Ex JUnj JGve NBe DFN

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/216883082
Cascaded Microresonator-Based Matrix Switch for Silicon On-Chip Optical

Interconnection
Article in Proceedings of the IEEE · August 2009

DOI: 10.1109/JPROC.2009.2014884
CITATIONS READS
187 938
4 authors, including:
Xianshu Luo Hui Chen

Institute of Microelectronics Sun Yat-Sen University
129 PUBLICATIONS 1,246 CITATIONS 63 PUBLICATIONS 785 CITATIONS
SEE PROFILE SEE PROFILE
Some of the authors of this publication are also working on these related projects:
modern optics View project
silicon nitride View project
All content following this page was uploaded by Hui Chen on 25 December 2013.
The user has requested enhancement of the downloaded file.

INVITED
PAPER
Cascaded Microresonator-Based
Matrix Switch for Silicon
On-Chip Optical Interconnection
Experiments have demonstrated proof of the concept that cascaded arrays of
optical microresonators can switch interconnections as desired between
multiple signal inputs and outputs without significant distortion.
By Andrew W. Poon, Member IEEE , Xianshu Luo, Student Member IEEE ,
Fang Xu, and Hui Chen
ABSTRACT | This paper reviews developments in cascaded I . INTRODUCTION

microresonator-based matrix switches for silicon photonic Networks-on-chip (NoC) architectures for next-generation
interconnection networks in many-core computing applica- high-performance many-core computing systems have
tions. Specifically, we emphasize our recently proposed 5 5 attracted substantial research interest in recent years [1],
matrix switch comprising two-dimensionally cascaded micror- [2]. Many-core computing has now reached eight cores in
ing resonator-based electrooptic switches coupled to a wave- 2007 [3], and the number of cores per chip is expected to
guide cross-grid on a silicon chip. The cross-grid adopts
be doubled every 18 months [4]. Therefore, one particular
low-loss low-crosstalk multimode-interference-based wave-
interest is whether one may partially replace the conven-
guide crossings. Such a microresonator-based matrix switch tional metallic interconnection by alternative intercon-
offers nonblocking interconnections among multiple inputs nection technologies [5]–[7]. As pointed out in the
and multiple outputs, with the key merits of i) a tens to hun- International Technology Roadmap for Semiconductors
dreds of micrometers-scale footprint, ii) gigabit/second-scale (ITRS),1 optical communications on a chip is one
data transmission, iii) nanosecond-speed circuit-switching, alternative interconnection technology that promises
iv) 100-W-scale dc power consumption per link, and high-data-rate signal transmission and low power con-
v) large-scale integration for networks-on-chips applications.
sumption. On-chip optical interconnection thus promises
We analyze in detail the microring resonator-based cross-grid
a new paradigm in computer system design and also
switch design for high-data-rate signal transmission in the opens up an entirely new set of associated fundamental
context of our proposed 5 5 matrix switch. We also study the problems and engineering challenges spanning materials
feasibility of large-scale integration of the matrix switch. We science, nanoelectronics and nanophotonics, and on-chip
report proof-of-concept experiments of a single cross-grid networking [8]–[12].
switch element and a 2 2 matrix switch, propose design To this end, we believe that silicon photonics offers a
guidelines, and discuss future engineering challenges. promising technology platform. Over the past five years,
along with a growing number of research groups, we have
KEYWORDS | Matrix switch; microresonator; networks on chips;
witnessed an unprecedented explosion in the field of silicon
optical interconnect; silicon waveguide photonics or Group IV photonics [13]–[20]. Thanks to the
enabling complementary metal–oxide–semiconductor
(CMOS) fabrication technologies, we now have created a
Manuscript received November 13, 2008. Current version published June 12, 2009. myriad of essential silicon photonic passive and active
This work was supported by the Research Grants Council of Hong Kong SAR, components including low-loss submicrometer-sized wave-
China, under Projects HKUST6254/04E and 618707.
The authors are with the Photonic Device Laboratory, Department of Electronic and guides [21], micrometer-scale optical filters [21]–[24],
Computer Engineering, The Hong Kong University of Science and Technology, gigabit/second-speed low-power-consumption (pJ/bit)
Hong Kong SAR, China (e-mail: eeawpoon@ust.hk; eexsluo@ust.hk; xufzdece@ust.hk;
chenh@ust.hk). 1
See http://www.itrs.net/Links/2007ITRS/2007_Chapters/2007_
Digital Object Identifier: 10.1109/JPROC.2009.2014884 Interconnect.pdf.
1216 Proceedings of the IEEE | Vol. 97, No. 7, July 2009 0018-9219/$25.00 2009 IEEE
Authorized licensed use limited to: CAMBRIDGE UNIV. Downloaded on July 21, 2009 at 06:22 from IEEE Xplore. Restrictions apply.
Poon et al.: Cascaded Microresonator-Based Matrix Switch for Silicon On-Chip Optical Interconnection
modulators [25]–[28], hybrid-integrated evanescent lasers

