Opti-
misation of adders can be realised either at the logic or at the circuit level. Logic
optimisation generally comes from rewriting the logic equations which define the
adder in such a way that their implementation results in a cheaper or faster realisa-
tion. Circuit optimisation is based on considering some of the realisation electrical
aspects such as topology and transistor sizing.
A bottleneck in the design of adders is the carry generation. N-bit ripple-carry
adders have a propagation delay linear with N and these are only useful for addition
of numbers with a relatively small word length. Nowadays, adders for a word length
up to 128 are required which makes the use of N -bit ripple-carry adders impractical.
One of the most common techniques to avoid this problem, and to design very fast
logic adders, uses the carry lookahead principle. The latter allows the optimisation
of the adder at a logic level resulting in implementations which effectively eliminate
the undesired ripple effect and exhibit a constant time delay. However, the physical
implementation of these carry lookahead adders can become impractical because of
the very high fan-ins required. To solve this problem a hierarchical approach is usually
taken, and a logarithmic time delay is so obtained.
This section will show how to implement a very fast carry lookahead by using
a design style based on TGs implemented as νMOS circuits. The technique is very
effective to cut-off the time delay and to increase the maximum operating frequency
of these adders.
8.4.1.1 Carry lookahead adders

Let A = An−1 , . . . , A1 , A0 and B = Bn−1 , . . . , B1 , B0 be the augend and addend inputs
to an n-bit adder, Ci−1 the carry input to the i-th bit position and C−1 the corresponding
input to the least significant position. In a carry lookahead adder, the carry-out of
each stage, Ci , is directly defined as a function of Ai , Ai−1 , . . . , A0 , Bi , Bi−1 , . . . , B0
and C−1 . By defining two auxiliary functions, the carry generate Gi and the carry
propagate Pi functions, the sum and carry outputs of the i-th stage can be obtained as
Si = Pi ⊕ Ci−1 Gi = Ai · Bi
(8.12)
Ci = Gi + Pi · Ci−1 Pi = Ai ⊕ Bi
These equations mean that if all the carry inputs Cn−2 , . . . , C1 , C0 , C−1 are avail-
able simultaneously, then all the sum bits Si for i = n − 1, . . . , 1, 0 can be generated
in parallel. By expanding the recursive formula for Ci , the set of eqs. (8.13) for the
carries can be obtained, which can be realised in two logic levels:
C0 = G0 + P0 · C−1
C1 = G1 + P1 · C0 = G1 + P1 · (G0 + P0 · C−1 )
= G1 + P1 · G0 + P1 · P0 · C−1
C2 = G2 + P2 · C1 = G2 + P2 · (G1 + P1 · C0 )
= G2 + P2 G1 + P2 P1 G0 + P2 P1 P0 C−1
································· (8.13)
Ck = Gk + Pk · Gk−1 = Gk + Pk · (Gk−1 + Pk−1 · Ck−2 )
= Gk + Pk Gk−1 + Pk Pk−1 Gk−2 + · · ·
+ Pk Pk−1 · · · P2 P1 G0 + Pk · · · P1 P0 C−1
····································
Cn−1 = Gn−1 + Pn−1 Gn−2 + · · · + Pn−1 · · · P1 P0 C−1
The final sum bits are available after a delay equal to
= PG + CLA + S (8.14)
where PG is the circuit delay due to the carry generate/propagate unit generating
the Pi and Gi signals, CLA is the delay of the CLA producing the Ci signals and
S corresponds to the generation of the Si sum signals (summation unit). Hence, for
example, in a 4-bit realisation, under the simplifying assumption that a gate delay is
k–1 k–1 k–1 k–1

P 4j+3 G 4j+3 P 4j+1 G 4j+1
k–1 k–1 k–1 k–1

P4j+2 G4j+2 P 4j G 4j
BCLA ( j, k) C k+1
4 j–1
C k+1 k
4 j + 4 –1
k k C k+1 k
Pj Gj 4 j + 2 . 4 –1
C k+1 k
4 j + 3 . 4 –1
Figure 8.4 4-bit block carry lookahead (BCLA) unit
constant and equal to δ, the final sum bit is available after 8δ:3δ (corresponding to
the circuit delay to generate the Pi and Gi signals, PG ) plus 2δ (for CLA ) plus 3δ
(for S ). However, direct calculation of carries beyond 4 bits becomes impractical
because of the very high fan-ins required in the implementation, and a hierarchical
approach is needed [249].
The hierarchical solution uses the so-called block carry lookahead (BCLA) unit.
Each one of such units computes its own ‘group’ carry propagate and generate from
the signals coming on it. Figure 8.4 shows the 4-bit BCLA unit corresponding to a
generic j-th column and k-th row in an adder tree implementation.
Equations defining the outputs of this BCLA are the carry signals
C4k+1 j+4k −1 , C4k+1 j+2·4k −1 , C4k+1 j+3·4k −1 , the group propagate signal Pjk , which
determines whether a carry into the block would result in a carry out of the block, and
the group generate signal Gjk , corresponding to the condition that the carry generated
out of the most significant position of the block was originated within the block itself:
k−1 k−1
C4k+1 j+4k −1 = G4j + P4j · C4k+1 j−1
k−1 k−1 k−1 k−1 k−1
C4k+1 j+2·4k −1 = G4j+1 + G4j · P4j+1 · P4j+1 · P4j · C4k+1 j−1
k−1 k−1 k−1 k−1 k−1 k−1
C4k+1 j+3·4k −1 = G4j+2 + G4j+1 · P4j+2 + G4j · P4j+2 · P4j+1
k−1 k−1 k−1
+ P4j+2 · P4j+1 · P4j · C4k+1 j−1 (8.15)
k−1 k−1 k−1 k−1 k−1 k−1
Gjk = G4j+3 + G4j+2 · P4j+3 + G4j+1 · P4j+3 · P4j+2
k−1 k−1 k−1 k−1
+ G4j · P4j+3 · +P4j+2 · P4j+1
k−1 k−1 k−1 k−1
Pjk = P4j+3 · P4j+2 · P4j+1 · P4j
Let us assume an n-bit adder and a b-bit BCLA. The delay corresponding to this
hierarchical solution is
= PG + (logb n − 1) · BCLA + CLA + S (8.16)
where the additional term BCLA is the circuit delay of the block carry lookahead
unit, (logb n − 1) is the depth of the adder tree and the term CLA is now the delay
of the final CLA unit.
The next section shows an alternative representation of the carry lookahead adder
using TGs.
8.4.1.2 Threshold gate based implementation

Optimised high-performance adders can be obtained by resorting to threshold logic.
The application of the threshold gate design methodology allows the reduction of
the delay for the carries in a carry lookahead implementation. This is because it
is possible to obtain the carry signals C0 , C1 , C2 and ‘group’ carry propagate and
generate functions as TGs whose weights and thresholds are as follows:

k−1 k−1
C4k+1 j+4k −1 C4k+1 j−1 , P4j , G4j = [1, 1, 2; 2]

k−1 k−1 k−1 k−1
C4k+1 j+2 · 4k −1 C4k+1 j−1 , P4j , G4j , P4j+1 , G4j+1 = [1, 1, 2, 3, 5; 5]

k−1 k−1 k−1 k−1 k−1 k−1
C4k+1 j+3 · 4k −1 C4k+1 j−1 , P4j , G4j , P4j+1 , G4j+1 , P4j+2 , G4j+2
= [1, 1, 2, 3, 5, 8, 13; 13] (8.17)

k−1 k−1 k−1 k−1 k−1 k−1 k−1
Gjk G4j , P4j+1 , G4j+1 , P4j+2 , G4j+2 , P4j+3 , G4j+3
= [1, 1, 2, 3, 5, 8, 13; 13]

k−1 k−1 k−1 k−1
Pjk P4j , P4j+1 , P4j+2 , P4j+3 = [1, 1, 1, 1; 4]
These results are not specific for these carry signals. It can be proven that all the
elements of the set of carry signals C0 , C1 , . . . , Cn−1 are threshold functions which
can be implemented in one logic level. Weights and threshold of the TGs obtained
are related with the Fibonacci’s number series in the following way:
Cj (C−1 , P0 , G0 , P1 , G1 , . . . , Pj−1 , Gj−1 , Pj , Gj )
= [1, 1, 2, 3, 5, . . . , F2j , F2j+1 , F2j+2 , F2j+3 , F2j+3 ] (8.18)
where Fk is the k-th Fibonacci’s number, i.e. each Fk is the sum of the two adjacent
Fibonacci’s numbers on its immediate left, Fk = Fk−1 + Fk−2 . The total weight for
the TG implementing Cj , WT (Cj ), is given by
WT (Cj ) = 1 + 1 + 2 + · · · + F2j+2 + F 2j+3 = F 2j+5 − 1 (8.19)
The availability of all the carry signals in only one level in a TG based implemen-
tation is theoretically important but of a limited practical interest since its usefulness
depends on the physical availability of TGs implementing the total weight. This will
be discussed in the next section. In general, there is a bound to the maximum total
weight physically implementable on a TG, and so a hierarchical approach will be
also needed. The practical difference with a traditional solution is that the BCLA
unit can be implemented by one level of TGs, while it requires a two-level network
when implemented with traditional gates. The delay for the n-bit adder, TG , when
a hierarchical approach is followed becomes
TG = PG + (logb n − 1) · TG

BCLA + CLA + S (8.20)
where CLA is the delay of the final CLA unit (the same two-level circuit for both
approaches) and TGBCLA is the circuit delay of the BCLA unit based on TGs. From
equations (8.16) and (8.20) and assuming a delay of δ units per level, the difference
between both delays is given by (logb n − 1) · δ; i.e.
= TG + (logb n − 1) · δ (8.21)
That is, the delay improves as n increases in an n-bit adder. To actually evaluate the
reduction of the delay, it is necessary to exactly know the design style used in the
implementation of the TGs. The next section analyses how the choice of technology
affects the implementation of the carry lookahead adder.
8.4.1.3 Physical implementation

The feasibility of the TG approach discussed in Section 8.3 depends on the com-
plexity of the TG that can be built with νMOS transistors. Results obtained from an
extensive set of simulations showed that carry signals implemented by TGs up to a
maximum number of bits equal to 4 (b = 4) can be built. If both 4-bit BCLAs, the
traditional BCLAs and the νMOS based one, are designed, the delay of the νMOS
implementation is around half of the delay of the traditional implementation. This
reduction is of the greatest importance because this delay multiplies the depth of the
network implementing the adder. The higher the number of bits of the adder the larger
the reduction of the delay. This is illustrated in the following with the design of a
64-bit adder.
8.4.1.4 The 64-bit adder

An example of the application is the 64-bit adder shown in Fig. 8.5. The adder uses
the 4-bit BCLA unit in Fig. 8.4. The carry generate/propagate unit and the summation
unit are not shown, but they are a direct implementation of eq. (8.12). The 64-bit adder
has two levels of 4-bit BCLAs and a final level with a 4-bit CLA; j = 0, k = 0 is the
block at the right top position. The results of performance shown are obtained in a
prototype designed and laid out in a 5 V 0.8 μm double poly CMOS process [214].
Correct operation under process, mismatch and ambient parameter variations
are validated through extensive Monte Carlo HSPICE simulations of the extracted
circuit. Transient characteristics and average power are measured on post-layout
simulation results using typical device parameters at a supply voltage of 3 V. The
worst case delay time is 5.6 ns and the power consumption is 19.6 mW at 100 MHz.
P G P G P G P G P15:12 G P G 11:8 P 7:4 G 7:4 P3:0 G 3:0

63:60 63:60 59:56 59:56 55:52 55:52 51:48 51:48 15:12 11:8
(15,0) (14,0) (13,0) (12,0) (3,0) (2,0) (1,0) (0,0)

C59 C 55 C 51 C 47 C 11 C7 C3 C– 1
P 13 G 13
P 14 G 14
P 12 G 12
0
0
P 15 G 15
0 0
P3 G 3
0 0
P2 G2
0 0
P1 G1
0 0
P0 G0
0
C 62:60 C58:56 C54:52 C50:48 C14:12 C 10:8 C 6:4 C 2:0
0
0
0
0
(3,1) (0,1)
C 47 C–1
1 1
P3 G3
1 1
P0 G0
1 1
P2 G2
1 1
P1 G1
C 59 C 55 C 51 C 11 C 7 C 3
CLA C–1
( j,k)
BCLA( j,k) C 63 C 47 C 31 C 15
Figure 8.5 Core of a 64-bit adder using the carry lookahead principle
However, the intrinsic nature of the νMOS approach makes this power consumption
very independent of the frequency.
In order to validate the proposed circuit, a comparison with another implementa-
tion is required. The 64-bit adder is also designed and laid out following a conventional
approach (NAND gates were used in the two-level implementation of the BCLA) and
with the same technological process. The worst case delay for this conventional design
is over 11 ns and the power consumption at 50 MHz is equal to the power consumption
of the TG-based implementation. Also, unlike the proposed design this one is very
dependent on frequency.
8.4.2 νMOS-based compressor designs

Partial-product reduction circuits (compressors) are of capital importance in the
design of high performance parallel multipliers. This section focuses on compressor
design based on TGs implemented as νMOS circuits.
The dominant delay in the critical signal path of a parallel multiplier is due to the
partial-product summation tree (PPST) [250]. A reduction of this delay is important
in order to obtain high performance parallel multipliers. It is well known that it is
hard to do layout for tree structures and hence large wiring channels are required. Full
adders have been widely used as the basic building block but its 3:2 nature (it takes
three inputs of the same weight and produces two outputs) makes it impossible to
design a completely regular layout. When compressors are used as building blocks,
more regular tree structures with fewer levels are obtained. Thus, PPSTs built with
compressors instead of with full-adders have been preferred and so the design of high-
performance compressors appears as a key problem in the design of high-performance
multipliers.
An (N , 2) compressor is defined to be a combinational circuit which operates in a
single column of the PPST taking as input N bit signals, (x0 , x1 , . . . , xN −1 ), from the
x0 x1 x2 x3 x0 x1 x2 x3
T1 (4, 2) t1 FA
T1 t1
y1 y0
(a) FA
, 2)
(4,
y1 y0 (b)
Figure 8.6 (a) The (4,2) compressor. (b) Full-adder implementation.
column and a number l = N − 3 of carry-in signals, (t1 , t2 , . . . , tl ), from the previous

column. It outputs two signals, (y1 , y0 ), in its column as well as l carry-out signals,
(T1 , T2 , . . . , Tl ), which are connected in the same order to the input lines of the next
compressor to preserve the signal ordering. Additionally, the significance of an (N , 2)
compressor comes from the fact that horizontal as well as vertical propagation of the
signals is allowed, instead of the traditional vertical direction of the full-adder tree.
In this section the problem of optimal design of compressors is addressed in two
ways: first, the logic circuit which evaluates the PPST is improved using TGs obtain-
ing an efficient logical redesign for the compressor basic building block, and second,
such TGs are implemented in a very efficient way by using the νMOS principle. It
is shown how the resulting νMOS basic compressors compare very favourably in
power consumption and speed with previous conventional circuits.
8.4.2.1 The (4,2) compressor-design approach

A (4,2) compressor has five inputs of the same weight, (x0 , x1 , x2 , x3 , t1 ), and three
outputs, (T1 , y1 , y0 ), and it is generally implemented from two full adders with outputs
y1 and T1 having the same weight (Fig. 8.6). Output T1 is not a function of the input
t1 , otherwise a ripple carry could occur.
The (4,2) compressor described here is based on TGs. Logic synthesis of an opti-
mum threshold gate network for the (4,2) compressor has been carried out using
integer linear programming. Figure 8.7 shows the corresponding threshold gate net-
work which requires only two levels of TGs. The main characteristic of this network
is the capability to obtain signals y1 and y0 with a delay which is the same and equal
to the delay of two TGs.
8.4.2.2 Conventional via νMOS TG-based compressors – performance

This section illustrates the performance of both a (4,2) and a (6,2) compressor designed
and laid out in a 0.8 μm double poly CMOS process. The correct operation under
x0 1 (4, 2)
x1 1 T1
x2 1 2
TG1 –2
x0 1
x1 1
x0 1 y1
x2 1
x1 1 x3 1
x2 1 t1 1 2
x3 1 TG4
t1 1 4
TG2 –2
–2
x0 1 x0 1
x1 1 x1 1 y0
x2 1 x2 1
x3 1 x3 1
t1 1 2 t1 1 1
TG3 TG5
TG 1 = [1,1,1;2]
TG2 = [1,1,1,1,1;4]
TG3 = [1,1,1,1,1;2]
TG4 = [–2,1,1,1,1,1;2]
TG5 = [–2,–2,1,1,1,1,1;1]
Figure 8.7 TG implementation of a (4,2) compressor
process and ambient parameter variations is validated using HSPICE simulations of

the extracted circuit that include Monte Carlo simulations and simulations using dif-
ferent standard worst case device parameters. Figure 8.8 shows simulated waveforms
for the extracted (4,2) compressor for a sample of inputs. A new pattern is applied each
2.5 ns. The first five waveforms are the input signals and the next three ones are the
output signals. For the purpose of comparison, a (4,2) compressor built from CMOS
full adders is also designed and laid out. A well-known CMOS full adder topology
with only 28 transistors is chosen as it requires a smaller circuit size and performs
better in terms of speed and power dissipation than other CMOS implementations
[251]. Sizing of transistors is carried out to optimise the transient performance of the
full adder. The area for these two implementations is similar. Transient characteris-
tics and average power are measured from post-layout simulation results using typical
device parameters at a 5 V supply voltage. Table 8.1 compares the transient perfor-
mance and the power-delay product of the two approximations for a (4,2) compressor.
The results of a (6,2) compressor are also shown in Table 8.1. The delay corresponds to
the worst case scenario; this is, situations where the inputs or the input sequence cause
the slowest circuit operation. The power is measured using a random generated input
sequence with 100 vectors. The table shows that a TG-based implementation is clearly
advantageous over a conventional one. The operating frequency for both νMOS TG-
based compressors is at least doubled when compared with previous conventional
Panel 1
Wave Symbol
Voltage (lin) Voltage (lin) Voltage (lin) Voltage (lin) Voltage (lin) Voltage (lin) Voltage (lin) Voltage (lin)
v(t1) 4
v(y1) 2
v(y0)
v(T1) 0
v(x3)
v(x2) 4
v(x1)
v(x0) 2
0
4
2
0
4
2
0
4
2
0
4
2
0
4
2
0
4
2
0
20n 25n 30n 35n 40n 45n 50n 55n 60n 65n 70n 75n 80n 85n 90n
Time (lin)
Figure 8.8 Simulated waveforms for a (4,2) compressor
approaches. Faster operation of the νMOS TG-based compressors does not increase
the power dissipation too much. Power Delay Product (PDP) figures of the νMOS
versions are also better. According to the obtained results, the νMOS designs are
very efficient when compared with other reported implementations. They occupy a
similar area, but exhibit a better time performance and power-delay product.
8.4.3 Sorting networks implemented as νMOS circuits

This section describes a hardware realisation of a binary sorting network (SN) based
on the use of νMOS transistors. An n-input SN is a switching network with n outputs
that generates an output which is a sorted (non-increasing order) permutation of
inputs. Binary SNs are built from comparator cells which have two inputs and two
Table 8.1 Comparison of performance
Compressor Conventional TG-based
(4,2) Delay time 3.1 ns 1.2 ns

PDP @200 MHz 8.6 pJ 6.9 pJ
(6,2) Delay time 4.5 ns 1.6 ns
PDP @200 MHz 20.3 pJ 15.8 pJ
x1
{
O1
{
x2 k first
O2 outputs
to 1
binary n-tuple Sorting
with k 1’s Ok
Network
n-k following
x n–1
xn
(a)
O n–1
On { outputs to 0
x1 O1
x1 2-input 2-input
+ x1 + x2
x2 SN SN O2
2-input
x1 & x2 x3 SN O3
x2 & 2-input 2-input
x4 SN SN O4
(b) (c)
Figure 8.9 (a) Sorting Network with k binary signal inputs equal to 1. (b) Logic
gate implementation of the comparator cell (2-input SN). (c) Batcher’s
implementation of a 4-input SN
outputs: one of them provides the maximum of both inputs and the other the minimum.
The internal structure of the comparator depends on the application. The inputs can
be either binary numbers, the comparator being then a complex element, or binary
signals, in which case maximum and minimum become OR and AND operations
respectively. Figure 8.9(a) shows the operation of an n-input sorter with k inputs
equal to 1. Figure 8.9(b) shows the logic implementation of the comparator cell. The
efficient implementation of an SN has for many years been an important subject of
research [252]. The method proposed by Batcher was a milestone [253]. Figure 8.9(c)
shows a 4-input SN implemented following Batcher’s method.
A different approach to the implementation of a sorter will be next presented,

which is based on the fact that each output of an n-input sorter depends only on the
number of inputs equal to 1. The i-th output will be 1 if and only if at least i of the
n inputs are 1. In this way, an n-input sorter can be seen as a cascaded two-block
circuit. The first block provides an output which depends linearly on the number of
1s applied at the inputs. The second block takes this output signal and compares it
with a set of n fixed values using a battery of comparators, thus providing the set of
n output functions of an n-input sorter.
An electrical realisation of an 8-input sorter can be carried out following the
above two-block approach and resorting to the νMOS transistor principle for its
implementation, as it requires counting the number of 1s in the inputs, or equivalently
performing their arithmetic addition. This is presented in the following subsection,
together with a comparison with the traditional solution.
8.4.3.1 Electrical realisation of a sorter circuit

Figure 8.10 shows the two-stage schematic diagram of a νMOS based n-input sorter.
The implementation of the first block resorts to the νMOS principle and to current
mirroring to provide an analog output voltage, V1 , which increases proportionally,
in a staircase shape, to the number of binary inputs equal to 1. This operation is
performed by transistors M1 –M4 which are biased in the strong inversion saturation
region. Transistors M2 and M4 are equally sized n-channel νMOS transistors. M1 and
M3 are equally sized pMOS transistors. The sorter inputs are the M2 input gates capac-
itively coupled to its FG with identical coupling capacitances, Cu , which produce an
O1
IO1
O2
M1 IO2
M3
V0
ΦR V1
x1
x2
xn M2
M4
ΦR I1
On
ΦR V FG IOn
second stage
ΦR
first stage
Figure 8.10 Two-stage schematic of a νMOS based n-input sorter

FG voltage, VFG , linearly dependent on the sum of the inputs. However, with this
circuit several input combinations with different numbers of 1s can give floating gate
voltages below the threshold voltage of the n-channel MOS transistor which would
not be properly distinguished. In order to avoid such a situation, the inverter I1 as
well as two additional inputs to transistor M2 with coupling capacitances Cu /2 and
Co have been included. During the initialisation phase, R = 1 switches that are on
during this phase connect together M2 FG as well as the output and input of I1 . Simul-
taneously the input terminals x1 , x2 , . . . , xn are connected to ground (input switches
not shown in Fig. 8.9). After initialisation, during processing mode, when R = 0
n

VFG = xi · VDD · Cu /CT + Uinv(I1 ) − VDD · (Cu /2)/CT (8.22)
i=1
where Uinv(I1 ) is the threshold of inverter I1

CT = (n + 1/2)Cu + Cchan + Co (8.23)
Capacitance Co is introduced by the extra grounded input in order to maintain
M2 saturated, even when the n inputs of the sorter are at logic ‘1’ (see Section 3.3.5),
and Cchan accounts for the transistor parasitic capacitances. VFG controls the current
through M1 and M3 . Since M4 is equal to M2 the voltage produced at the M4 drain
terminal, V1 = VFG . With this scheme, the obtained analog output staircase shape
voltage V1 is robust to process parameter variations.
The second block comprises a set of comparators which are implemented as
inverters. Each inverter is sized in such a way that its switching threshold falls between
two given consecutive steps of the aforementioned staircase. For example, the output
O1 must be a logical ‘1’ if there is at least an input at a logical ‘1’ and so the switching
threshold of inverter IO1 is fixed to (V1 (0) + V1 (1))/2, where V1 (0) stands for the
voltage at node V1 when an all zero input vector is applied. V1 (1) corresponds to the
voltage at node V1 when an input vector with only one ‘1’ is applied.
Due to process parameter variations the voltages V1 (i), i = 0, . . . , n, as well as
the switching thresholds of the comparator-inverters can change from their nominal
value. In order to reduce this sensitivity I1 is made identical to IO1 . The role of
capacitor Cu /2 becomes now clear; it assures that with all inputs at logical zero the
V1 voltage is under Uinv(IO1 ) , the switching threshold of inverter IO1 .
8.4.3.2 Design and evaluation of an 8-input sorter

This section illustrates the performance of an 8-input νMOS sorter designed and laid
out in a 0.8 μm double poly CMOS process. In order to minimise excessive load of
the comparators over the first block, the M3 -M4 branch is replicated. Comparators
are made of chains of inverters in order to regenerate the output signal to full logic
swing. The operation is validated through HSPICE simulations of the extracted circuit
including Monte Carlo simulations and simulations using different standard worst
case device parameters. Figure 8.11 shows the simulation results for nodes V0 and V1
as functions of the number of inputs at logic ‘1’. It can be seen how the changes at
V0 are significantly reduced at V1 .
Wave Symbol
vout 4
vout
vout
vout V0
vout
3
Voltage (V)
3 2
1
300 n Time (s) 350 n

Wave Symbol
vout1
4
vout1
vout1
vout1
vout1
V1
Voltage (V)
300 n Time (s) 350 n

1 worst case (zero)
2 Monte Carlo
3 worst case (one)
Figure 8.11 Simulated behaviour of nodes V0 and V1 as functions of the number of

inputs equal to 1 showing the stabilising action of the circuit
For the purpose of comparison, an 8-input sorter following the Batcher’s conven-
tional approach consisting of a network of comparator cells is also designed and laid
out. There are 23 of such cells in an 8-input SN. Table 8.2 compares the area, the time
performance and the power consumption of both sorters. Transient characteristics and
average power are measured on post-layout simulation results using typical device
parameters at a supply voltage of 5 V. The worst case delay time happens in situations
when the inputs or the input sequence are such that the circuit operation is the slowest.
In the conventional design an input vector exciting the true longest path has been used
to measure that delay. In the νMOS counterpart an input vector consisting of only
‘1s’ followed by an input vector consisting of only ‘0s’ is employed. The power is
measured using a random generated input sequence with 100 vectors.
Table 8.2 Area, time performance and power consumption of

νMOS and conventional sorters
Area Worst case delay Power consumption

(@175 MHz)
νMOS 5625 μm2 4.1 ns 7.8 mW

Conventional 45400 μm2 5.2 ns 8.2 mW
Compared with a conventional gate-based implementation, the νMOS design is

very efficient in terms of area. It occupies an area that is nearly an order of magnitude
smaller than in its conventional counterpart and, at the same time, it exhibits a better
transient performance. Furthermore, the power consumption is found to be very
dependent on the frequency in the conventional design as opposed to the νMOS
design. For frequencies of 170 MHz and above, the νMOS sorter consumes less
power than the conventional implementation.
Finally, it could be claimed that the transistor used in the sorter is not a real νMOS
since the FG is not really floating (see Section 2.4.2). However, the FG is effectively
floating for the range of operating frequencies of this circuit. It is true that the gate
needs to be reset every few milliseconds; however, that should not be a problem as
in that time the circuit is able to perform hundreds of thousands of operations.
8.4.3.3 νMOS sorters for arithmetic applications

The threshold functions produced at the outputs of the sorter circuit are involved in
many arithmetic-like operations. This is illustrated below with two examples. The
first example refers to the implementation of an (8 × 8) multiplier, while the second
one describes a (15,4) counter which is used in the summation of the partial products
in a parallel multiplier.
The (8 × 8) multiplier:
The (8 × 8) multiplier is based on the architecture shown in Fig. 8.12, originally

proposed in [254]. The main component of this design, apart from the peripheral cir-
cuitry necessary for data scheduling, is a combinational functional block (F_Block)
with 16 inputs and 9 outputs. Eight of the outputs correspond to threshold func-
tions T216 , T416 , T616 , T816 , T10
16 , T 16 , T 14 and T 16 . The ninth is the parity function. The
12 16 16
F_Block circuit is realised using a two-level network of capacitive TGs (17 gates).
The F_Block described here uses the νMOS sorter circuit as the key component.
Figure 8.13 shows the logic diagram. It consists only of a 16-input sorter, T 16 , and a
TG, T916 , realised using the ideas explained in the previous section. The output of the
TG implements the parity function following the Muroga’s method [255] as
parity = T116 − T216 + T316 − · · · − T14

16 16
+ T15 16
− T16 (8.24)
Serial Data Parallel Data

CLK
Input Input T16
2
8 T16
4
& T16
6
CLK 8 bit
Shift 8 8 T16 Register
F_Block 8
Register
T 16
10
T 16
12
8
T 16
14
T 16
16
parity
Serial Data
Output
Figure 8.12 Serial/Parallel multiplier architecture
T 16
1
T 16
2
T 16
3
T 16
4
16 input parity
sorter T 16
9
16 (T16)
16
T13
16
T14
16
T15
16
T16
Figure 8.13 Logic diagram for the F_Block
The F_Block circuit using the νMOS sorter is designed and laid out in a 0.8 μm
5 V double poly CMOS process [214]. Figure 8.14 shows the simulated wave-
forms for the parity output of the extracted F_Block. The inputs correspond to a
sequence of input patterns with an increasing number of ones: (x1 , x2 , . . . , x16 ) =
{(0, 0, . . . , 0), (0, 0, . . . , 1), . . . , (1, 1, . . . , 1)} starting at time t = 60 ns. A new pat-
tern is applied each 7.5 ns. Clearly, the parity of the 16 input signals is correctly
evaluated.
Transient characteristics and average power are measured on post-layout simu-
lation results using typical device parameters at a supply voltage of 5 V. The power
is measured using a randomly generated input sequence with 100 vectors. The worst
case delay time is 4.5 ns and the power consumption is 13 mW at 100 MHz.
In order to validate the proposed circuit, its performance is compared with other
solutions. Simulation results for the threshold-gate-based implementation of the
Wave Symbol 5
v(outs)
3
Voltage (V)
60 n Time (s) 180 n
Figure 8.14 HSPICE simulation results for the parity output of F_Block
architecture in Fig. 8.12 [254] show a maximum clock frequency of around 30 MHz
for the multiplier when implemented in a 1.2 μm technology. On the other hand, the
multiplier based on a sorter circuit can work at frequencies in excess of 175 MHz.
This is because the clock frequency of the former is mainly limited by the signal
propagation through the F_Block.
Also, in order to compare these two approaches, the F_Block is designed and laid
out following a conventional approach (NOR and NAND gates were used) using the
same technological process. The worst case delay for this conventional design is over
11 ns and the power consumption at 66 MHz is 13 mW. Besides, it occupies an area
between one and two orders of magnitude larger.
The (15,4) counter:
A second example illustrates a possible application of the sorter as a counter. This

section describes the design methodology of a (15,4) counter.
A counter is a combinational circuit which outputs a binary number equal to the
number of input lines that are at logical one. The summation of partial products in a
parallel multiplier is traditionally performed using a full adder tree (full adders are
a particular case of counters, the (3,2) counter). However, because the routing can
be complicated, high-order counters are preferred. High-order counters are usually
implemented using (3,2) counters because of the disadvantages of a direct implemen-
tation [250]. With the approach described a counter can be built directly from its logic
equations. Let (x0 , x1 , . . . , x14 ) be the fifteen numbers to add in a (15,4) counter, and
(y3 , y2 , y1 , y0 ) be the counter output. Signals y3 , y2 , y1 and y0 are symmetric functions
that can be related to the sorter outputs as

y0 = T115 · T215 + T315 · T415 + T515 · T615 + T715 · T815
+ T915 · T10
15 15
+ T11 15
· T12 15
+ T13 15
· T14 15
+ T15
y1 = T215 · T415 + T615 · T815 + T10
15 15
· T12 15
+ T14 (8.25)
y2 = T415 · T815 15
+ T12
y3 = T815
The implementation of these equations can be optimised by implicit computations,
using arithmetic operators. This can be done in only one level of threshold logic, as
shown in Fig. 8.15. Outputs y0 , y1 and y2 are implemented through functions T815 , T47
and T23 , realised as TGs:
y0 = T115 − T215 + T315 − T415 + T515 − T615 + T715 − T815 + T915 − T10
15
15 15 15 15 15
+ T11 − T12 + T13 − T14 + T15
y1 = T215 − T415 + T615 − T815 + T10
15 15
− T12 15
+ T14 (8.26)
y2 = T415 − T815 15
+ T12
y3 = T815
Two (15,4) counters, one using the νMOS sorter circuit and νMOS TGs and
another one following a conventional approach, are designed and laid out in the same
technological process. Transient characteristics and average power are measured.
T 15
1
T 15
2 T 415 T 15 y0
8
T 15
3
15 input
sorter
15 (T 15 ) T 815
T 15
4
T 47 y1
15
15
T 12
T 14
15
T 15
T 15
4
T 23 y2
T 15
12
y3
Figure 8.15 Logic diagram of the (15,4) counter