[29], [30], and gigahertz-bandwidth SiGe photodetectors
[31]–[33]. It is these demonstrated silicon photonic devices
and technologies that make on-chip optical interconnection
viable.
In order to realize on-chip optical interconnection, it is
of essence to consider the following performance metrics
from system to device levels:
• bandwidth (capacity);
• power consumption;
• footprint;
• scalability.
Silicon wire waveguides are broadband and low-loss in
the 1550-nm telecommunication wavelengths. Silicon opto-
electronic devices such as modulators, switches, and
photodetectors have attained gigahertz bandwidths. Power
consumption on a chip concerns both optical and electrical
power consumption. One major source of optical power
dissipation is optical link loss, particularly in the case that
the link comprises many waveguide intersections (crossings)
that are typically lossy. In order to attain low electrical power
consumption, modulators, switches, and photodetectors all Fig. 1. Schematic hierarchy of on-chip optical interconnections for a
need to be powered at low voltage. Devices in optical many-core computing chip from network to switch elements.
interconnects should occupy a compact footprint and enable (a) Conceptual schematic of optical network-on-chip in mesh topology.
large-scale integration on a chip. In light of these The network is separated into an electronic layer (plane in gray) and an
considerations, we begin this paper with an overview of optical layer (plane in white). In the electronic layer, blue dashed lines
represent metal wire interconnections for control message exchange.
photonic NoCs in a top-down approach from the network In the optical layer, green lines represent bus optical waveguides. Each
level to the switch device level. Given such a wide scope, it is tile in the many-core chip has a corresponding 5 5 router (blue box)
necessary that the review here be brief and nonexhaustive. and an E/O interface (red box) in the optical layer. The electronic layer
communicates data with the optical layer through the E/O interface
A. Photonic Networks-on-Chip Architectures (red line), and controls the router through the control link (blue dashed
line). (b), (c) Router topologies. (b) Cross-grid array. Red lines show the
We classify photonic NoC architectures [10], [34]–[39] optical links established through the turned-on switches (in green).
into two classes: i) passive networks using fixed wave- (c) Cascaded 2 2 switch array. (d)–(f) Typical 2 2 switch element
length assignment and ii) switching networks. While structures. (d) A microring resonator coupled to a waveguide
passive networks offer only limited scalability and cross-grid (crossing). (e) Second-order coupled-microring resonators
complicated design [36], switching networks [39] are in coupled to two parallel waveguides. (f) Waveguide-based
Mach–Zehnder interferometer.
general more favorable for high-bandwidth on-chip
communications. It is, however, necessary to integrate
microelectronic control technology and photonic signal
transmission in the switching NoC design, as it is difficult controlling the photonic switching network. The photonic
to realize signal processing in optics. network routes packets of optical signals across the chip,
To date, it is not clear how to integrate the electronic with potentially low latency and low power dissipation. The
and photonic device layers on a single computer chip. It is router is a 5 5 matrix switch that interconnects four
generally believed that the integration may be in the form orthogonal directions with local injection/ejection ports. We
of multilayer stack architecture. The photonic network note that optical routers (switches) only need to be switched
layer design may follow the same topology as the on when establishing an optical circuit (discussed below) and
underlying electronic control network [1], [2], [34]. do not need to be switched on and off for every optical data
Fig. 1(a) depicts the schematic of a photonic NoC bit at high data rates. Furthermore, ultra-high-bandwidth
comprising a photonic interconnection layer vertically optical signal transmission can be attained by means of
integrated with an electronic control layer. The photonic wavelength-division multiplexing (WDM) technology [10],
device layer includes: [40]. Unlike electrical interconnections, there is no practical
• optical bus waveguides; on-chip photonic buffer that enables packet-switching as in
• optical routers; fiber-optic networks. This is likely to limit the on-chip
• electrical–optical (E/O) interfaces. photonic signal transmission to an end-to-end transmission
The electronic network is responsible for exchanging once the light path is established without buffering. This is a
short messages among the processor cores and also circuit-switched on-chip photonic network [34], [35].
Vol. 97, No. 7, July 2009 | Proceedings of the IEEE 1217

It is helpful to highlight two specific recent studies in based switches, the FCD effect shifts the resonance
literature on silicon photonic NoCs for many-core wavelength upon only a small phase change. This
computing applications. Notably, Shacham et al. [41] in translates into a highly sensitive and power-efficient
2007 proposed a modified two-dimensional (2-D) mesh switching for microresonators with sharp resonances.
photonic NoC using 4 4 blocking routers. Beausoleil et al. Moreover, microresonator-based switches are typically
[10] in 2008 proposed a nanophotonic interconnect design compact in size (a few tens of micrometers), and thus in
comprising waveguide buses connecting to all the cores, general require lower power for the required phase
targeting a 256-core chip. Both of these optical intercon- change. Typical silicon microresonator-based switch power
nect designs utilize silicon microring resonators as critical consumption reported in the literature is about a few
building blocks. W=m [25]. For MZI-based switches, the FCD effect
needs to -phase shift one arm of the MZI. This translates
B. Optical Router and Switch Structures into large device footprint (hundreds to thousands of
There are two major topologies for on-chip optical micrometers) due to the relatively small FCD effect in
routers: i) cross-grid and ii) Banyan network. Fig. 1(b) silicon [see Section II-B (1)]. Thus, compared with
illustrates a cross-grid matrix switch acting as a 5 5 microresonator-based switches, MZI-based switches are
nonblocking router. At each intersection of the horizontal generally larger in size and impose relatively larger power
(input) and vertical (output) waveguides is a 2 2 switch. consumption for the phase tuning.
The switch at a cross-state (green box) routes the optical In terms of scalability, the microresonator cross-grid
signal from the input to output waveguides, whereas the switch enables topologically simple crossbar architecture
switch at a bar-state (white box) keeps the optical signal for integration. The tradeoff, however, is that the
propagating straight through the bus waveguide. The additional insertion loss and crosstalk at the waveguide
optical switches can be in the form of a micrometer-scale crossing may end up limiting the scalability. The key
ring (microring) resonator coupled to a waveguide therefore is to minimize the cross-grid insertion loss and
crossing, as shown in Fig. 1(d). crosstalk. While the coupled-microresonator switch with
Fig. 1(c) illustrates the Banyan network acting as a 4 4 both the bar- and cross-state in the same direction as the
nonblocking router comprising six 2 2 switches [42]. input light offers a natural architecture to form a Banyan
The link path in this Banyan network passes through only network without waveguide intersections, the tradeoff is
three cascaded switches. Typical optical switch structures the more demanding fabrication resolution needed in
are a coupled-microring resonator-based switch and order to precisely define the mutually coupled micro-
Mach–Zehnder interferometer (MZI)-based switch, as resonators for each switch element. The MZI, however, is
shown in Fig. 1(e) and (f). in general much more tolerant to fabrication imperfections
Fig. 1(d)–(f) illustrates the three most common types of due to its broadband characteristic, and thus is scalable.
integrated 2 2 optical switch elements applicable to matrix However, the MZI-based network can take up a relatively
switches: the microresonator-coupled waveguide cross-grid, large footprint due to the size requirement for the FCD-
coupled-microresonators with parallel bus waveguides, and based -phase tuning.
MZI-based switch. These switches can be based on refractive All in all, we believe the microresonator-based switch
index change induced by slow (millisecond-speed) thermal is more favorable for realizing compact size and low-power
effect or fast (nanosecond-speed or faster) free-carrier consumption switching for on-chip interconnection. For
dispersion (FCD) in silicon (see Section II-B). For optical the rest of this paper, we will focus on microresonator-
interconnections in many-core computing chips, FCD-based based matrix switches.
switching should be preferred.
Here, we contrast these three types of switch structures C. Microresonator-Based Array and Matrix Switches
in terms of bandwidth, power consumption, footprint, and Cascaded microresonator-based array filters and
scalability. Microresonator-based switches (cross-grid and switches for multiple-input multiple-output (MIMO)
coupled microresonators) are inherently narrow-band due optical interconnects have long attracted research interest.
to the sharp resonances, while the MZI switch is It is only in recent years that researchers have proposed
inherently broadband. For high-data-rate optical signal cascaded microresonator-based architectures specifically
transmission, we must pay attention to carefully match the for multicore communications in a computer chip.
data bandwidth with the switch bandwidth in order to Table 1 provides a partial list of representative
minimize band-limiting signal distortion. While the cross- microresonator array structures relevant to optical inter-
grid switch assuming a single microresonator exhibits only connects in different materials including silicon [36],
a Lorentzian passband, the coupled microresonators allow [43]–[47]. There are again two main types: i) cross-grids
more room to obtain a box-like passband by tailoring the and ii) Banyan networks. For the cross-grid type,
coupling coefficient between the two microresonators. Little et al. in 2000 [44] first demonstrated the basic
The switch power consumption relates to both the architecture of microring resonators coupled to an 8 8
switching mechanism and the footprint. For microresonator- waveguide cross-grid array on a glass substrate. Thereafter,
1218 Proceedings of the IEEE | Vol. 97, No. 7, July 2009

Table 1 Microresonator Arrays for Multiple-Input Multiple-Output Optical Interconnects
researchers have demonstrated such type of microring Most recently, a thermal-optically tunable 8 8 array
resonator-based cross-connect array filters and their microring resonator-based optical matrix switch was
building blocks in various materials, including compound demonstrated [45]. The matrix switch comprises 64
glasses [48], [49], silicon nitride-on-silica substrates [50], third-order coupled-microring resonators coupled to a
and silicon-on-insulator (SOI) substrates [51], [52]. waveguide crossing array on a silicon oxynitride (SiON)