The worst case delay time for the νMOS solution is 8 ns and the power consumption
is 12 mW at 66 MHz and it is very independent of the frequency. The conventional
design has the same power consumption at 66 MHz, but the worst case delay is
11.25 ns.
8.4.4 A multi-input Muller C-element

A Muller C-element (where the C stands for concurrence) is a circuit widely used in
the design of self-timing circuits to perform the function ‘and’ of events (transitions
1 → 0 or 0 → 1). The output of a Muller C-element is equal to the value of the input
after all the inputs reach the same value; otherwise, it remains the same. It has been
proven [235] that an m-input Muller C-element can be implemented using a single
threshold gate with (m + 1) inputs, and the simplest solution is obtained when the
primary inputs have an associated weight of 1, the (m + 1)-th input (the feedback
input) is affected by a weight of (m − 1), and the threshold of the TG is m, as shown
in Fig. 8.16(a).
The complexity of this logic element is high enough to use it as a demonstrator
of a design approach based on νMOS-TG. On one hand, it shows that a complex
functionality can be implemented by a single inverter. On the other hand, the existence
of a feedback loop with a high associated weight (m − 1) allows the testing of the
signal regeneration capability that would guarantee the correct operation of threshold
networks.
Figure 8.16(b) depicts the νMOS realisation of the logic diagram shown in
Fig. 8.16(a), for m = 8. This circuit is designed and fabricated in a 5 V 0.8 μm
double poly CMOS technology. Operation under process and ambient parameters
was validated through extensive HSPICE simulations of the extracted circuit includ-
ing Monte Carlo simulations and simulations using different standard worst case
device parameters. Figure 8.17 shows the experimental responses. The waveform at
VDD
C ( 115 f F)
V x1
4/2
U: [1,1,...,1,m-1;m] Vout
x1 V x8
x2 Q
TG
xm 3.5/2
7C
(a) (b)
Figure 8.16 (a) Threshold-gate-based m-input Muller C-element realisation;

(b) Electrical diagram of a νMOS-based 8-input Muller C-element
(transistor sizes and capacitance values are shown in the figure)
0
0
0
0
0
1
0
1
Figure 8.17 Experimental waveforms of the νMOS-based 8-input Muller C-element
Table 8.3 Performance parameters for νMOS and conven-

tional 8-input Muller C-elements
Area Worst case delay Power consumption

(@100 MHz)
νMOS 4927 μm2 1.8 ns 0.03 mW

Conventional 9942 μm2 3.2 ns 0.54 mW
the bottom trace is the circuit output and the remaining waveforms correspond to the
circuit input signals. Correct operation is obtained with a supply voltage down to 3 V.
A conventional 8-input Muller C-element based on the efficient structure pro-
posed in [256] is also designed and laid out for the purpose of comparison. Table 8.3
compares the area, transient performance and power consumption of both Muller
C-elements. The νMOS design performs much more efficiently than a conventional
gate-based implementation. It occupies half the area occupied by its conventional
counterpart, has better transient performance and consumes significantly less power.
The realisation of certain digital functions is greatly simplified by using TGs as build-
ing blocks. The implementation of a TG using the νMOS transistor has been presented
in this chapter. The νMOS TG is much simpler than other reported TG realisations.
All this makes a design style based on combining the TG gate approach with the
implementation of this building block using νMOS devices markedly interesting.
The use of TGs as well as νMOS devices reduces the overall power × delay figure of
merit of certain digital systems. This chapter has illustrated how to achieve this with
several design examples. The reader can extend the illustrated design approach much
further, to other digital systems that operate with even tighter design constraints in
terms of the power consumption and voltage supply, and also adapt these circuits to
new improved technological processes.
Notation
Delay of the final sum bits in an n-bit CLA (eq. (8.14)). Also
delay difference between two realisations, traditional and
hierarchical, of an n-bit adder and b-bit BCLA given by
eq. (8.21)
BCLA Circuit delay of a carry lookahead unit block (eq. (8.16))
CLA Delay of a CLA block (eqs. (8.14), (8.16) and (8.20))
PG Delay of an n-bit CLA due to carry generate/propagate unit
generating Pi and Gi signals (eqs. (8.14) and (8.16))
S Delay corresponding to the generation of Si sum signals
(eqs. (8.14) and (8.16))
TG Delay of the n-bit adder in a hierarchical approach (eq. (8.20))
TG
BCLA Circuit delay of the BCLA unit based on TGs (eq. (8.20))
δ Number of delay units per level (eq. (8.21))
R Switch signal in the n-input sorter in Fig. 8.10 (has the value 1
during initialisation and the value 0 during the processing
mode)
A= Augend input to an n-bit CLA
An−1 , . . . A1 , A0
B= Addend input to an n-bit CLA
Bn−1 , . . . B1 , B0
BCLA(j, k) 4-Bit BCLA unit corresponding to a generic j-th column and
k-th row in an adder three implementation (Fig. 8.4 and
Fig. 8.5)
C Input capacitances (connected to Vxi ) of the νMOS-based
8-input Muller C-element in Fig. 8.16(b)
C−1 Carry input to the least significant position of an n-bit CLA
Cchan Sum of M2 parasitic capacitances in the n-input sorter in
Fig. 8.10
CControl Control input capacitance in Fig. 8.3
Ci Input capacitance corresponding to the input xi in Fig. 8.3
(eq. (8.2)). Also used in a different section to refer to the carry
output of the i-th stage in an n-bit CLA obtained using Gi , Pi
and Ci−1 (see eq. (8.12))
Ci−1 Carry input to the i-th bit position of an n-bit CLA
Co M2 input capacitance connected to ground in the n-input sorter
in Fig. 8.10
CT Total capacitance seen by the FG (eq. (8.2)). The same value

has been assumed for all the devices unless the opposite is said
Cu Input capacitances connecting input gates xi of M2 (sorter
inputs) and its FG (Fig. 8.10)
c Extra control input of the νMOS TG in Fig. 8.3 used for the
logic threshold adjustment
Fk : Fk−1 + Fk−2 k-th Fibonacci’s number
f (x) Threshold function defined by [w; T ] where
w = [w1 , w2 , . . . , wn ] (Fig. 8.2)
f (x) Complement of a threshold function defined by [−w; 1 − T ]
(Fig. 8.2)
f (x̄)
: [−w; T − Threshold function of complemented inputs (Fig. 8.2)
wi ]
˙
f (x̄)
: [w; 1 − T + Complement of a threshold function of complemented inputs
wi ] (Fig. 8.2)
Gi Carry generate function (eq. (8.12))
Oi for i = [1, n] Outputs of the sorting network
Pi i-th Carry propagate function (eq. (8.12))
Si i-th Sum of the bits in an n-bit CLA (eq. (8.12))
T TG threshold value (eq. (8.1))
T1 Carry-out signal of a (4,2) compressor (Fig. 8.7)
T 16 16-Input sorter (Fig. 8.12)
Ti16 for
i = [2, 4, 6, 8,
10, 12, 14, 16] Outputs of the F_Block in Fig. 8.13 corresponding to the
threshold functions
TGi for i = [1, 5] TGs used for implementing the (4,2) compressor (Fig. 8.7)
t1 Carry-in signal to the (4,2) compressor (Fig. 8.7)
Uinv(I1 ) Switching threshold voltage of the inverter I1 in Fig. 8.10
Uinv(IO1 ) Switching threshold voltage of the inverter IO1 in Fig. 8.10
Uinv Switching threshold voltage of the inverter (νMOS TG) in
Fig. 8.3
Vc Voltage applied to the control input c of the νMOS threshold
gate in Fig. 8.3
VFG FG voltage at the FG of the νMOS TG in Fig. 8.3
Vxi : xi VDD Voltage applied to the input xi of the νMOS TG in Fig. 8.3
(eq. (8.2))
V0 M2 drain voltage in the first block of the n-input sorter in
Fig. 8.10
V1 Analogue output voltage of the first block in the n-input sorter
in Fig. 8.10
Vout Output of the νMOS-based 8-input Muller C-element in
Fig. 8.16(b)
Vxi for i = [1, 8] Inputs of the νMOS-based 8-input Muller C-element in

Fig. 8.16(b)
WT (Cj ) Total weight for the TG implementing Cj given by eq. (8.19)
wi i-th Input weights (eqs. (8.1) and (8.4))
xi for i = [1, n] Inputs
yi i-th Output signals of the (4,2) compressor (Fig. 8.6). Also
counter outputs (Fig. 8.15)
Chapter 9
Summary and conclusions
This chapter summarises the most relevant results and conclusions that have been
presented throughout the book.
Chapter 2 focused on the characterisation and modelling of the FGMOS transistor.
From the analysis introduced in this chapter, the FGMOS transistor can be seen as a
MOS transistor with the following characteristics:
(1) Multiple inputs (as many as there are input capacitors).
(2) Output resistance strongly dependent on the ratio between the capacitive cou-
pling between the FG and the drain terminal and the total capacitance seen by
the FG. The smaller this ratio, the larger the output resistance is.
(3) Reduced effective transconductance, which is scaled down by the effective input
weight.
The first characteristic can be considered as an advantage with respect to the normal
MOS transistor, whereas the reduction in the output resistance and transconductance
may, in certain situations, represent drawbacks.
The second part of the chapter described some of the initial issues that designers
encounter when starting to work with FGMOS devices, as well as solutions to them.
The main two problems observed were as follows:
(1) The foundries do not provide models to simulate the FGMOS.
(2) There is an unknown amount of charge that can remain trapped at the FG during
the fabrication process.
A very simple technique that does not require special device modelling to simulate
the device was explained and compared with other existing methods that do need to
model the transistor beforehand. Also, different techniques to control the charge at
the FG were reviewed, together with a very simple layout technique, consisting of
adding a contact to metal on the FG, which removes the initial trapped charge.
274 Summary and conclusions
One modification to the FGMOS transistor for which the ‘initial charge’ is not
a problem is the pseudo-floating-gate MOS device. This is an FGMOS device with
a very high value resistor connected to the gate, filtering out low frequency signals.
Since the trapped charge is equivalent to an initial DC condition, in this way it becomes
irrelevant.
One of the main negative points of FGMOS transistors is the area required by the
added input capacitors. Hence, ideally, these capacitors should be designed as small
as possible. The problem is that the smaller these are, the higher the ratios between
the parasitic capacitances and the total capacitance seen by the FG become. This may
affect the performance of the circuits in different ways (as was shown in following
chapters). Chapter 2 provided an analysis on how to choose the values of the input
capacitances and an estimate of the overhead in terms of area.
Chapter 3 introduced basic circuit blocks that illustrated some of the interesting
features of the FGMOS transistor. The most important ones are as follows:
(1) The threshold voltage is controllable. This property can be used whenever the
voltage requirements in a given design are too restrictive for a given nominal
threshold voltage and a certain power supply voltage.
(2) The transistor is a powerful computational element. It performs a weighted sum
of its inputs in a high impedance node. The coefficients can be chosen by sizing
the input capacitances appropriately.
(3) The operating point at the gate of the transistor can be shifted as required by the
circuit topology without the need for adding extra circuitry.
(4) The transistor biased in the weak inversion saturation region can behave as a
current controlled current source79 , which is a nonlinear function between the
currents, with exponents dependent on the ratios between the input capacitances.
In addition, the gain of the current source can be programmed by means of
voltage sources connected to the transistor inputs.
The chapter also shows how to reduce the minimum required power supply voltage
in a cascode current mirror with FGMOS transistors by using the voltage supply in
one of the inputs to bias the devices and sizing the input capacitances accordingly.
Another interesting simple block is the FGMOS inverter in which an input can
be used to program the switching threshold.
In an FGMOS comparator the versatility of the device was illustrated with four-
input FGMOS devices in which each input was used for a different aim: effective
input, offset cancellation, compensation of the degradation in output resistance caused
by the gate to drain parasitic capacitance and reset.
The FGMOS transistor used in a switched current memory cell permits to program
the cell in a wide range without the need of any other extra circuitry.
Finally some ideas for the implementation of very simple D/A converters making
use of the summation property at the FG were also presented.
79 If its inputs are voltages generated by currents flowing through other transistors.
Summary and conclusions 275
Chapters 4 and 5 demonstrated the potential of FGMOS devices biased in the

strong inversion region for the design of low power and low voltage programmable
analog Gm -C filters. The designs presented showed how:
(1) The use of the FGMOS transistor simplified the designs enormously due to the
addition of the input voltages at the FG.
(2) An immediate consequence of (1) is that fewer devices are needed, which means
fewer internal nodes. This is advantageous from the point of view of frequency
response.
(3) The noise is lower as the circuit topologies get simpler.
(4) The voltage supply can be scaled down by optimising the voltage ranges at the
devices. This can be achieved by shifting down the effective threshold voltage
with the help of one of the transistor’s inputs. The input range is also increased
this way, due to the attenuation of the effective input voltage at the FG.
(5) The power consumption decreases as a consequence of the lower power supply
voltage and the smaller number of active devices.
(6) If the transistor operates in the strong inversion ohmic region the Total Harmonic
Distortion is lower than in a MOS transistor with equivalent voltage swings at the
gate. The reason for this is that the CGD term appears subtracted in the quadratic
VDS term, which reduces the nonlinear contribution to the current function.
There are also some drawbacks in using the device, which can be summarised as
follows:
(1) As was mentioned before, the output resistance is reduced because of the capac-
itive coupling between the FG and the drain terminal. A possible way to deal
with this if necessary is to add cascode transistors.
(2) If the effective threshold voltage is scaled down by using extra inputs connected
to the voltage supplies the PSRR degrades. It also becomes dependent on the
mismatch between the capacitances.
(3) The CMRR is affected by the mismatch between the input capacitances.
(4) The transconductance and the frequency range are lower than the transcon-
ductance and frequency range that could be obtained with equally sized MOS
transistors driving the same values of currents. However, whether this causes
a problem or not depends on the application. Also, in general, assumption of
having a similar circuit only with MOS transistors is not realistic since using
the FGMOS transistor only makes sense whenever normal MOS cannot be
used to obtain the required performance in terms of voltage, power and signal
ranges.
(5) Although a double poly technology is not strictly required, if low qual-
ity capacitors are used instead of the FGMOS the circuit performance will
degrade.
Chapters 6 and 7 explored the design of low power and low voltage Gm -C and
log-domain filters, taking advantage of the functionality of the FGMOS transistor
biased in the weak inversion saturation region and using the Translinear Principle,
276 Summary and conclusions
which was reformulated for the FGMOS. The most relevant conclusions extracted
from these chapters were as follows:
(1) Translinear loops created with FGMOS devices simplify the circuit topologies.
This brings with it a reduction in the number of internal nodes, which improves
the frequency response, and also lowers noise and power consumption.
(2) The bulk and source terminal do not need to be used to implement the translinear
equation. Because of this they can be connected to constant voltages, which
reduces the risk of instabilities.
(3) There is no need of stacking transistors to implement the translinear equation,
which is beneficial for low power supply voltage operation.
(4) There are many possible alternatives when programming the design, thanks to
the controllability property of the FGMOS device.
The main disadvantage of using FGMOS in weak inversion are as follows:
(a) The mismatch between input capacitances together with the parasitic coupling
between the drain and the FG affects the Total Harmonic Distortion.
(b) In log-domain implementations, the signal is not as compressed in the FG as
when MOS transistors are used instead because of the smaller than one capacitive
input weight.
Finally, Chapter 8 moved away from analog design, illustrating the advantages of
using FGMOS devices in digital circuits. The chapter showed how logic circuit design
based on Threshold Gates implemented with FGMOS transistors is an alternative to
traditional logic gate design. Using threshold gates as main building blocks in a system
realisation enables a reduction in both the number of gates and the gate levels, with
the consequent saving in power consumption, mainly for high-frequency operation.
So, where do we go from here? Although this book has a chapter on digital design
and reviews the main issues related to power and voltage reduction in digital circuits,
it is mainly a text on analog design. As many have admitted before, analog design is
an art based on the combination of the different degrees of freedom in the transistors’
behavioural equations. If properly combined, the result is a beautiful piece that meets
the consumers’ specifications. The FGMOS transistor is like a wonderful new colour
palette, a transistor with added degrees of freedom. This opens up a whole new world
of possibilities for the designer. The specifications are the same, but the limitations on
how to achieve them are fewer. This book has shown just a tiny sample of topologies
that can be created by playing with the extra degrees of freedom, achieving a far
superior performance than the one that could have been obtained with normal MOS
devices, mostly in terms of voltage ranges and power. There are still many other
circuit topologies to invent and explore in the context of low power and low voltage
electronics. Also, due to the problem of the trapped charge at the FG (which has only
recently been solved), the device has not been widely used. As a consequence, not
many electronic systems had been reported that made use of the device in a context
other than memories. It is now time to start exploiting the full potential of the transistor
in more complex systems.
Summary and conclusions 277
A very interesting property of the FGMOS transistor is that theoretically each