substrate (which does not allow electrooptic tuning). This wavelength tuning on the resonators. Such design im-
work, to our knowledge, demonstrated the largest number poses multiple wavelength channels for wavelength-agile
(192) of two-dimensionally cascaded microring resonators. switching.
It also indicates that a cascaded microresonator-based
matrix switch for silicon on-chip optical interconnection D. Paper Outline
should be feasible. The rest of this paper is organized as follows. Section II
Perhaps the most relevant work was done by Bergman’s discusses the nuts and bolts of our earlier proposed
group. Shacham et al. [41] in 2007 proposed and nonblocking microresonator-based 5 5 matrix switch
numerically analyzed a 4 4 blocking photonic switch and the basic switching mechanism of a microresonator-
on SOI comprising eight silicon microring resonators based switch coupled with a multimode-interference
laterally coupled to a waveguide cross-grid. There each of (MMI) waveguide crossing. Section III analyzes in detail
the waveguide crossings is laterally coupled to two the design of the microring resonator-based switching
identical microring resonators at two diagonal quadrants, element to meet the bandwidth and low-power consump-
and each crossing is designed to simultaneously route two tion requirements. Section IV presents the analysis of
links. Nonetheless, such design was internally blocking, large-scale integration of the 5 5 matrix switch taking
requiring advanced routing algorithms to improve the into account Gaussian distributed geometrical deviations
performance. Later on [53], [54], they proposed another among the microring resonators. Section V shows our
design by increasing the number of internal paths and the proof-of-concept experimental demonstration of a switch
crossing numbers within the switch in order for nonblock- element and a 2 2 matrix switch. Section VI summarizes
ing performance while still keeping eight microring this paper and suggests possible design guidelines and
resonators. further engineering challenges.
Notably, in collaboration with Bergman’s group,
Sherwood-Droz et al. from Cornell University demon-
strated such 4 4 nonblocking silicon router [39]. They I I. NONBLOCKING MICRORE SONATOR-
adopted eight microring resonators with 10-m radius and BASED 5 5 MATRI X SW IT CH
ten waveguide crossings, with a footprint of 0.07 mm2 . The
dynamic switching was realized by thermal-optical (TO) A. Waveguide Wires and Waveguide Crossings
tuning. Their design enabled bidirectional communications Waveguide wires are basic photonic components for
among four input and four output ports. The waveguide on-chip optical interconnection. A microresonator needs
crossings showed a measured insertion loss of 0.51 dB to be coupled to waveguides or waveguide crossings to act
(0.18 dB from simulation). as a photonic network component. The waveguide optical
However, the shortcoming of using the TO effect is loss largely depends on the waveguide dimensions and
that the switch on/off times would be limited to a few sidewall roughness [55]. Low-loss submicrometer-sized
microseconds, and the power consumption for the silicon wire waveguides have been demonstrated [21],
resonance wavelength tuning would be relatively high [56], [57]. Notably, the IBM Inc. group in 2007 [21]
(0.25 nm/mW in their demonstration). Moreover, in a 2-D reported silicon wire waveguides with propagation loss
mesh/torus network, a 5 5 nonblocking switch function down to 1.7 dB/cm. Such a propagation loss is consid-
is required in order to allow simultaneous two-way ered sufficiently low to support global interconnection
communications in four orthogonal directions with data on a chip.
injection and ejection. Groups of such 4 4 routers are Waveguide crossings are potential building blocks for
needed and that should further complicate the switch photonic NoCs. Nonetheless, there exists additional
network design. insertion loss and light leakage to the intersecting
For the Banyan network, Soref et al. [43] in 1998 waveguide (crosstalk) due to light scattering of the
first proposed an N-wavelength M M cross-connect expanded mode field at the waveguide sidewall intersec-
switch using coupled-microring resonator-based 2 2 tions. The scattering loss and crosstalk problems are
electroabsorptive switches on a semiconductor chip. particularly severe in high-index-contrast submicrometer-
The cross-connect architecture is nonblocking and poten- sized waveguide crossings. Reference [38] shows that
tially scalable without imposing waveguide crossings. silicon submicrometer-sized wire waveguide crossings
Recently, Goebuchi et al. [46] demonstrated such a type exhibited an insertion loss exceeding 1-dB and about
of optical cross-connect circuit using a coupled-microring 10-dB crosstalk. Such levels of loss and crosstalk are
resonator-based 2 2 thermooptic wavelength channel detrimental for forming a large-scale-integrated waveguide
switching on a doped silica glass substrate. Furthermore, cross-grid network.
Kazmierczak et al. [36] in 2007 proposed an alternative For NoC applications, we believe there are six essential
structure using coupled-microdisk resonator-based switches criteria for waveguide crossings design:
coupled to waveguide crossings. There the passive matrix i) low loss;
switch design is nonblocking, and it avoids active resonance ii) low crosstalk (to the intersecting waveguide);

iii) broadband; induced scattering loss and crosstalk. Our MMI crossing
iv) nonblocking; features:
v) compact footprint; i) small micrometer-scale footprint (the demon-
vi) high fabrication tolerance. strated MMI waveguide width is 1.1 m and
Broadband characteristic is essential for multiple length is 4.3 m in [61]);
wavelength transmission should WDM technology be ii) broadband transmission response;
adopted for the NoC. The nonblocking characteristic iii) single-layer fabrication process with high tolerance.
means that the different light paths intersecting can Fig. 2(a) illustrates the numerically simulated mode-
simultaneously cross over the junction without blocking field pattern of a silicon MMI crossing using 2-D
each other. finite-difference time-domain (FDTD) method. We see
In order to mitigate the loss and crosstalk induced by the self-imaged mode field is confined in the central region
the crossing junction, one proposal is to employ vertical of the crossing, and the mode field only induces a low
coupling between crossed waveguides at different layers level of scattering loss and crosstalk to the intersecting
separated by a thin low-refractive-index spacer layer [48]. waveguide. The inset depicts the simulated mode-field
However, this three-dimensional (3-D) multilayer archi- pattern of a conventional plain waveguide crossing of
tecture necessarily imposes more complicated fabrication the same bus waveguide width. There the scattering at
than 2-D monolithic architecture. the crossing junction results in significant radiation loss,
For planar monolithic devices, researchers have long crosstalk to the intersecting waveguide, and backward
investigated several schemes to improve the integrated reflection as suggested by the standing-wave field pattern
waveguide crossing performances for low loss and low along the input waveguide.
crosstalk [38], [48], [58]–[60]. Table 2 summarizes some Fig. 2(b) and the inset show our simulated transmis-
representative crossing structures of submicrometer-sized sions and crosstalks of the MMI crossing and the plain
waveguide crossings in high-index-contrast materials. crossing. The MMI crossing reveals 0.12 dB insertion
Previously [61], some of us demonstrated MMI-based loss (0.7 dB improvement from the plain crossing) and
waveguide crossings in SOI substrates. The concept overall better than 40 dB crosstalk (20 dB improve-
follows the MMI-based waveguide crossings, which were ment from the plain crossing) in the 1500–1600 nm
first numerically analyzed in low-index-contrast silica wavelengths [52]. Our MMI crossing design thus basically
waveguides [59]. Two identical MMI waveguides orthog- meets the above essential criteria for NoC applications.
onally intersect to form the waveguide crossing. We
leverage the self-imaging property of the multimode B. Microresonator-Based Switch Elements
waveguides in order to focus the mode field at the center Fig. 3(a) illustrates the switching mechanism of a 2 2
of the crossing, thereby minimizing the junction sidewalls microresonator-based switch element. Our switch element
Table 2 Low-Loss High-Index-Contrast Wire Waveguide Crossings