device can be independently controlled. This could be a very promising possibility
in certain applications that require a high level of programmability, tunability or
calibration. An example could be making robust the circuit drivers in polysilicon
technologies for TFT (Thin Film Transistors) displays. The main problem of this
kind of technologies is related to the large variations among the devices. A calibration
mechanism could be devised to adjust the drivers by changing the bias voltages in
one of the inputs of TFT devices with floating gates. Another example is low power
Analog Field Programmable Arrays (FPAA), in which different analog blocks could
be programmed by changing the bias voltages in certain FGMOS devices.
In the digital arena, many new topologies, different from the ones presented here,
could be invented to take advantage of compact threshold gate topologies built with
the FGMOS device, based on the threshold decision principle.
A point worth stressing is that the basic property of the FGMOS transistor, its
floating gate, may also be its most important limitation in any future technologies in
which a leaky gate oxide might render the transistor useless for certain applications.
In this case, the pseudo-FGMOS described in Chapter 2 could be used bearing in
mind the limitations it has with respect to the normal FGMOS, as was previously
explained.
Ultimately, there are plenty of electronic systems that could benefit from exploit-
ing the capabilities of the FGMOS device and it is left to the creative mind of the
circuit designer to decide how to combine all the degrees of freedom and come up
with an optimum design. Having more variables to play with might make the process
of finding the solution more difficult; nevertheless, the benefits make the challenge
worthwhile in the end.
References
1 Vittoz, E. A.: ‘Analog VLSI signal processing: why, where, and how?’, Analog
Integrated Circuits and Signal Processing, 1994 8, pp. 27–44
2 Luryi, S., Xu, J., and Zaslavski, A., (eds.): Future trends in microelectronics.
The road ahead (Wiley-Interscience, 1999) ISBN: 0 471 32183 4
3 Lyon, R. F., and Mead, C.: ‘An analog electronic cochlea’, IEEE Transactions
on Acoustics, Speech and Signal Processing, 1988 36(7), pp. 1119–34
4 Stotts, L. J.: ‘Introduction to implantable biomedical IC design’, IEEE Circuits
Devices Mag., 1989 5 (1), pp. 12–18
5 Von Kaenel, V., Macken, P., and Degrauwe, M. G. R.: ‘A voltage reduction
technique for battery operated systems’, IEEE Journal of Solid-State Circuits,
1990 25, pp. 1136–40
6 Yúfera, A., Rueda, A., Munoz, J. M., Doldan, R., Léger, G., and Rodríguez
Villegas, E. O.: ‘A tissue impedance measurement chip for myocardial ischemia
detection’, IEEE Transactions on Circuits and Sytems I: Fundamental Theory
and Applications’, 2005 52(12), pp. 2620–8
7 Baltes, H., and Brand, O. (invited keynote): ‘CMOS integrated microsystems
and nanosystems’. Proceedings of the SPIE Conference on Smart Electronics
and MEMS, 1999 3673, pp. 2–10
8 ‘The National Technology Roadmap for Semiconductors. Technology Needs’,
Technical report, Semiconductor Industry Association, 2000.
9 El-Adawy, A. A., and Soliman, A. M.: ‘A low-voltage single input class AB
transconductor with rail-to-rail input range’, IEEE Transactions on Circuits and
Systems-I, 2000 47, pp. 236–42
10 Duque-Carrillo, J. F., Ausín, J. L., Torelli, G., Valverde, J. M., and
Domínguez, M. A.: ‘1V rail-to-rail operational amplifiers in standard CMOS
technology’, IEEE Journal of Solid State Circuits, 2000 35, pp. 33–44
11 Lin, C., and Ismail, M.: ‘Robust design of LV/LP low-distortion CMOS rail-to-
rail input stages’, Analog Integrated Circuits and Signal Processing, 1999 21,
pp. 153–61
12 Ferri, G., Sansen, W., and Peluso, V.: ‘A low-voltage fully-differential
constant-Gm rail-to-rail CMOS operational amplifier’, Analog Integrated Cir-
cuits and Signal Processing, 1998 16, pp. 5–15
280 References
13 deLangen, K. J., and Huijsing, J. H.: ‘Compact low-voltage power-efficient

operational amplifier cells for VLSI’, IEEE Journal of Solid State Circuits, 1998
33, pp. 1482–96
14 Gregorian, R., and Temes, G. C.: Analog MOS integrated circuits for signal
processing (Wiley-Interscience, 1986) ISBN: 0 471 09797 7
15 Setty, S., and Toumazou, C.: ‘N-folded cascode technique for high frequency
operation of low voltage opamps’, IEE Electronics Letters, 1996 32, pp. 955–7
16 Xu, G., Embabi, S. H. K., Hao, P., and Sanchez-Sinencio, E.: ‘A low voltage fully
differential nested Gm capacitance compensation amplifier: analysis and design’,
Proceedings of the IEEE International Symposium on Circuits and Systems, 1999
2, pp. 606–9
17 Wu, J., and Chang, K.: ‘MOS charge pumps for low-voltage operation’, IEEE
Journal of Solid State Circuits, 1998 33, pp. 592–7
18 Duister, T. A. F., and Dijkmans, E. C.: ‘A -90dB THD rail-to-rail input opamp
using a new local charge pump in CMOS’, IEEE Journal of Solid State Circuits,
1998 33, pp. 947–55
19 St. Pierre, R.: ‘Low-power BiCMOS opamp with integrated current mode charge
pump’. Proceedings of the IEEE European Solid-State Circuits Conference, 1999
pp. 70–3
20 Shin, J., Chung, I., Park, Y., and Min, H.: ‘A new charge pump without degra-
dation in threshold voltage due to body effect’, IEEE Journal of Solid State
Circuits, 2000 35, pp. 1227–30
21 Tsividis, Y.: ‘Externally linear, time-invariant systems and their application to
companding signal processors’, IEEE Transactions on Circuits and Systems-II,
1997 44, pp. 65–85
22 Frey, D., Tsividis, Y. P., Efthivoulidis, G., and Krishnapura, N.: ‘Syllabic-
companding log domain filters’, IEEE Transactions on Circuits and Systems-
II: Analog and Digital Signal Processing, 2001 48 (4), pp. 329–39
23 Mulder, J., Serdijn, W. A., van der Woerd, A. C., and van Roermund, A. H. M.:
‘An instantaneous and syllabic companding translinear filter’, IEEE Transac-
tions on Circuits and Systems-II, 1998 45, pp. 150–54
24 Tsividis, Y.: ‘Minimizing power dissipation in analogue signal processors
through syllabic companding’, IEE Electronics Letters, 1999 35, pp. 1805–7
25 Khoury, J. M.: ‘On the design of constant settling time AGC circuits’, IEEE
Transactions on Circuits and Systems-II, 1998 45, pp. 283–94
26 Toumazou, C., Ngarmnil, J., and Lande, T. S.: ‘Micropower log-domain filter
for electronic cochlea’, IEE Electronic Letters, 1994 30(22), pp. 1839–41
27 El-Masry, E. I., and Wu, J.: ‘CMOS micropower universal log-domain biquad’,
IEEE Transactions on Circuits and Systems I, 1999 46, pp. 389–92
28 Fragnière, E., and Vittoz, E.: ‘A log-domain CMOS transcapacitor: design,
analysis and applications’, Analog Integrated Circuits and Signal Processing,
2000 22, pp. 195–208
29 Fox, R. M., and Nagarajan, M.: ‘Multiple operating points in a CMOS log-
domain filter’, IEEE Transactions on Circuits and Systems-II: Analog and
Digital Signal Processing, 1999 46(6), pp. 705–10
References 281
30 Degrauwe, M. G., Rijmenents, J., Vittoz, E. A., and DeMan, H. J.: ‘Adap-
tive biasing CMOS amplifiers’, IEEE Journal of Solid State Circuits, 1982 17,
pp. 522–8
31 Klinke, R., Hosticka, B. J., and Pfleiderer, H. J.: ‘A very-high-slew-rate CMOS
operational amplifier’, IEEE Journal of Solid State Circuits, 1989 24, pp. 744–6
32 Vittoz, E., and Fellrath, J.: ‘CMOS analog integrated circuits based on weak
inversion operation’, IEEE Journal of Solid State Circuits, 1977 12, pp. 224–31
33 Horowitz, M., Indermaur, T., and Gonzalez, R.: ‘Low-power digital design’.
Proceedings of the IEEE Symposium on Low Power Electronics, 1994 pp. 8–11
34 Chandrakasan, A., Sheng, S., and Brodersen, R. W.: ‘Low-power CMOS digital
design’, IEEE Journal of Solid-State Circuits, 1992 27, pp. 473–84
35 Indermaur, T., and Horowitz, M.: ‘Evaluation of charge recovery circuits and
adiabatic switching for low power CMOS design’, Symposium on Low-Power
Electronics, 1994 pp. 102–3
36 Hasler, P., and Lande, T. S.: ‘Overview on floating-gate devices, circuits, and
systems’, IEEE Transactions on Circuits and Systems II: Analog and Digital
signal Processing, 2001 48
37 Kahng, D., and Sze, S. M.: ‘A floating-gate and its application to memory
devices’, Bell Syst. Tech. J., 1967 46(4), pp. 1288–95
38 Holler, M., Tam, S., Castro, H., and Benson, R.: ‘An electrically trainable
artificial neural network with 10240 ‘floating gate’ synapses’, Proceedings of
the IEEE International Joint Conference on Neural Networks, Vol. II, 1989,
pp. 191–6
39 Lai, S.: ‘Flash memories: where we were and where we are going’, Proceedings
of the IEEE International Electron Devices Meeting, 1998, pp. 971–4
40 Sackinger, E., and Guggenbuhl, W.: ‘An analog trimming circuit based on
a floating-gate device’, IEEE Journal of Solid-State Circuits, 1988 23(6),
pp. 1437–40
41 Yu, C. G., and Geiger R. L.: ‘Very low voltage operational amplifiers using
floating gate Mos transistor’. Proceedings of the IEEE International Symposium
on Circuits and Systems, Vol. II, May 3–6 1993, pp. 1152–55
42 Lenzlinger, M., and Snow, E. H.: ‘Fowler-Nordheim tunneling into thermally
grown SiO2 ’, Journal of Applied Physics, 1969 40(1), pp. 278–83
43 Thomsen, A., and Brooke, M. A.: ‘A floating gate MOSFET with tunneling
injector fabricated using a standard double-polysilicon CMOS process’, IEEE
Electronics Device Letters, 1991 12(3), pp. 111–13
44 Op’t Eynde, F., and Zorio, C.: ‘An EEPROM in a standard CMOS tech-
nology’. Proceedings of the 23rd European Solid-State Circuits Conference,
Southampton, UK, Sep.–Oct. 1997
45 Lanzoni, M., Briozzo, L., and Riccò, B.: ‘A novel approach to controlled
programming of tunnel-based floating-gate MOSFET’s’, IEEE Journal of
Solid-State Circuits, 1994 29(2), pp. 147–50
46 Gao, W., and Snelgrove, W. M.: ‘Floating gate charge-sharing: a novel circuit for
analog trimming’. Proceedings of the IEEE International Symposium on Circuits
and Systems, 1994, pp. 315–18
282 References
47 Thomsen, A., and Brooke, M. A.: ‘Low control voltage programming of floating
gate MOSFETs and applications’, IEEE Transactions on Circuits and Systems, I:
Fundamental Theory and Applications, 1994 41(6), pp. 443–52
48 Lanzoni, M., Tondi, G., Galbiati, P., and Ricco, B.: ‘Automatic and contin-
uous offset compensation of MOS operational amplifiers using floating-gate
transistors’, IEEE Journal of Solid-state Circuits, 1998 33, pp. 287–90
49 Jackson, S. A., Killens, J. C., and Blalock, B. J.: ‘A programmable current mirror
for analog trimming using single-poly floating-gate devices in standard CMOS
technology’, IEEE Transactions on Circuits and Systems II: Analog and Digital
Signal Processing, 2001 48, pp. 100–6
50 Shibata, T., and Ohmi, T.: ‘A functional MOS transistor featuring gate-
level weighted sum and threshold operations’, IEEE Transactions on Electron
Devices, 1992 39(6), pp. 1444–55
51 Shibata, T., and Ohmi, T.: ‘Neuron MOS binary-logic integrated circuits- Part I:
design fundamentals and soft-hardware-logic circuit implementations’, IEEE
Transactions on Electron Devices, 1993 40(3), pp. 570–6
52 Shibata, T., and Ohmi, T.: ‘Neuron MOS binary-logic integrated circuits- Part II:
simplifying techniques of circuit configuration and their practical applications’,
IEEE Transactions on Electron Devices, 1993 40(5), pp. 974–9
53 Shibata, T., Kotani, K., and Ohmi, T.: ‘Real-time reconfigurable logic circuits
using Neuron MOS transistors’, ISSCC Tech. Dif., 1993, pp. 238–9
54 Hany, T., Teranishi, K., and Kameyama, M.: ‘Multiple-valued logic-in-memory
VLSI based on a floating-gate-MOS pass-transistor network’, Proceedings of
the IEEE International Solid-State Circuits Conference, 1998, pp. 194–5
55 Kaeriyama, S., Hanyu, T., and Kameyama, M.: ‘Arithmetic-oriented multiple-
valued logic-in-memory VLSI based on current-mode logic’, Proceedings of the
30th IEEE International Symposium on Multiple-Valued Logic, 2000, pp. 438–43
56 Hanyu, T.: ‘Challenge of a multiple-valued technology in recent deep-submicron
VLSI’. Proceedings of the 31st IEEE International Symposium on Multiple-
Valued Logic, 2001, pp. 241–4
57 Hanyu, T., Kameyama, M., Shimabukuro, K., and Zukeran, C.: ‘Multiple-valued
mask-programmable logic array using one-transistor universal-literal circuits’,
Proceedings of the 31st IEEE International Symposium on Multiple-Valued
Logic, 2001, pp. 167–72
58 Weber, W., Prange, S. J., Thewes, R., Wohlrab, E., and Luck, A.: ‘On the
application of the neuron MOS transistor principle for modern VLSI design’,
IEEE Transactions on Electron Devices, 1996 43(10)
59 Fujita, O., and Amamiya, Y.: ‘A floating-gate analog memory device for neural
networks’, IEEE Transactions on Electron Devices, 1993 40(11), pp. 2029–35
60 Shima, T., and Rinnert, S.: ‘Multiple-valued memory using floating gate
devices’, IEICE Transaction on Electronics, 1993 E76-C(3), pp. 393–402
61 Kosaka, H., Shibata, T., Ishii, I., and Ohmi, T.: ‘An excellent weight-updating-
linearity EEPROM synapse memory cell for self-learning neuron-MOS neural
networks’, IEEE Transactions on Electron Devices, 1995 42(1), pp. 135–43
References 283
62 Shibata, T., Kosaka, H., Ishii, H., and Ohmi, T.: ‘A neuron-MOS neural network
using self-learning-compatible synapse circuits’, IEEE Journal of Solid-State
Circuits, 1995 30(8), pp. 913–22
63 Card, H. C., and Moore, W. R.: ‘EEPROM synapses exhibiting pseudo Hebbian
plasticity’, IEE Electronic Letters, 1989 25, pp. 805–6
64 Lee, B. W., Sheu, B. J., and Yang, H.: ‘Analog floating gate-synapses for general
purpose VLSI neural computation’, IEEE Transactions on Circuits Systems,
1991 38, pp. 654–7
65 Durffee, C. A., and Shoucair, F. S.: ‘Comparison of floating gate neural net-
work memory cells in standard VLSI CMOS technology’, IEEE Transactions
on Neural Networks, 1992 3(3), pp. 347–53
66 Hasler, P., Diorio, C., Minch, B. A., and Mead, C.: ‘Single transistor learn-
ing synapse with long term storage’, Proceedings of the IEEE International
Symposium on Circuits and Systems, 1995, pp. 1660–3
67 Diorio, C., Hasler, P., Minch, B. A., and Mead, C.: ‘A single-transistor
silicon synapse’, IEEE Transactions on Electron Devices, 1996 43(11),
pp. 1972–80
68 Hasler, P., Minch, B. A., Diorio, C., and Mead, C.: ‘An autozeroing ampli-
fier using pFET hot-electron injection’, Proceedings of the IEEE International
Symposium on Circuits and Systems, 1996 3, pp. 325–9
69 Diorio, C., Hasler, P., Minch, B. A., and Mead, C.: ‘A complementary pair
of four-terminal silicon synapses’, Analog Integrated Circuits and Signal
Processing, 1997 13, pp. 153–66
70 Diorio, C., Hasler, P., Minch, B. A., and Mead, C. A.: ‘Floating-gate MOS
learning array with locally computed weight updates’, IEEE Transactions on
Electronics Devices, 1997 44, pp. 2281–9
71 Hasler, P., Minch, B. A., and Diorio, C.: ‘An autozeroing floating-gate band-
pass filter’. Proceedings of the IEEE International Symposium on Circuits and
Systems, 1998
72 Hasler, P., Diorio, C., and Minch, B. A.: ‘Continuous-time feedback in floating-
gate MOS circuits’. Proceedings of the IEEE International Symposium on
Circuits and Systems, Vol. 3, Monterey, CA, 1998
73 Apsel, A., Stanford, T., and Hasler, P.: ‘An adaptive front end for olfaction’.
Proceedings of the IEEE International Symposium on Circuits and Systems,
Vol. 3, Monterey, CA, 1998
74 Hasler, P., and Dugger, J.: ‘Correlation learning rule in floating-gate pFET
synapses’. Proceedings of the IEEE International Symposium on Circuits and
Systems, 1999, pp. 387–90
75 Hasler, P., Minch, B. A., and Diorio, C.: ‘Adaptive circuits using pFET floating-
gate devices’. Proceedings of the Advanced Research VLSI Conference, 1999,
pp. 215–31
76 Hasler, P., Kucic, M., and Minch, B. A.: ‘A transistor-only circuit model
of the autozeroing floating-gate amplifier’. Proceedings of the IEEE Midwest
Conference on Circuits and Systems, 1999, pp. 157–60
284 References
77 Hasler, P.: ‘Continuous-time feedback in floating-gate MOS circuits’, IEEE

Transactions on Circuits and Systems II: Analog and Digital Signal Processing,
2001 48, pp. 56–64
78 Hasler, P., Minch, B. A., and Diorio, C.: ‘An autozeroing floating-gate ampli-
fier’, IEEE Transactions on Circuits and Systems II: Analog and Digital Signal
Processing, 2001 48, pp. 74–82
79 Cohen, M., and Cauwenberghs, G.: ‘Floating-gate adaptation for focal-plane
online nonuniformity correction’, IEEE Transactions on Circuits and Systems
II: Analog and Digital Signal Processing, 2001 48, pp. 83–9
80 Kucic, M., Low, A., Hasler, P., and Neff, J.: ‘A programmable continuous-time
floating-gate Fourier processor’, IEEE Transactions on Circuits and Systems II:
Analog and Digital Signal Processing, 2001 48, pp. 90–99
81 Hasler, P., Stanford, T., and Minch, B.: ‘A second-order section built from
autozeroing floating-gate amplifiers’, IEEE Transactions on Circuits and
Systems II: Analog and Digital Signal Processing, 2001 48, pp. 116–20
82 Yang, K., and Andreou, A. G.: ‘A multiple input differential amplifier based
on charge sharing on a floating-gate MOSFET’, Analog Integrated Circuits and
83 Glasser, L. A.: ‘A UV write-enabled PROM’, Chapel Hill Conference on VLSI,
Computer Science Press, (Rockville, MD, 1985) pp. 61–5
84 Kerns, D. A., Tanner, J., Sivilotti, M., and Luo, J.: ‘CMOS UV-writable non-
volatile analog storage’, Sequin, C. H. (ed.) Advanced Research in VLSI, MIT
Press (Cambridge, MA, 1991) pp. 245–61
85 Benson, R. G., and Kerns, D. A.: ‘UV-activated conductances allow for mul-
tiple time scale learning’, IEEE Transactions on Neural Networks, 1993 4(3),
pp. 434–40
86 Yang, K., and Andreou, A. G.: ‘Subthreshold analysis of floating-gate
MOSFET’s’. Proceedings of the 10th Biennial University Government Industry
Microelectronics Symposium, 1993, pp. 141–4
87 Hasler, P., Minch, B., and Diorio, C.: ‘Floating-gate devices: they are not just for
digital memories anymore’, Proceedings of the IEEE International Symposium
on Circuits and Systems, 1999 2, pp. 388–91
88 Minch, B. A., Diorio, C., Hasler, P., and Mead, C. A.: ‘Translinear circuits
using subthreshold floating-gate MOS transistors’, Analog Integrated Circuits
and Signal Processing, 1996 9(2), pp. 167–79
89 Minch, B. A.: ‘Analysis, Synthesis, and Implementation of Networks of
Multiple-Input Translinear Elements’, Ph.D. Thesis, Caltech, Pasadena, CA,
1997
90 Harrison, R., Hasler, P., and Minch, B.: ‘Floating-gate CMOS analog memory
cell array’. Proceedings of the IEEE International Symposium on Circuits and
Systems, 1998, pp. 204–7
91 Harrison, R., Bragg, J. A., Hasler, P., Minch, B. A., and Deweerth, S. P.:
‘A CMOS programmable analog memory-cell array using floating-gate cir-
cuits’, IEEE Transactions on Circuits and Systems II: Analog and Digital Signal
Processing, 2001 48, pp. 4–11
References 285
92 Hasler, P.: ‘Foundations of learning in analog VLSI’, Ph.D. Dissertation,

California Institute of Technology, Pasadena, CA, 1997
93 Hasler, P., Andreou, A. G., Diorio, C., Minch, B. A., and Mead, C. A.: ‘Impact
ionization and hot-electron injection derived consistently from Boltzmann
transport’, VLSI Design, 1998 8(1–4), pp. 455–61
94 Ramírez-Angulo, J., Choi, S. C., and Gonzalez-Altamirano, G.: ‘Low-voltage
OTA architectures using multiple input floating gate transistors’, IEEE Trans-
actions on Circuits and Systems, 1995 42(12), pp. 971–4
95 Ramírez-Angulo, J., Choi, S. C., and González Altamirano, G.: ‘Ultracompact
lo-voltage analog CMOS multiplier using multiple-input floating gate tran-
sistors’, Proceedings of the European Solid-State Circuits Conference, 1996,
pp. 99–103
96 Ramírez-Angulo, J.: ‘A +−0.75 V BiCMOS quadrant analog multiplier with
rail-rail input signal swing’, Proceedings of the IEEE International Symposium
on Circuits and Systems, Atlanta, Georgia, 1996, pp. 242–5
97 Reza Mehrvarz, H., and Yee Kwok, C.: ‘A novel multi-input floating-gate MOS
four-quadrant analog multiplier’, IEEE Journal of Solid-State Circuits, 1996
31(8), pp. 1123–31
98 Chen, J-J., Liu, S-I., and Hwuang, Y-S.: ‘Low-voltage single supply four-
quadrant multiplier using floating-gate MOSFETs’, Proceedings of the IEEE
International Symposium on Circuits and Systems, June 1997, pp. 237–48
99 Wong, L. S., Kwok, C. Y., and Rigby, G. A.: ‘1-V CMOS D/A converter with a
multi-input floating-gate MOSFET’, IEEE Journal of Solid-State Circuits, 1999
34(10), pp. 1386–90
100 Inoue, T., Nakane, H., Fukuju, Y., and Sánchez-Sinencio, E.: ‘A low-voltage
fully-differential current-mode analog CMOS integrator using floating-gate
MOSFETS’, IEEE International Symposium on Circuits and Systems, Vol. IV,
Geneva, June 2000, pp. 145–8
101 Ramírez-Angulo, J., and López, A. J.: ‘MITE circuits: the continuous-time
counterpart to switched-capacitor circuits’, IEEE Transactions on Circuits and
102 Muñoz, F., Torralba, A., Carvajal, R. G., Tombs, J., and Ramírez-Angulo, J.:
‘Floating-gate-based tunable CMOS low-voltage linear transconductor and its
application to HF GM-C filter design’, IEEE Transactions on Circuits and
103 Ramírez-Angulo, J., Carvajal, R. G., Tombs, J., and Torralba, A.: ‘Low-voltage
CMOS op-amp with rail-to-rail input and output signal swing for continuous-
time signal processing using multiple-input floating-gate transistors’, IEEE
Transactions on Circuits and Systems II: Analog and Digital Signal Processing,
2001 48, pp. 111–16
104 Ramírez-Angulo, J., Quintana, V., and Choi, S. C.: ‘Analog VLSI fuzzy proces-
sor architectures based on multiple-input floating gate transistors’, Proceedings
of the World Automation Congress, 1996, pp. 539–44
105 Ramírez-Angulo, J., Quintana, V., Choi, S. C., Zrilic, D., and de Luca, A.:
‘Charge mode defuzzifiers using multiple input floating-gate transistors’,
286 References
Proceedings of the International Journal of Intelligent Automation and Soft

Control, Special Issue on Autonomous Control Engineering, 1997 3(1), pp. 5–12
106 Yu, N. M., Shibata, T., and Ohmi, T.: ‘A real-time center-of mass tracker
circuit implemented by neuron MOS technology’, IEEE Transactions on Cir-
cuits and Systems II: Analog and Digital Signal Processing, 1998 45(4),
pp. 495–503
107 Berg, Y., and Lande, T. S.: ‘Programmable floating-gate MOS logic for low-
power operation’, Proceedings of the IEEE International Symposium on Circuits
and Systems, Hong-Kong, 1997, pp. 1792–5
108 Lande, T. S., and Berg, Y.: ‘Ultra low-voltage transconductance amplifier’, Pro-
ceedings of the IEEE International Conference on Electronics, Circuits and
Systems (ICECS), 1997 1, pp. 333–6
109 Berg, Y., and Lande, T. S.: ‘Ultra low-voltage current mirrors and pseudo dif-
ferential pairs’, Proceedings of the Eleventh Annual IEEE International ASIC
Conference, 1998, pp. 109–14
110 Berg, Y., and Lande, T. S.: ‘Area efficient circuit tuning with floating-gate
techniques’, Proceedings of the IEEE International Symposium on Circuits and
Systems, 1999, pp. 396–9
111 Berg, Y., and Lande, T. S.: ‘Tunable current mirrors for ultra low voltage’,
1999, pp. 17–20
112 Berg, Y., Wisland, D. T., and Lande, T. S.: ‘Ultra low-voltage/low-power dig-
ital floating-gate circuits’, IEEE Transactions on Circuits System, 1999 46(7),
pp. 930–6
113 Berg, Y., and Lande, T. S.: ‘Area efficient circuit tuning with floating-gate
techniques’, Proceedings of the IEEE International Symposium on Circuits and
Systems, 1999 2, pp. 396–9
114 Berg, Y., and Lande, T. S.: ‘Ultra low voltage current multiplier/divider’, Pro-
ceedings of the IEEE International Conference on Electronics, Circuits and
Systems (ICECS), 1999, pp. 1369–72
115 Berg, Y., Lande, T. S., and Naess, O.: ‘Programming floating-gate circuits with
UV-activated conductances’, IEEE Transactions on Circuits and Systems II:
Analog and Digital Signal Processing, 2001 48, pp. 12–19
116 Berg, Y., Lande, T. S., Naess, O., and Gundersen, H.: ‘Ultra-low-voltage
floating-gate transconductance amplifiers’, IEEE Transactions on Circuits and
117 Tsividis, Y.: Operation and modeling of the MOS transistor (McGraw and Hill,
1987)
118 Foty, D.: MOSFET modelling with SPICE: principles and practice (Prentice
Hall, 1997)
119 Allen, P. E., and Holberg, D. R.: CMOS analog circuit design (Oxford University
Press, 2002)
120 Star-Hspice Manual, Release 2002.4, Avant! software V2002.4 Copyright
121 Ramírez-Angulo, J., González-Altamirano, G., and Choi, S. C., ‘Mod-
eling multiple-input floating-gate transistors for analog signal processing’,
References 287

1997, pp. 2020–3
122 Yin, L., Embabi, S. H. K., and Sánchez-Sinencio, E.: ‘ A floating-gate MOSFET
D/A converter’, Proceedings of the IEEE International Symposium on Circuits
and Systems, 1997, pp. 409–12
123 Tombs, J., Ramírez-Angulo, J., Carvajal, R. G., and Torralba, A.: ‘Integration
of multiple-input floating-gate transistors into a top-down CAD design flow’,
Conference on Design of Circuits and Integrated Systems (DCIS), Palma de
Mallorca, 1999, pp. 767–71
124 Rodríguez, E., Huertas, G., Avedillo, M. J., Quintana, J. M., and Rueda, A.:
‘Practical digital circuit implementations using νMOS threshold gates’, IEEE
Transactions on Circuits and Systems II, 2001 48, pp. 102–6
125 Cadence Software
126 Kahng, D., and Sze, S. M.: ‘A floating-gate and its application to memory
devices’, The Bell System Technical Journal, 1967 46(4), pp. 1288–95
127 Kotani, K., Shibata, T., Imai, M., and Ohmi, T.: ‘Clock-controlled neuron-MOS
logic gates’, IEEE Transactions on Circuits and Systems-II: Analog and Digital
128 Lee, C., Hou, T-H., and Kan, E. C.-C.: ‘Nonvolatile memory with a metal
nanocrystal/nitride heterogeneous floating-gate’, IEEE Transactions on Electron
Devices, 2005 52(12), pp. 2697–702
129 Bandyopadhyay, A., Serrano, G. J., and Hasler, P.: ‘Programming analog compu-
tational memory elements to 0.2% accuracy over 3.5 decades using a predictive
method’, Proceedings of the IEEE International Symposium on Circuits and
Systems, 2005 3, pp. 2148–51
130 Rodriguez-Villegas, E., and Barnes, H.: ‘Solution to the trapped charge in
FGMOS transistors’, IEE Electronic Letters, 2003 39(19), pp. 1416–17
131 Urquidi, C., Ramirez-Angulo, J., Gonzalez-Carvajal, R., and Torralba, A.:
‘A new family of low-voltage circuits based on quasi-floating gate transis-
tors’, Proceedings of the IEEE Midwest Symposium on Circuits and Systems.
MWSCAS-2002, 2002 1, pp. 93–6
132 Naess, O., Olsen, E. A., Berg, Y., and Lande T.S.: ‘ A low voltage second
order biquad using pseudo floating gate transistors’, Proceedings of the IEEE
International Symposium on Circuits and Systems, 2003, pp. 125–8
133 Ramirez-Angulo, J., Lopez-Martin, A. J., Gonzales Carvajal, R., and
Munoz Chavero, F.: ‘Very low-voltage analog signal processing based on quasi-
floating gate transistors’, IEEE Journal of Solid-State Circuits, 2004 39(3),
pp. 434–42
134 Jespers, P. G. A.: Integrated converters - D to A and A to D Architectures,
Analysis and Simulation (Oxford University Press Inc., New York, 2001)
135 Toumazou, C., Lidgey, F. J., and Haigh, D. G. (eds.): Analogue IC design:
The current mode approach (IEE Circuits and Systems Series 2, 1989)
ISBN: 0 86341 2157
136 Goncalves, R.T., Fiho, S.N., Schneider, M.C., and Galup-Montoro, C.: ‘Pro-
grammable switched current filters using MOSFET-only current dividers’,
288 References
Proceedings of the IEEE Midwest Symposium on Circuits and Systems, 1995