Lorentzian lineshape filter profile. Upon a change in the

microresonator refractive index, the resonance center
wavelength shifts accordingly from a so-called electrical
off-state resonance wavelength to an electrical on-state
resonance wavelength. The resonance wavelength tuning
enables a switch of light route at a probe (carrier)
wavelength c aligned to the on-state resonance wave-
length. In the case that the microresonator switch is at an
off-state, the light wavelength c is off-resonance and
propagates straight through from the input- to throughput-
port. In the case that the resonance wavelength is shifted
to align with c , the microresonator switch is at an on-state
and the light is resonantly transferred from the input- to
drop-port via the microresonator.
In silicon microresonators, fast resonance wavelength
tuning has been realized using the electrooptic ef-
fectVnamely, the plasma dispersion effect [64]Vby
means of carrier injection [25], [62], [65]–[67] and
depletion [27]. For the carrier-injection type, the micro-
resonators are integrated with a lateral p-i-n diode that
wraps around the microresonator. Carriers are injected
into the microresonator mode volume upon forward bias,
and the injected carriers are typically swept out upon
reverse bias to obtain a fast switch off without being
limited by the slow carrier recombination. For the carrier-
depletion type, the microresonators are integrated with a
p-n diode that wraps around the microresonator. Carriers
are depleted from the moderately doped microresonator
mode volume upon reverse bias.
Fig. 2. (a) FDTD-simulated TE-polarized intensity profile of the MMI Here, we adopt carrier-injection type microring
crossing at wavelength 1550 nm with a bus waveguide width of 0.4 m. resonator-based switches. Fig. 3(b) shows the top-view
Inset: simulated intensity profile of a plain crossing with a bus schematic of our switch design. The microring resonator is
waveguide width of 0.4 m. C: cross-port, T: throughput-port.
nearly wrapped completely by a ring-shaped lateral p-i-n
(b) FDTD-simulated TE-polarized throughput-port transmission
spectra of the silicon MMI crossing (red solid line) and the plain diode, following the pioneering work by Xu et al. [25]. The
crossing (blue dashed line). Inset: simulated cross-port transmissions inset shows the cross-sectional view of the microring
of the MMI crossing (solid line) and the plain crossing (dashed line). waveguide mode [numerically simulated by the beam
(Reproduced with permission from OSA [52].) propagation method (BPM)]. We note that in order to
minimize the light scattering loss due to the spatial overlap
between the waveguide mode evanescent field and the
comprises a single silicon microresonator-based electro- highly doped pþ and nþ regions, a separation spacing of
optic switch side-coupled to an MMI crossing. The 0.3–0.5 m between the highly doped region edge and
microresonator is positioned in the waveguide crossing the waveguide sidewall is essential. This constitutes an
quadrant that allows light to be routed via the micro- intrinsic region of 1 m for the lateral p-i-n diode.
resonator prior to encountering the crossing. This We note that the intrinsic region width is one key
minimizes the effects of the crossing junction loss and factor that affects the switching times [68]. The electrical
crosstalk to the routed light channel. The microresonator switching times can be limited by the carrier transit times
can be in the form of a microring [25] or microdisk [62], when the microresonator is switched on upon forward-
[63]. Here, we adopt microring resonators due to their biasing and when it is switched off upon zero bias or
simple single-mode resonances and the ease of phase- reverse bias [25]. The narrower the intrinsic region width
matching between the microring and the coupled is, the shorter the transient time needed to reach the
waveguides of the same waveguide cross-sections. steady-state of the carrier concentration change upon a
The filter response follows an add–drop filter with the certain bias voltage [65].
throughput-port transmission’s exhibiting resonance dips According to the seminal work of Soref and Bennett,
and the drop-port transmission’s exhibiting resonance the empirical equations to describe the changes in the
peaks at the corresponding resonance wavelengths. The refractive index ðnÞ and the accompanied absorption
single-order microresonator-based filter gives rise to only coefficients ðÞ as functions of changes in free electron