2, pp. 1046–9
137 Cheung, Y.L., and Buchwald, A.: ‘A sampled-data switched-current analog
16-tap FIR filter with digitally programmable coefficients in 0.8 μm CMOS’,
Proceedings of the IEEE International Solid-State Circuits Conference, 1997
429 pp. 54–5
138 Rueda, A., Yufera, A., and Huertas, J.: ‘Switched-current wave analogue filters’,
Chapter 11 in Toumazou, C., Hughes, J. B. C., and Battersby, N. C. (eds.),
Switched-currents: an analogue technique for digital technology
139 De Lima, J. A., and Cordeiro, A. S.: ‘ A low-voltage low-power analog memory
cell with built-in 4-quadrant multiplication’, IEEE Transactions on Circuits and
Systems II: Analog and Digital Signal Processing, 2003 50(4), pp. 191–5
140 de la Vega, A. S., de Queiroz, A. C. M., and Diniz, P. S. R.: ‘Adaptive filter
implementation using switched-current technique’, Proceedings of the IEEE
International Symposium on Circuits and Systems Vol. I, May 2001, pp. 17–20
141 Mazurek, A., and Wawryn, K.: ‘Programable current mode circuits’, The 8th
IEEE International Conference on Electronics, Circuits and Systems (ICECS
2001), 2001 2, pp. 553–6
142 Rodriguez-Vasquez, A., and Espejo, S.: ‘Switched-current cellular neural
networks for image processing’, Chapter 17 in Switched-currents an ana-
logue technique for digital technology, Toumazou, C., Hughes, J. B. C., and
Battersby, N. C. (eds.)
143 Steyaert, M., Crols, J., and Gogaert, S.: ‘Low-voltage analog CMOS filter
design’ in Low-voltage/low-power integrated circuits and systems, Sánchez-
Sinencio, E., and Andreou, A. G. (eds.) (IEEE Press Series on Microelectronic
Systems, New York, 1999)
144 Sánchez-Sinencio, E.: ‘Continuous-time low-voltage current-mode filters’ in
Low-voltage/low-power integrated circuits and systems, Sánchez-Sinencio, E.,
and Andreou, A. G. (eds.) (IEEE Press Series on Microelectronic Systems, New
York, 1999)
145 Schaumann, R., Ghausi, M. S., and Laker, K. R.: Design of analog filters:
passive, active RC and switched capacitor (Englewood Cliffs, NJ, Prentice-Hall,
1990)
146 Castello, R., Montecchi, F., Rezzi, F., and Baschirotto, A.: ‘Low-voltage analog
filters’, IEEE Transactions on Circuits and Systems-I, 1995 42(11), pp. 827–40
147 Tsividis, Y. P., and Voorman, J. O. (eds.): Integrated continuous-time filters:
principles, design and applications (IEEE Press, New York, 1993)
148 Banu, M., and Tsividis, Y.: ‘Fully integrated active RC filters in MOS
technology’, IEEE Journal Solid-State Circuits, 1983 18, pp. 644–51
149 Tsividis, Y., Banu, M., and Khoury, J.: ‘Continuous-time MOSFET-C filters in
VLSI’, IEEE Journal of Solid-State Circuits, 1986 21(1), pp. 15–30
150 Park, C. S., and Schaumann, R.: ‘Design of a 4-MHz analog integrated CMOS
transconductance-C bandpass filter’, IEEE Journal of Solid-State Circuits, 1988
23, pp. 987–96
References 289
151 Snelgrove, W. M., and Shoval, A.: ‘A balanced 0.9 μm CMOS

transconductance-C filter tunable over the VHF range’, IEEE Journal of
Solid-State Circuits, 1992 27(3), pp. 314–23
152 Nauta, B.: ‘A CMOS transconductance-C filter technique for very high
frequencies’, IEEE Journal of Solid-State Circuits, 1992 27(2), pp. 142–53
153 Alini, R., Baschirotto, A., and Castello, R.: ‘Tunable BiCMOS continuous-time
filter for high-frequency applications’, IEEE Journal of Solid-State Circuits,
1992 27, pp. 1905–15
154 Stefanelli, B., and Kaiser, A.: ‘A 2 μm CMOS fifth-order low-pass continuous-
time filter for video-frequency applications’, IEEE Journal of Solid-State
Circuits, 1993 28(7), pp. 713–18,
155 Rezzi, F., Bietti, I., Cazzaniga, M., and Castello, R.: ‘A 70 mW seventh-
order filter with 7-50 MHz cutoff frequency and programmable boost and
group delay equalization’, IEEE Journal of Solid-State Circuits, 1997 32(12),
pp. 1987–99
156 Ganti, R., Carley, L. R., and Myers, B. A.: ‘A low distortion high frequency
transconductor structure’. Proceedings of the IEEE International Symposium on
Circuits and Systems, II, 1999 pp. 216–19
157 Pavan, S., Tsividis, Y. P., and Nagaraj, K.: ‘Widely programmable high-
frequency continuous-time filters in digital CMOS technology’, IEEE Journal
of Solid-State Circuits, 2000 35(4), pp. 503–11
158 Giustolisi, G., De, G., and Pennisi, S.: ‘High-linear class AB transconductor for
high-frequency applications’, Proceedings of the IEEE International Symposium
on Circuits and Systems, 2000, pp. 169–72
159 Voorman, H., and Veenstra, H.: ‘Tunable high-frequency Gm-C filters’, IEEE
Journal of Solid-State Circuits, 2000 35(8), pp. 1097–108
160 Silva-Martínez, J., Steyaert, M., and Sansen, W.: High-performance CMOS
continuous-time filters (Kluwer Academic Publishers, 1993)
161 Nedungadi, A., and Viswanathan, T. R.: ‘Design of linear CMOS transcon-
ductance elements’, IEEE Transactions on Circuits and Systems, 1984 31,
pp. 891–4
162 Pennock, J. L.: ‘CMOS triode transconductor for continuous-time active
integrated filters’, IEE Electronic Letters, 1985 21, pp. 817–18
163 Seevinck, E., and Wasenaar, R. F.: ‘ A versatile CMOS linear transconductor
square-law function circuit’, IEEE Journal of Solid-State Circuits, 1987 22,
pp. 366–77
164 Szczepanski, S., Schaumann, R., and Wu, P.: ‘Linear transconductors based on
cross-coupled CMOS pair’, IEE Electronic Letters, 1991 27, pp. 783–5
165 Tanimoto, H., Koyama, M., and Yoshida, Y.: ‘Realization of a 1V active filter
using linearization technique employing plurality of emitter-couple pairs’, IEEE
Journal of Solid State Circuits, 1991 26, pp. 937–45
166 Silva-Martinez, J., Steyaert, M., and Sansen, W.: ‘ A large-signal very low-
distortion transconductor for high-frequency continuous-time filters’, IEEE
Journal of Solid-State Circuits, 1991 SC-26, pp. 946–55
290 References
167 Koyama, M., Arai, T., Tanimoto, H., and Yoshida, Y.: ‘A 2.5 V active low-pass
filter using all n-p-n Gilbert cells with a 1-Vpp linear input range’, IEEE Journal
of Solid State Circuits, 1993 28, pp. 1246–53
168 Yang, F., and Enz, C. C.: ‘A low distortion BiCMOS seventh-order Bessel
filter operating at 2.5 V supply’, IEEE Journal Solid-State Circuits, 1996 31,
pp. 321–30
169 Punzenberger, M., and Enz, C.: ‘A new 1.2 V BiCMOS log-domain integrator
for companding current mode filters’. Proceedings of the IEEE International
Symposium on Circuits Systems, Vol. 1, Atlanta, GA, 1996, pp. 125–8
170 Punzenberger, M., and Enz, C. C.: ‘A compact low-power BiCMOS log-domain
filter’, IEEE Journal of Solid-State Circuits, 1998 33(7), pp. 1123–9
171 Szczepanski, S., Jakusz, J., and Schaumann, R.: ‘A linear fully balanced
CMOS OTA for VHF filtering applications’, IEEE Transactions on Circuits
and Systems-II: Analog and Digital Signal Processing, 1997 44(3), pp. 174–87
172 Rodríguez, E. O., Yúfera, A., and Rueda, A.: ‘A low-voltage floating-gate MOS
transconductor’, Proceedings of XIV Design of Integrated Circuits and Systems
Conference, Palma de Mallorca, Spain, 1999, pp. 539–44
173 Lee, T-S., and Liu, C-C.: ‘Design techniques for low-voltage VHF BiCMOS
transconductance-C filter with automatic tuning’, Proceedings of International
Symposium on Circuits and Systems, II, 2000, pp. 601–4
174 Rodríguez-Villegas, E. O., Rueda, A., and Yúfera, A.: ‘A 1.5 V second-order
FGMOS filter’, Proceedings of XV Design of Integrated Circuits and Systems
Conf. (DCIS’00), 2000, pp. 358–62
175 Rodríguez, E. O., Yúfera, A., and Rueda, A.: ‘A Gm -C floating-gate MOS
integrator’, Proceedings of the IEEE International Symposium on Circuits and
Systems, 2000, pp. 153–6
176 Rodríguez, E. O., Yúfera, A., and Rueda, A.: ‘A low-voltage sqr-root floating-
gate Mos integrator’, Proceedings of the IEEE International Symposium on
Circuits and Systems, 2000, pp. 184–7
177 Rodríguez-Villegas, E. O., Rueda, A., and Yúfera A.: ‘A 1.5 V second-
order FGMOS filter’, Proceedings of the IEEE European Solid-State Circuits
178 Rodríguez-Villegas, E., Yúfera, A., and Rueda, A.: ‘A low-voltage floating-gate
MOS biquad’, VLSI Design, 2001 12(3), pp. 407–14
179 Adams, R. W.: ‘Filtering in the log domain’, Proceedings of 63rd Convention
A. E. S., May 1979, pp. 1470–6
180 Seevinck, E.: ‘Companding current-mode integrator: a new circuit principle
for continuous-time monolithic filters’, IEE Electronic Letters, 1990 26(24),
pp. 2046–7
181 Tsividis, Y., Gopinathan, V., and Tóth, L.: ‘Companding in signal processing’,
IEE Electronic Letters, 1990 26, pp. 1331–2,
182 Frey, D. R.: ‘Log-domain filtering: an approach to current-mode filtering’, IEE
Proceedings-G, 1993 140(6), pp. 406–16
183 Frey, D. R.: ‘A general class of current mode filters’, Proceedings of the IEEE
International Symposium on Circuits and Systems, 1993 2, pp. 1435–8
References 291
184 Frey, D. R.: ‘Exponential state filters: a generic current mode design strategy’,
IEEE Transactions on Circuits Systems I, 1996 43, pp. 34–42
185 Mulder, J., van der Woerd,
√ A. C., Serdijn, W. A., and Roermund, A. M.: ‘Current-
mode companding x domain integrator’, IEE Electronic Letters, 1996 32(3),
pp. 198–9
186 Frey, D.: ‘Future implications of the log domain paradigm’, IEE Proceedings
on Circuits Devices System, 2000, 147(1) pp. 65–72
187 Eskiyerli, M., and Payne, A. J.: ‘Square root domain filter design and
performance’, Analog Integrated Circuits and Signal Processing 2000 22,
pp. 231–43
188 López-Martín, A. J., and Carlosena, A.: ‘A 1.5 V CMOS square-root domain fil-
ter’, Proceedings of the IEEE International Symposium on Circuits and Systems,
2001, pp. 1465–8
189 Frey, D. R.: ‘Log domain filtering for RF applications’, IEEE Journal of Solid
State Circuits, 1996 31(10), pp. 1468–75
190 El-Gamal, M., Leung, V., and Roberts, G. W.: ‘Balanced log-domain filters
for VHF applications’, Proceedings of the IEEE International Symposium on
Circuits and Systems, 1997, pp. 493–6
191 El-Gamal M. N., and Roberts, G. W.: ‘Very high-frequency log-domain
bandpass filters’, IEEE Transactions on Circuits and System II, 1998 45,
pp. 1188–98
192 Payne, A., Khumsat, P., and Thanachayanont, A.: ‘Design and implementation
of high frequency log-domain filters’, Proceedings of the IEEE International
Symposium on Circuits and Systems, 1999, pp. 972–5
193 El-Gamal, M. N., Baki, R. A., and Bar-Dor, A.: ‘30-100 MHz NPN-only
variable-gain class AB instantaneous companding filters for 1.2 V applications’,
IEEE Journal of Solid-State Circuits, 2000 35, pp. 1853–64
194 Serra-Graells, F.: ‘VLSI CMOS low-voltage log companding filters’, Proceed-
ings of the IEEE International Symposium on Circuits and Systems, Vol. I, 2000,
pp. 172–5
195 Masmoudi, D., Serdijn, W. A., Mulder, J., van der Woerd, A. C., Tomas, J.,
and Dom, J. P.: ‘A new current-mode synthesis method for dynamic translinear
filters and its applications in hearing instruments’, Journal of Analog IC and
196 Fried, R., Python, D., and Enz, C. C.: ‘Compact log-domain current mode inte-
grator with high transconductance to bias current ratio’, IEE Electronic Letters,
1996 32(11), pp. 952–3
197 Python, D., Punzenberger, M., and Enz, C.: ‘A 1 V CMOS log-domain integra-
tor’, Proceedings of the IEEE International Symposium on Circuits and Systems,
Vol. II, 1999, pp. 685–8
198 Python, D., and Enz, C.: ‘A micropower class AB CMOS log-domain filter for
DECT applications’, Proceedings of the IEEE European Solid-State Circuits
199 El-Masry, E. I., and Wu, J.: ‘Low voltage micropower log-domain filters’,
Journal of Analog IC and Signal Processing, 2000 22, pp. 209–20
292 References
200 Serra-Graells, F.: ‘All-MOS subthreshold log filters’, Proceedings of the