Fig. 3. (a) Illustration of the switching mechanism of the 2 2 cross-connect switch element using a microring resonator side-coupled to a
MMI waveguide crossing. The throughput- and drop-port transmission spectra illustrate the microring resonance wavelength blue-shift
between the electrical off-state (orange solid lines) and the electrical on-state (blue dotted lines). c : carrier wavelength aligned to the on-state
resonance; : resonance linewidth (at the electrical off-state); 0 , 1 : waveguide-to-microring and microring-to-waveguide coupling coefficients
(assumed symmetrical); t0 , t1 : bus waveguide transmission coefficients at the in/output-couplers (assumed symmetrical); A: microring round-trip
amplitude transmission factor. I: input-port, T: throughput-port, D: drop-port. (b) Schematic of a silicon microring resonator-based
carrier-injection-type MMI cross-grid switch using a lateral p-i-n diode. Inset: simulated waveguide mode field distribution showing minimal
spatial overlap with the nþ - and pþ -doped regions.
and hole concentrations (Ne cm3 and Nh cm3 ) at C. 5 5 Matrix Switch Configuration
1550 nm wavelength are [64] Previously [69], we proposed a 5 5 matrix switch
using cascaded microring resonators laterally coupled to a
MMI cross-grid array with 25 microring resonators and
n ¼ ne þ nh 25 waveguide crossings. If we allow no self-communications
(e.g., communications between Ieast and Oeast , Iinjection and
¼ 8:8 1022 Ne þ 8:5 1018 ðNh Þ0:8 (1)
Oejection ), we can remove the five microring resonators along
¼ e þ h the diagonal direction. Therefore, the total microring
¼ 8:5 1018 Ne þ 6:0 1018 Nh : (2) resonator number in the matrix switch is reduced to 20.
Fig. 4(a) shows the modified 5 5 matrix switch with
20 microring resonators. The configuration is a generic
Carrier-injection based switches enable a refractive layout comprising five input-ports and five output-ports,
index change n of up to 103 range with a carrier enabling simultaneously two-way communications among
injection of 1017 to 1018 cm3 following (1). This four orthogonal directions and local injection and ejection.
enables a microcavity resonance wavelength blue-shift of All 20 microring resonators are designed to be centered at
the order of 1 nm. Equation (2) suggests that the the same resonance wavelengths within a relatively wide
injected free carriers also induce additional power loss subnanometer resonance linewidth (giving a Q-factor of
known as the free-carrier absorption (FCA) loss. FCA is the order of 103 ). The subnanometer-wide resonance
the main loss mechanism affecting the drop-port trans- linewidth (tens of gigahertz bandwidth in 1550-nm
mission (see Section III-C). wavelengths) is of the essence in order to transmit

Fig. 4(b) illustrates one possible physical layout of the

matrix switch for actual interconnection in a 2-D mesh/
torus network. In this design, the throughput-ports are
terminated by means of antireflection or absorption. We
assume 6-m-radius microring resonators (limited by our
current MMI crossing design). Such a matrix switch oc-
cupies less than 0.1 0.1 mm2 area, enabling 10 s 10 s
of such matrix switches in a 20 20 mm2 silicon chip.
We note that an additional waveguide crossing is used at
the intersection between ports Isouth and Oeast , making a
total of 26 waveguide crossings.
With proper design outlined below, we believe that the
following matrix switch performance metrics should be
feasible:
i) broadband transmission: optical signal transmis-
sion up to 5-Gbit/s data rate per wavelength
channel;
ii) low power consumption: few tens–100 mW-range
dc electrical power consumption;
iii) fast circuit-switching: nanosecond-speed response
using carrier injection;
iv) compact footprint: 100 100 m 2 scale
footprint;
v) network scalability for many-core interconnections.
I II . MICRORING RESONATOR-BASED
SWITCH DESI GN
Microresonator-based switches constitute the basic ele-
ment of our matrix switch. Each optical path comprises
electrical on-state and off-state switch elements. It is thus
of the essence to consider the cascaded filter effect when
Fig. 4. (a) Schematic of a 5 5 matrix switch comprising 20 identical designing the microresonator filter response. Here, we
microring resonator-based switch elements coupled to a MMI crossing discuss in detail the single-order microresonator-based
grid array with 25 crossings. The dashed arrows illustrate some of the
switch design as a building block for the matrix switch. We
possible light paths connecting 5 different input-ports (Iwest , Ieast ,
injection, Inorth , and Isouth ) to five different output-ports (Owest , Oeast , first review the modeling of a single microresonator-based
ejection, Onorth , and Osouth ). (b) Physical layout of our proposed 5 5 filter transmission and bandwidth and consider the effect
photonic matrix switch comprising 20 identical silicon microring of cascading multiple filters along an optical link routed
resonator-based switch elements coupled to a MMI crossing grid array through multiple matrix switches. Then we discuss the
with 26 crossings. The pink arrow depicts one of the longest links
optical link loss, switching electrical power consumption,
(injection-to-east). The blue arrow depicts one of the shortest links
(west-to-ejection). An additional MMI crossing is employed at the and response time. Furthermore, we briefly discuss
intersection between Isouth - and Oeast -ports. Black solid dots represent possible improvements in the switch performances using
termination at the unused ports. a third-order coupled-microring switch element with a
box-like filter lineshape.
gigabit/second data with minimum distortion [70]–[72] A. Filter Transmission and Bandwidth
and also enhance tolerance to fabrication imperfections. The waveguide-to-microresonator coupling efficiency is
The matrix switch works as a nonblocking crossbar one key parameter that determines the microresonator-
network. Each data path connecting one input-port to one based filter transmission and bandwidth. For microring
output-port is established by switching on only one switch resonator-based filters, the coupling efficiency depends
element (e.g., Isouth -to-Owest and Iwest -to-Osouth ). In such on the bus waveguide dimensions, the waveguide-to-
design, different data paths that crossover at a single grid microring interaction length, and the gap spacing
point can be simultaneously established [e.g., injection-to- between the bus waveguide and the microring. Given
Oeast and Inorth -to-ejection; see Fig. 4(a)], with low signal the coupling efficiency (in terms of field coupling and
crosstalk imposed at the MMI junction between the two transmission coefficients), we can model the field
data paths. transmission coefficients of the microresonator coupled

with a waveguide crossing by the transfer matrix method carriers to the microring, we express the resulting
[73], [74]. The formal expressions of the electric field electrical on-state round-trip phase as
transmissions at the throughput- and drop-ports ðTt ; Td Þ
are as follows:
ON ¼ 2ðneff þ neff ÞLr = (5)
Et t0 t20 t1 A expði Þ þ 20 t1 A expði Þ

Tt ðÞ ¼ ¼ (3) where neff is the effective refractive index change upon
Ein 1 t0 t1 A expði Þ
0:25
the FCD-induced refractive index change n [see (1)]. We
Ed 0 1 A expði0:25 Þ obtain ðneff þ neff Þ numerically using BPM. The FCA
Td ðÞ ¼ ¼ (4)
Ein 1 t0 t1 A expði Þ suppresses the electrical on-state round-trip amplitude
transmission factor as
where Et ðEd Þ is the electric field amplitude at the

throughput- (drop-) port, Ein is the electric field at the 1
input-port, x and tx (x ¼ 0, 1) are the field coupling and AON ¼ exp Lr A (6)
2
transmission coefficients of the input/output couplers (for
lossless coupling, jx j2 þ jtx j2 ¼ 1), A is the microring
round-trip amplitude transmission factor (A ¼ 1 means where is given by (2).
lossless microresonator), and is the microring round-trip Fig. 5(a) shows the calculated throughput- and drop-port
phase change [see Fig. 3(a)]. The term ðA expði ÞÞ0:25 in transmission spectra at electrical off- and on-states for a
(4) represents the field component that propagates a single switch element using (3)–(6). We choose a 6-m-
quarter of the microring between the horizontal and radius microring resonator in our study. The calculation
vertical bus waveguides. parameters are as follows: neff ¼ 2:65, A ¼ 0:9993,
We relate to the microring round-trip length Lr as 0 ¼ 1 ¼ 0:3. The throughput lineshape shows the rejec-
¼ 2neff Lr =, where neff is the waveguide effective tion filter response (inverted Lorentzian), whereas the
refractive index and is the free-space wavelength. We complementary drop-port transmission shows the band-
obtain neff numerically using the BPM. Upon injecting pass filter response (Lorentzian). Upon carrier injection, the
Fig. 5. (a) Calculated off-state throughput-port transmission (red line) and on-state drop-port transmission (blue line) of a single cross-grid switch
element. (b) Calculated off-state throughput loss ðÞ and on-state drop-port loss ðÞ with on-state resonance wavelength shifts. Calculated carrier
concentration change ðhÞ to attain such on-state resonance shifts. Lines are for visual aid. (c) Calculated transmission spectrum of optical link Iwest -
to-Osouth . (d) Calculated waveform of a 5-Gbit/s NRZ signal transmission upon optical link Iwest -to-Osouth . The input waveform is shown in black.