International Symposium on Circuits and Systems, Vol. II, 2001, pp. 137–40
201 Perry, D., and Roberts, G. W.: ‘The design of log-domain filters based on the
operational simulation of LC ladders’, IEEE Transaction on Circuit and System
II: Analog and Digital Signal Processing, 1996 43, pp. 763–74
202 Drakakis, E. M., Payne, A. J., and Toumazou, C.: ‘Bernoulli operator: a low
level approach to log-domain signal processing’, IEE Electronic Letters, 1997
33(12), pp. 1008–9
203 Fox, R., Nagarajan, M., and Harris, J.: ‘Practical design of single-ended log-
domain filter circuits’, Proceedings of the IEEE International Symposium on
Circuits and Systems, 1997 1, pp. 341–4
204 Himmelbauer, W., and Andreou, A. G.: ‘Log-domain circuits in subthresh-
old MOS’, Proceedings of the IEEE Midwestern Symposium on Circuits and
Systems, 1997 1, pp. 26–30
205 López, A., and Carlosena, A.: ‘The exponential CCII-, a building block for
log-domain circuits’, Proceedings of the IEEE International Conference on
Electronics, Circuits, and Systems, 1998 2, pp. 389–92
206 Baswa, S., Bikumandla, M., Ramirez-Angulo, J., Lopez-Martin, A.J.,
Carvajal, R. G., and Ducoudray-Acevedo, G.: ‘Low-voltage low-power super
class-AB CMOS op-amp with rail-to-rail input/output swing’, Proceedings of the
Fifth IEEE International Caracas Conference on Devices, Circuits and Systems,
2004 1, pp. 83–6
207 Rodriguez-Villegas, E., Yufera, A., and Rueda, A.: ‘A 1-V micropower log-
domain integrator based on FGMOS transistors operating in weak inversion’,
IEEE Journal of Solid-State Circuits, 2004 39(1), pp. 256–9
208 Rodriguez-Villegas, E., Yufera, A., and Rueda, A.: ‘A 1.25-V micropower
Gm-C filter based on FGMOS transistors operating in weak inversion’, IEEE
Journal of Solid-State Circuits, 2004 39(1), pp. 100–11
209 Chawla, R., Haw-Jing Lo, Basu, A., Hasler, P., and Minch, B. A.: ‘A fully
programmable log-domain bandpass filter using multiple-input translinear ele-
ments’, Proceedings of the IEEE International Symposium on Circuits and
Systems, 2004 1, pp. I-33 - I-36
210 Rodriguez-Villegas, E.: ‘A 0.8 V, 360 nW Gm-C biquad based on FGMOS
transistors’, Proceedings of the IEEE International Symposium on Circuits and
Systems ISCAS 2005, May 2005, pp. 2156–9
211 Steyaert, M., and Sansen, W.: ‘Power-supply rejection ratio in operational
transconductance amplifiers’, IEEE Transaction on Circuits and Systems, 1990
37, pp. 1077–84
212 Johns, D. A., and Martin, K.: Analog integrated circuit design (Wiley, New York,
1997)
213 Vittoz, E.: ‘Micropower techniques’ in Design of VLSI circuits for telecommu-
nication and signal processing, Franca, J., and Tsividis, Y. (eds.) (Prentice-Hall,
Englewood Cliffs, NJ, 1993)
214 Austria Mikro Systeme International, ‘0.8 μm CMOS process parameters’,
9933006, rev_a, Nov. 1996
References 293
215 Rodríguez-Villegas, E. O., Rueda, A., and Yúfera, A., ‘A 10.7 MHz FGMOS
low power sixth order bandpass filter’, Proceedings of Design Circuits and
Integrated Systems Conference (DCIS), 2001, pp. 68–71
216 Silva-Martinez, J., Steyaert, M. S. J., and Sansen, W.: ‘Design techniques for
high-performance full-CMOs OTA-RC continuous-time filter’, IEEE Journal
of Solid-State Circuits, 1992 26, pp. 993–1001
217 Austria Mikro Systeme International, ‘0.35 μm CMOS process parameters’,
9933016, rev_2.0, Nov. 1998
218 Gilbert, B.: ‘Translinear circuits: a proposed classification’, Electronic Letters,
11(1), pp. 14–16. See also errata, Vol.11, no.6, pp. 136, 1975
219 Serrano-Gotarredona, T., Linares-Barranco, B., and Andreou, A. G.: ‘General
translinear principle for subthreshold MOS transistors’, IEEE Transactions on
Circuits and Systems I, 1999 46(5), pp. 607–16
220 Shepherd, L., and Townazou, C.: ‘Towards direct biochemical analysis with
weak inversion ISFET’, IEEE International Workshop on Biomedical Circuits
and Systems, 2004, pp. S1/5–S5-8
221 Yin, Q., Eisenstadt, W. R., and Fox, R. M.: ‘A translinear-based RF RMS
detector for embedded test’, Proceedings of the 2004 International Symposium
on Circuits and Systems, 2004. ISCAS ’04, Vol. 1, pp. 245–8
222 Abuelma’atti, M. T.: ‘A translinear circuit for sinusoidal frequency divi-
sion’, IEEE Transactions on Instrumentation and Measurement, 2005 54(1),
pp. 10–14
223 Psychalinos, C., Fragoulis, N., and Haritantis, I.: ‘Log-domain wave filters’,
IEEE Transactions on Circuits and Systems II: Analog and Digital Signal
Processing, 2004 51(6), pp. 299–306
224 Baki, R. A., and El-Gamal, M. N.: ‘A low-power 5-70-MHz seventh-order log-
domain filter with programmable boost, group delay, and gain for hard disk drive
applications’, IEEE Journal of Solid-State Circuits, 2003 38(2), pp. 205–15
225 Gerosa, A., Maniero, A., and Neviani, A.: ‘A fully integrated dual-channel
log-domain programmable preamplifier and filter for an implantable cardiac
pacemaker’, IEEE Transactions on Circuits and Systems I: Fundamental Theory
and Applications, 2004 51(10), pp. 1916–25
226 Serra-Graells, F.: ‘VLSI CMOS subthreshold log companding analog cir-
cuit techniques for low-voltage applications’, Ph.D. dissertation, Universidad
Autónoma Barcelona, 2001
227 Cilingiroglu, U.: ‘A purely capacitive synaptic matrix for fixed-weight neural
networks’, IEEE Transactions on Circuits and Systems, 1991 38, pp. 210–17
228 Lerch, J. B.: ‘Threshold gate circuits employing field-effect transistors’,
US03715603, Feb. 6, 1973
229 Johnson, M. G.: ‘A symmetric CMOS NOR gate for high speed applications’,
IEEE Journal of Solid State Circuits, 1998 23, pp. 1233–6
230 Schultz, K. J., Francis, R. J., and Smith, K. C.: ‘Ganged CMOS: trading standby
power for speed’, IEEE Journal of Solid-State Circuits, 1990 25(3), pp. 870–3
231 Goodwin-Johansson, S.: ‘Circuit to perform variable threshold logic’, 1990,
US048906059
294 References
232 Varshavsky, V., ‘Beta-driven threshold elements’, Proceedings of Great Lakes

Symposium on VLSI, 1998, pp. 52–8
233 Prange, S., Thewes, R., Wohlrab, E., and Weber, W.: Circuit arrangement for
realizing logic elements that can be represented by threshold value equations’,
1999, US 05991789
234 Lee, C. L., and Jen, C-W.: ‘CMOS threshold gate and networks for order statistic
filtering’, IEEE Transactions on Circuits and Systems I: Fundamental Theory
and Applications, 1994 41(6), pp. 453–6
235 Quintana, J. M., Avedillo, M. J., Rueda, A., and Baena, C.: ‘Practical low-cost
CMOS realization of complex logic functions’, Proceedings of the European
Conference on Circuit Theory and Design (ECCTD), 1995, pp. 51–4
236 Goser, K. F., Pacha, C., Kanstein, A., and Rossmann, M. L.: ‘Aspects of sys-
tems and circuits for nanoelectronics’, Proceedings of the IEEE, 1997 85(4),
pp. 558–73
237 Avedillo, M. J., Quintana, J. M., and Rueda, A.: ‘Threshold logic’, Wiley ency-
clopedia of electrical and electronics engineering, Webster, J. G., (ed.), 1999
22, pp. 178–90
238 Rodríguez-Villegas, E., Quintana, J. M., Avedillo, M. J., and Rueda, A.:
‘Threshold logic based adders using floating-gate circuits’ in Advances in
physics, electronics and signal processing applications (World Scientific and
Engineering Society Press, 2000)
239 Quintana, J. M., Avedillo, M. J., Jiménez, R., Rodríguez Villegas, E., and
Rueda, A: ‘Improved compressor designs bases on threshold logic’, Proceedings
of the Seminario Anual de Automática, Electrónica Industrial e Instrumentación
(SAEEI), 2001, pp. 1–5 (index)
240 Quintana, J. M., Avedillo, M. J., Rodríguez-Villegas, E., and Rueda, A.: ‘vMOS-
based compressor designs’, Proceedings of the 12th International Conference
on Microelectronics (ICM), 2000, pp. 33–6
241 Rodríguez-Villegas, E., Quintana, J. M., Avedillo, M. J., and Rueda, A.: ‘Sorting
networks implemented as νMOS circuits’, IEE Electronic Letters, 1998 34(23),
pp. 2237–8
242 Rodríguez-Villegas, E., Avedillo, M. J., Quintana, J. M., Huertas, G., and
Rueda, A.: ‘νMOS based sorters for arithmetic applications’, VLSI Design
Journal, 2000 11(2), Number j99309
243 Quintana, J. M., Avedillo, M. J., Rodríguez-Villegas, E., Rueda, A., and Huer-
tas, J. L.: ‘Efficient νMOS realization of threshold voters for self-purging
redundancy’, Proceedings of the XIII Symposium on Integrated Circuits and
Systems Design, 2000, pp. 321–6
244 Rodríguez-Villegas, E., Avedillo, M. J., Quintana, J. M., Huertas, G., and
Rueda, A.: ‘νMOS-based sorters for multiplier implementations’, Proceedings
of the IEEE International Symposium on Circuits and Systems (ISCAS’99),
Orlando, 1999
245 Zhuang, N., and Wu, H.: ‘A new design of the CMOS full adder’, IEEE Journal
of Solid-State Circuits, 1992 27(5), pp. 840–4
References 295
246 Abu-Khater, I. S., Bellaouar, A., and Elmasry, M. I.: ‘Circuit techniques for
CMOS low-power high-performance multipliers’, IEEE Journal of Solid-State
Circuits, 1996 31(10), pp. 1535–46
247 Gustafsson, O., Dempster, A.G., and Wanhammar, L.: ‘Multiplier blocks
using carry-save adders’, Proceedings of the IEEE International Symposium
on Circuits and Systems, 2004 2, pp. 473–6
248 Quintana, J. M., Avedillo, M. J., Jimenez, R., and Rodriguez-Villegas, E.: ‘Low-
power logic styles for full-adder circuits’. Proceedings of the IEEE International
Conference on Electronics, Circuits and Systems (ICECS), 2001 3, pp. 1417–20
249 Hwang, K.: Computer Arithmetic, Principles, Architecture and Design (John
Wiley & Sons, 1979) ISBN:0471052000
250 Song, P. J., and De Micheli, G.: ‘Circuit and architecture trade-offs for high-speed
multiplication’, IEEE Journal of Solid-State Circuits, 1991 26(9), pp. 1184–98
251 Zimmermann, R., and Fichtner, W.: Low-power logic styles: CMOS versus pass-
transistor logic’, IEEE Journal of Solid-State Circuits, 1997 32(7), pp. 1079–90
252 Lamagna, E. A.: ‘The complexity of monotone networks for certain bilin-
ear forms, routing problems, sorting, and merging’, IEEE Transaction on
Computing, 1979 C-28, pp. 773–82
253 Batcher, K. E., and Ohmi, T.: ‘Sorting networks and their applications’.
Proceedings of the AFIPS Spring Joint Computer Conference, 1968 32,
pp. 307–14
254 Leblebici, Y., Özdemir, H., Kepkep, A., and Çilingiroglu, U.: ‘A compact (8×8)-
bit serial/parallel multiplier based on capacitive threshold logic’, Proceedings
of the European Conference on Circuit Theory and Design (ECCTD), 1995
pp. 55–8
255 Muroga, S.: Threshold logic and its applications (John Wiley & Sons, New York,
1971)
256 Wuu, T.-Y., and Vrudhula, S. B. K.: ‘A design of a fast and area efficient
multi-input Muller C-element’, IEEE Transactions on VLSI Systems, 1993 1(2),
pp. 215–19
Index
Page numbers in italics show pages with tables or figures that are separated
from text.
νMOS: see neuron MOS (νMOS) hot electron injection 29–30

poly 1 usage 30–3
adaptive biasing 4 pseudo- (quasi-) FGMOS transistors
adders 33–5
FGMOS as MOS transistor plus adder UV cleaning 28–9
equivalent 43–4 charge pumps 3
see also neuron MOS (νMOS) threshold circuit equivalents: see FGMOS circuit
gate (TG) adders equivalents
analog/digital, choice between 1 CMFB (common mode feedback)
applications: see FGMOS applications with log-domain filtering 223–4, 226,
area aspects of FGMOS 35–9 227
input capacitor areas 36 and PSSR 113–15
matching issues 37 in the strong inversion ohmic region
MOS device area 36 95–6, 102, 109–11, 113–15
see also input capacitance of FGMOS in the strong inversion saturation region
ASICs (application-specific integrated 145–6, 147–8, 150, 159–62
circuits) 1 in the weak inversion region 193, 195
CMOS (complementary metal oxide
Batcher’s method, sorters 259, 262 semiconductor), benefits of 2
BCLA (block carry lookahead) unit 252–3 CMRR (common mode rejection ratio)
BJTs (bipolar junction transistors) 215, 216 102–3, 139–43, 148, 161–2
companding, instantaneous 4
capacitance of input: see input capacitance comparators, offset compensated 60–6
of FGMOS about FGMOS comparators 60–1
capacitors, metal-poly: see metal-poly comparison phase 62–3
capacitors (single poly technology) feedback issues 62–4
cascading stages, for voltage reduction 3 feedthrough 64
cascode current mirrors: see current mirrors input capacitances mismatch 64
charge accumulation problems/techniques input transistor threshold voltage
28–35 mismatch 61–4
about charge problems 28 offset compensation phase 62–3
electrical switch solution 30 parasitic capacitance 64–5
Fowler-Nordheim tunnelling process 29 PMOS transistor offsets 65–6
298 Index
compressing/expanding/companding 4, D/A (digital to analog) converter

223 design; input capacitance;
see also log-domain filtering; neuron log-domain filtering; neuron MOS
MOS (νMOS) threshold gate (TG) (νMOS) threshold gate (TG);
based compressor designs OPAMP (operational amplifier),
construction/fabrication of FGMOS devices two-stage; simulation/simulators
15–17 digital/analog, choice between 1
continuous-time analog filters: see filtering, DR (dynamic range) 21, 103, 144, 162
analog continuous-time
CONTOUR function in Matlab program
227 EEPROMs 7, 8, 9, 28
converters: see D/A (digital to analog) electric fields, and low voltages 2
converter design ELIN (externally linear-internally nonlinear)
current mirrors integration/topologies 88, 215
input/output characteristics 48–9 EPROMs 7
minimum voltage requirements 49–50 ESS (exponential state space) circuits 215
MOS and FGMOS comparisons 46–50 ETANN chip (Intel) 7
and PSIFG cells 74
current multipliers, FGMOS in weak fabrication/construction of FGMOS devices
inversion as a 45–6 15–17
current reduction techniques feedback: see CMFB (common mode
adaptive biasing 4 feedback)
adiabatic circuits 6 FGMOS
operand isolation 6
about floating gate MOS 6–7, 40
parallelism 6
for circuit complexity reduction 7
precomputation 6
construction/fabrication 15–17
problem redefinition 6
equivalent schematic 17
subthreshold biasing 4
for facilitating tuning 7
technology scaling 6
and flexibility, controllability and
transistor sizing 5–6
tunability 7
transition reduction 6
history of 7–11
voltage scaling 5
notation 40–2
for signal level shifting 7
D/A (digital to analog) converter design for simplifying signal processing 7
66–74 symbol and voltage definitions 17
basic converter 66–7 FGMOS applications 46–85
biasing capacitors 67–70 about applications 43, 79–80
boundary conditions 69 cascode current mirrors 46–50
with cascade transistors 70–2 D/A (digital-to-analog) converters
design 2 converter 72–3 66–73
linearity 70 notation 81–5
low voltage/low speed applications 72 offset compensated FGMOS comparator
Monte Carlo simulation results 70 60–6
operational examples 70, 71 programmable FGMOS inverter 58–60
output current 74 PSIFG (programmable switched current
output voltage 70, 71 floating gate) cell 73–9
performance summary 71 two-stage OPAMP (operational
transistors, keeping saturated 68 amplifier) 50–8
DDA (differential difference amplifier) 10 FGMOS circuit equivalents 43–6
design with FGMOS about circuit equivalents 43, 79–80
about design with FGMOS 22 as a MOS transistor with controllable
see also area aspects of FGMOS; charge threshold voltage 44–5
accumulation problems/techniques; as a MOS transistor plus adder 43–4
Index 299
in weak inversion as a current multiplier operating point 92–5

45–6 threshold voltage shift 93–4
FGMOS design HD2 (second harmonic distortion) 108,
about FGMOS design 22 109
see also area aspects of FGMOS; charge intermodulation test (IM3) 104
accumulation problems/techniques; noise 108–11
D/A (digital to analog) converter small signal equivalent circuit
design; input capacitance; 109, 110
log-domain filtering; neuron MOS notation 123–7
(νMOS) threshold gate (TG); PSRR (power supply rejection ratio)
OPAMP (operational amplifier), 111–15
two-stage; simulation/simulators small signal equivalent circuits
FGMOS theory 111–14
dynamic range (DR) 21 second order effects 103–11
gain 20–1 small signal considerations 96–103
input transconductance 19 THD (total harmonic distortion)
large signal DC behaviour 18–19 103–11, 116–17, 119, 120
modelling 19 transconductance cell (N1) 97–9
noise considerations 21 tuning amplifier (N2) 99–101
output resistance 19–21 filtering, analog continuous-time, based on
small signal equations 19–21 FGMOS in the strong inversion
strong inversion ohmic 20 saturation region 129–84
strong inversion saturation 20 about FGMOS in the strong inversion
weak inversion saturation 20 saturation region 129, 175–7
filtering, analog continuous-time comparative performances 175–7
about continuous-time filters 87–9 Gm -C filter for audio example 152–7
MOSFET-C filters 88 design performance 154–7
OTA-C filters 88 HSPICE simulations summary
RC active filters 88 156
transconductor basic blocks 89–90 programming 156, 157
filtering, analog continuous-time, based on THD (total harmonic distortion)
FGMOS in the strong inversion 155–6
ohmic region 87–127 Gm -C integrator for audio 130–57
about FGMOS in the strong inversion advantages 131
ohmic region 87–9, 117–23 circuit operation 130–1
CMFB (common mode feedback) CMFB (common mode feedback)
95–6, 102, 109–11, 113–15 131, 135–43, 145–6, 147–8
CMRR (common mode rejection ratio) CMMR (common mode rejection
102–3 ratio) 139–43, 148
common mode DC response 101–3 design trade-offs 132
comparative performances 175–7 DR (dynamic range) 144
drain voltage 104 frequency response/bandwidth
feedback circuits 98 137–9
filter example 115–17, 118–21 gain 139–43
lowpass (LP) and bandpass (BP) noise 145–8, 148–52
transfer functions and responses power consumption features
116, 118–19 132–5
THD (total harmonic distortion) PSRR (power supply rejection
116–17, 119, 120 ratio) 148–52
Gm -C multiple-input linear integrator second order effects 143–8
91–5 small signal equivalent circuit 136
cascode output stage, output swing THD (total harmonic distortion)
94–5 143–4, 145, 146
300 Index
filtering, analog continuous-time, based on gain of FGMOS 20–1

FGMOS in the strong inversion gain-programming, Gm -C FGMOS
saturation region (continued) integrator 207, 209
transconductance independence GBW (gain bandwidth) 50, 52–3
131 GIDL (gate-induced-drain-leakage) 8
transconductance limits 134 Gm -C FGMOS integrator
intermediate frequency filter 157–68 for audio, with strong inversion
cascode load 159–60 saturation region 130–57
CMFB (common mode feedback) multiple-input linear with strong
159–61 inversion ohmic region 91–5
CMRR (common mode rejection see also TP (translinear principle)
ratio) 161–2 Gm -C FGMOS integrator using translinear
cut-off frequency 165 circuits in the weak inversion region
DR (dynamic range) 162 188–207
FGMOS-based transconductor β parameter dependence on gate voltage
158, 159 195–6
performance summary 168 about using translinear circuits 188,
PSRR (power supply rejection 207–10
ratio), neg supplies 163–4 advantages 207–8
PSRR (power supply rejection CMFB (common mode feedback block)
ratio), pos supplies 163–4 193, 195
disadvantages 208
quality factor 165, 167
FGMOS blocks 189–93
second order filter prototype
filter (second-order) example 202–7
165–8
about the filter 202–4
notation 177–84
experimental results 205–7
second order filter in single poly
gain-programming 207, 209
technology 168–75
noise floor 207, 210
cut-off frequency 173, 175 performance summary 210
effects of the integrating Q-programming 207, 208
capacitance 172–3 THD (total harmonic distortion)
effects of metal-poly capacitors 205–7
168–73 tuning 204
experimental results/performance gate-to-drain parasitic coupling problems
173–5, 175–6 194–5
frequency response of the OTA input capacitances mismatch 193–4
171–2 notation 211–14
gain tuning 173, 176 offsets from mismatches 196, 197
offset effects 169–71 output resistance degradation from drain
THD (total harmonic distortion) gate capacitive coupling 195
173, 175 PSSR (power supply rejection ratio)
filtering, continuous time Gm -C FGMOS: 201–2
see Gm -C FGMOS integrator using second-order effect parameters for
translinear circuits in the weak calculations 197
inversion region second-order harmonics and mismatches
filtering, log domain: see log-domain 198
filtering state space circuits 193
filters, second order with PSIFG cells state-space integrator equations
77–9, 80 188–9
FLASH memories 7 THD (total harmonic distortion) effects
Fowler-Nordheim tunnelling process and 198–201
charge accumulation 29 third-order harmonics and
Fox’s method for stability analysis 227 intermodulation distortion 198
Index 301
transconductors 189–93 CMFB (common mode feedback)

about nonlinear transconductors 223–4, 226, 227
189–90 common mode control 225–6
differential transconductor 193, compressors 223, 226
194 CONTOUR function in Matlab program
transistor
√ mismatch problems 195 227
y/ x circuit 191–2 design example 235–7
cut-off frequency programming
235–7
hot electron injection and charge gain programming 238, 239
accumulation 29–30 harmonic distortion 240
HSPICE simulations 24–5, 156, 257, 265, performance summary 236
267–8 quality factor programming 237,
238
ILEL (internally linear-externally linear) signal to noise ratio 239
integration 88 expanders 223, 224, 225
initial transient analysis (ITA) simulation Fox’s method for stability analysis 227
technique 23–8 gate to drain capacitance effects 229–35
input capacitance of FGMOS 35–9 HD2 (second harmonic distortion)
area issues 35–7 233–5
intrinsic capacitance 38 HD3 (third harmonic distortion) 233–5
matching issues 37 input stage 225, 226
minimum capacitance 37 multiple operating points 226–7
overlap capacitance 38 NLBs (non-linear blocks) 223
parasitic capacitances 37 notation 241–4
simplified mathematical model 38–9 output resistance and parasitic
strong inversion ohmic 39 capacitance 228–9
strong inversion saturation 38 parasitic capacitances, effects of 228–9
weak inversion saturation 38 second-order filter design 222–5
see also area aspects of FGMOS stability analysis 227
input transconductance of FGMOS 19 THD (total harmonic distortion) 231–5
instantaneous companding 4 log-domain integration 216–22
integration advantages/disadvantages of using
ELIN (externally linear-internally FGMOS transistors 221–2
nonlinear) methods 88 basic circuits 217–19
ILEL (internally linear-externally linear) basic principle 216–17
methods 88 fully differential integrator 220, 221
see also log-domain integration input voltage from input current 218
Intel ETANN chip 7 logarithmic integrator 219–21
intermodulation test (IM3) 104 NLB (non-linear block)/divider block
inversion layer charge 185 218, 219–20, 223
inverters, programmable 58–60 low voltage, benefits of 1–2
input/output characteristics 59 LSI (large scale integration) 9
switching thresholds CMOS and FGMOS LTI (linear, time invariant) systems 215
58–9
trapped charge (Q) issues 60
inverters, simulation of 25–8 Matlab program, CONTOUR function 227
ITA (initial transient analysis) simulation MEMS (micro electro mechanical
technique 23–8 systems) 2
metal-poly capacitors (single poly
technology) 30–3, 168–75
log-domain filtering 215–44 construction and performance
about log-domain filtering 215–16 comparisons 30–3
302 Index
metal-poly capacitors (single poly about binary sorting networks 258–60

technology) (continued) arithmetic applications
experimental results for second-order (8 × 8) multiplier 263–5
biquad 173–5 (15,4) counter 265–7
frequency response with an OTA Batcher’s method 259, 262
171–2 design/evaluation of an 8-input sorter
integrating capacitor parasitics 172–3 261–3
offset effects 169–71 electrical realisation 260–1
microelectronics, progress of 1 NLB (non-linear block)/divider block 218,
micropower circuits, definition 185 219–20, 223
modelling FGMOS 19, 22, 38–9 noise 21, 108–11, 145–8, 148–52
see also FGMOS design see also PSRR (power supply rejection
MOSFET-C filters 88 ratio)
Muller C-element: see neuron MOS (νMOS) noise floor, Gm -C FGMOS integrator 207,
threshold gate (TG) multi-input 210
Muller C-element nonlinear processing of the signals, for
voltage reduction 3–4
nonlinear transconductors 189–90
netlists 25–7
neural networks 9–10
neuron MOS (νMOS), about 9 ohmic region: see filtering, analog
neuron MOS (νMOS) threshold gate (TG) continuous-time, based on FGMOS
247–71 in the strong inversion ohmic
about (νMOS) threshold gates 247 region
design 249–50 OPAMP (operational amplifier), two-stage
electrical simulation 250 50–8
notation 269–71 cascode transistors, effect of 56
signal regeneration 250 compensation capacitance 53–4
theory 247–9 FGMOS first version 51–4
neuron MOS (νMOS) threshold gate (TG) FGMOS second version 54–7
adders 250–5 gain 53–5
64-bit adder 254–5 GBW (gain bandwidth) 50, 52–3
carry generation problem 250–1 minimum voltage supply 51
carry lookahead adders 251–3 MOS version 51–2
4-bit block carry lookahead negative/positive design tradeoffs
(BCLA) unit 252–3 56–7
N -bit ripple-carry adders 250 slew rate (SR) 50
physical implementation 254 operand isolation 6
threshold gate based implementation OTA-C (operational transconductance
253–4 amplifier and capacitor)
neuron MOS (νMOS) threshold gate (TG) integrator/filter topologies 88
based compressor designs 255–8 frequency response 171–2
(4,2) design approach 256 output resistance of FGMOS 19–21
HSPICE simulations 257
PDP (power delay product) figures 258 parallelism 6
performance comparisons 256–8, 259 parasitic capacitances 37, 64–5, 97,
PPST (partial-product summation tree) 228–9
255–6 partial-product summation tree (PPST)
neuron MOS (νMOS) threshold gate (TG) 255–6
multi-input Muller C-element PDP (power delay product) figures 258
267–8 PMOS transistor offsets 65–6
HSPICE simulation 267–8 polysilicon thin film transistors (TFT) 49
neuron MOS (νMOS) threshold gate (TG) poly technology: see metal-poly capacitors
sorting networks (SNs) 258–67 (single poly technology)
Index 303
power slew rate (SR) 50

analog techniques for reduction SN (sorting network): see neuron MOS
2–4 (νMOS) threshold gate (TG) sorting
digital techniques for reduction 4–6 networks (SNs)
interconnect power dissipation 2 Spectre simulator 27
need for low power 2 square-root circuits: see Gm -C FGMOS
see also current reduction integrator using translinear
techniques; voltage reduction circuits in the weak inversion
techniques region
PPST (partial-product summation tree) stability analysis, Fox’s method 227
255–6 state space circuits 193
precomputation 6 strong inversion saturation (MOS and
programmable inverters: see inverters, FGMOS) 20
programmable subthreshold biasing 4
pseudo-FGMOS transistors 33–5 supply multipliers, for voltage reduction 3
PSIFG (programmable switched current switched-current (SI) memory cells: see
floating gate) cell 73–9 PSIFG (programmable switched
current mirrors 74 current floating gate) cell
differential PSIFG cell 76–7
second order filters 77–9, 80
switched-current (SI) cells 73–4, 75 TDDB (time-dependent-dielectric-
with transconductance amplifiers breakdown) 8
75 technology scaling 6
PSRR (power supply rejection ratio) TFT (thin film transistors) 49
111–15, 148–52, 163–5, 201–2 TG: see threshold gates (TG)
THD (total harmonic distortion)
with log-domain filtering 231–5
Q-programming, Gm -C FGMOS integrator in the strong inversion ohmic region
207, 208 103–11, 116–17, 119, 120
quality factor, IF filter 165, 167 in the strong inversion saturation region
quasi-FGMOS transistors 33–5 143–5, 146, 155–6
in the weak inversion region 198–201,
205–7
RC active filters 88
theory: see FGMOS theory
thin film transistors (TFT) 49
saturation region: see filtering, analog threshold gates (TG) 245–71
continuous-time, based on FGMOS about threshold gates 245–6, 268–9
in the strong inversion saturation definition 246–7
region notation 269–71
second harmonic distortion 108, 109 see also neuron MOS (νMOS) threshold
simulation/simulators 22–8 gate (TG)
(νMOS) threshold gates 250 TP (translinear principle) 186–8
convergence problems/techniques FGMOS translinear loop 186, 187
Ramírez-Angulo et al. solution see also Gm -C FGMOS integrator using
22–3 translinear circuits
Yin et al. solution 23, 24 transconductance/transconductors
initial transient analysis (ITA) technique basic blocks 89–90
23–8 FGMOS-based intermediate frequency
inverters 25–8 158, 159
iterative simulations 23 input transconductance 19
models for, lack of 22 transconductance amplifiers, with PSIFG
Spectre simulator 27 cells 75
single poly technology: see metal-poly transconductance cells 97–9
capacitors (single poly technology) with translinear circuits 189–93
304 Index
translinear circuits/principle: see Gm -C nonlinear processing of the signals 3–4

FGMOS integrator using translinear rail-to-rail operating range circuits 3
circuits in the weak inversion region supply multipliers 3
tunnelling/tunnel effect 8 voltage scaling 5
voltage scaling 5
ultraviolet conductance 10–11
UV (ultraviolet) cleaning 28–9 weak inversion
about weak inversion 185–6
VCVS (voltage controlled voltage sources) weak inversion saturation (MOS and
22 FGMOS) 20
VLSI (very large scale integration) 1, 2 see also Gm -C FGMOS integrator using
voltage reduction techniques translinear circuits in the weak
cascading stages 3 inversion region
Low Power and Low Voltage Circuit
Design with the FGMOS Transistor
Motivated by consumer demand for smaller, more portable electronic devices that Esther Rodriguez-Villegas
offer more features and operate for longer on their existing battery packs, cutting edge is a lecturer in the Circuits
electronic circuits need to be ever more power efficient. For the circuit designer, this and Systems group in the
requires an understanding of the latest low voltage and low power (LV/LP) techniques, Department of Electrical and
one of the most promising of which makes use of the floating gate MOS (FGMOS) Electronic Engineering at Imperial
College, London. Born in
transistor.
Spain, she graduated from the
Whilst a conventional MOS transistor has only one input, the FGMOS transistors often Department of Physics at the
have several. This fact, along with some other remarkable properties of this very University of Seville, receiving
interesting device, offers the designer many extra degrees of freedom that can significantly the San Alberto award. She
improve power efficiency. By using FGMOS transistors in the right way – establishing went on to achieve her PhD
degree from the same university.
appropriate relationships between their inputs – it is possible to achieve design trade-offs
After receiving a professional
that are not possible with conventional MOS devices. This is especially true when power research grant from the Spanish
consumption and supply voltage are the main design constraints. government she joined the
This book demonstrates how FGMOS transistors can be used in a low voltage and low Microelectronic National Centre,
power design context. The techniques shown provide innovative solutions, often in where she worked for almost six
years. Before moving to Imperial
situations where the limits of the technology in question have been pushed far below the
College Rodriguez-Villegas was
values recommended by the manufacturer. an Associate Professor at the
University of Seville for two years.
The Institution of Engineering and Technology

www.theiet.org
0 86341 617 9
978-086341-617-0

(Esther Villegas Low Power and Low Volt

Uploaded by

Copyright:

Available Formats

(Esther Villegas Low Power and Low Volt

Uploaded by

Document Information

Original Title

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

(Esther Villegas Low Power and Low Volt

Uploaded by

Copyright:

Available Formats

IET Circuits, Devices and Systems Series 20

Low Power and Low

Low Power and Low Voltage

The Institution of Engineering and Technology

British Library Cataloguing in Publication Data

ISBN (10 digit) 0 86341 617 9

Typeset in India by Newgen Imaging Systems (P) Ltd, Chennai, India

2 The Floating Gate MOS transistor (FGMOS) 15

3 FGMOS – Circuit applications and design techniques 43

4 Low power analog continuous-time filtering based on the FGMOS

5 Low power analog continuous-time filtering based on the FGMOS

5.2.2 The Common-Mode Feedback Circuit (CMFB): effects

6 Low power analog continuous-time Gm -C filtering using the

7 Low power log-domain filtering based on the FGMOS transistor 215

7.5 A design example 235

9 Summary and conclusions 273

The implantation of electronic devices into biological systems is an exciting new

1.1 The general framework

1.1.1 Why analog? Why digital?

1.1.2 Why low voltage?

1.1.3 Why low power?

1.1.4 Why CMOS?

1.2 Techniques to reduce the power consumption and the

1.2.1 Analog techniques

1.2.1.1 Techniques for voltage reduction

1.2.1.2 Techniques for current reduction

1.2.2 Digital techniques

2 t is the delay time. f is the operating frequency.

1.3 Why floating gate MOS?

Flexibility, Controllability, Tunability

Flexibility to implement both linear as well as complex nonlinear functions in a

1.4 FGMOS history

1.5 Structure of the book

Chapter 4: ‘Low power analog continuous-time filtering based on the

2.2 The Floating Gate MOS (FGMOS) device

2.2.1 Introducing the device

3 Fig. 2.1 is not drawn to scale.

Figure 2.2 Equivalent schematic for an n-channel N-input FGMOS transistor

2.2.2.2 Small signal equations

Table 2.1 MOS versus FGMOS model equations

Weak inversion saturation when VBS = 0

Strong inversion ohmic

Strong inversion saturation

(3) (2) (3) CGD (3) −1

2.2.2.3 Noise and dynamic range

where wi is the capacitive weight corresponding to the input i and (vin

2.3 Designing with FGMOS: problems and solutions

Figure 2.5 Ramírez-Angulo’s model for simulation of the FGMOS

Figure 2.6 Model in [122] for the FGMOS

VDD = 0 VSS = 0 VDD VSS

.ic v(fg2 )=v(fg2 )op

Figure 2.7 Simulation technique based on an initial transient analysis (ITA)

Floating gate Floating gate

.subckt transistorn vg vd v1 v2 v3 vs w=20u l=20u

.subckt transistorp vg vd v1 v2 v3 vs w=20u l=20u

x1 vg vd v1 v2 v3 0 transistorn w=20u l=20u c1=210f

x2 vg vd v1 v2 v3 vdd transistorp w=20u l=20u c1=210f

v1 v1 0 pulse(0 a 0.01 0.01 1 2 4)

v3 v3 0 pulse(0 2 0.01 0.01 1 2 4)