resonance wavelength blue-shifts and gives the electrical carefully design the value in order to balance the
on-state switch transmission response. The resonance throughput and drop loss. In our following analysis, we
wavelength shift is given by ¼ neff =neff . The choose a of 1.3 nm, giving an off-state throughput loss
resonance linewidth of 0.7 nm ðQ 2200Þ corresponds of 0.33 dB and an on-state drop-port transmission loss of
to a 3-dB bandwidth of 90 GHz. The filter bandwidth is a 1.64 dB. The drop-port transmission can only be optimized
critical design parameter for high-data-rate optical signal by balancing the in/out-coupling coefficients and the
transmission. The resonance linewidth of 0.7 nm in our microresonator total loss [76]. We note that in the
design is essential, considering the cascaded filter bandwidth literature, the typical 0.5 0.1 dB resonance drop-
narrowing effect (see Section IV). We note that the drop-port port transmission loss for a single switch element [23] is
response of the electrical on-state switch encounters extra only given by microresonators without FCA loss.
microresonator loss due to the FCA. As an example, we show here the spectrum and data
In our analysis, we assume a 5-Gbit/s nonreturn-to-zero transmission along the optical link of IWest -to-OSouth in a
(NRZ) data signal on a single wavelength channel. This 5 5 matrix switch [see Fig. 4(b)]. Fig. 5(c) shows
assumption, we believe, is reasonable for applications in the calculated transmission spectrum. Here, we account
many-core computing systems [4], [10], [75]. Therefore, for all the losses from the MMI crossings, the off-state
we target the switch element filtering bandwidth to be tens throughput-port losses, and the on-state drop-port loss. The
to 100 GHz (resonance Q-factor of 1000 s) in order to resulting transmission peak intensity suffers a power loss of
minimize the band-limiting signal distortion. Although 3.7 dB. The resonance linewidth remains 0.7 nm, as
single-order filters with lower Q-factors (larger band- there is only one on-state drop-port transmission in a single
widths) accommodate a higher-speed NRZ optical signal, optical link. We note that, however, the transmission
and are less sensitive to fabrication imperfections and lineshape is asymmetric due to the throughput transmissions
temperature variations, the drawbacks are that the at the tail of a Lorentzian lineshape. Such asymmetric filtering
Lorentzian lineshape exhibits a long tail that imposes passband may only cause waveform distortion when the
significant off-resonance crosstalk and also larger reso- filtering bandwidth is comparable to the signal bandwidth.
nance wavelength shift to switch on or off. Fig. 5(d) shows the calculated 5-Gbit/s NRZ signal
waveform routed through the optical link. We choose the
B. Optical Power Loss input signal rise- and fall-times of 72 ps. The intensity of
The optical power loss of the matrix switch mainly the NRZ pulse suffers insertion loss of 3.7 dB [see
comprises i) the off-state throughput-port transmission loss, Fig. 5(c)]. The signal rise- and fall-times remain 72 ps,
ii) the on-state drop-port transmission loss, and iii) the without any significant waveform distortion. Thus, the
waveguide crossing loss. Here, we neglect the waveguide switch 3-dB bandwidth of 90 GHz is sufficiently wide for
propagation loss (dB/cm [21]) for our submillimeter-sized 5-Gbit/s data transmission.
matrix switch. Thus, the optical power loss in a light path
routed by the matrix switches depends only on the number C. Electrical Power Consumption
of the on/off-state switch elements and the waveguide The electrical power consumption is dominated by the
crossings. In a long light path traversing many switch dc power consumption during the circuit switch-on time.
elements across the matrix switches, the number of off-state A practical dc power consumption for the carrier-
switch elements far exceeds the number of on-state switch injection-based switch element is in the range of tens to
elements. Therefore, the throughput loss accumulated hundreds of microwatts, given a subnanometer- to
across many off-state switch elements along a link is the nanometer-scale resonance wavelength blue-shift in a
dominating factor compared with the drop-port loss. few-micrometer-radius microresonator-based switch sur-
In our routing scheme, we adopt the signal wavelength as rounded with a p-i-n diode. The detuning determines
the on-state blue-shifted resonance wavelength. The on-state the refractive index change needed, and thereby the dc
drop-port transmission loss and the off-state throughput-port power consumption to setup and maintain a light circuit.
transmission loss at the signal wavelength depend on the We simulate the electrical characteristics of a carrier-
wavelength detuning from the off-state resonance injection-based switch by the MEDICI device simulator.
wavelength [see the pink circles denoted in Fig. 3(a)]. Fig. 6(a) shows the simulated steady-state refractive index
Fig. 5(b) illustrates the design tradeoffs of choosing change in silicon (blue line) following (1) and the
different values. The off-state throughput loss corresponding current density (red curve) as functions of
decreases with following the inverted Lorentzian applied voltages (the calculation method follows that in
lineshape. This suggests adopting a large . However, a [77] and [78]). A 1.3-nm wavelength blue-shift at around
large imposes a large free-carrier injection for the 1550-nm wavelengths requires a refractive index drop of
switching. The higher free-carrier concentration induces a about 2.2 103 in silicon, corresponding to a 1.4-A=m
higher FCA and raises the on-state drop-port transmission dc diode current. The inset shows the cross-section of the
loss. The on-state drop-port insertion loss thus increases simulated p-i-n diode integrated waveguide structure. We
linearly with wavelength blue-shifts. Hence, we need to choose the diode intrinsic region width as 1 m for a

Fig. 6(b) shows the modeled power consumption

imposed for a certain resonance wavelength shift in the
designed p-i-n diode switch. The power consumption
increases parabolically with the resonance wavelength
shift. For a 5-Gbit/s system that requires a 1.3 nm
resonance wavelength shift, the calculated power con-
sumption is 1.3 W=m.
D. Switching Times
We define the switch-on (off) time ton ðtoff Þ as the time
for the drop-port transmission to raise from 10% to 90%
(drop from 90% to 10%) of the peak transmission upon
forward bias (zero bias). For p-i-n diode-based carrier-
injection-type switches, the switching time is no less than a
nanosecond without a preemphasis driving scheme upon
such resonance wavelength shifts. The MEDICI-simulated
ton is 1-2 ns, and toff is also 1 ns depending on the
injected carrier concentration and the carrier recombina-
tion time. Here, we adopt only a low forward-bias voltage
of 1 V to setup the light path and zero-bias voltage to
close the drop-port transmission.
We note that although the demonstrated gigahertz-speed
carrier-injection-type modulators show subnanosecond
switching times (e.g., the 12.5-Gbit/s silicon microring
resonator-based modulator [65]), the high-speed modulators
always require preemphasis driving. There the rise-time is
shortened by a high forward-bias pulse to quickly inject the
carriers, while the fall-time is shortened by a high reverse-
Fig. 6. (a) Simulated steady-state carrier-injection-induced refractive biased voltage to quickly sweep out the injected carriers from
index change (blue curve) and current density (red curve) as functions
the diode intrinsic region. For matrix switch purposes, we
of the driving voltage. Inset: schematic cross-section of the lateral p-i-n
diode straddled across a silicon wire waveguide. J: current density. believe that the nanosecond-speed circuit switching is
BOX: buried oxide. (b) Simulated power consumption for carrier- sufficient, and therefore prefers a simple low-voltage
injection-based resonance wavelength blue-shifts. Dashed lines forward-only driving scheme.
denote the resonance wavelength shift needed and the corresponding
power consumption (per micrometer length) for switching signals with
a data rate of 5 Gbit/s.
E. Third-Order Coupled-Microring Switch Element
While the single microring-based switch element
promises the key merits of easy design and small footprint,
the shortcomings are the cascaded filter bandwidth
waveguide width of 0.4 m. We assume the diode pþ - narrowing and the excessive crosstalk from the off-state
doped region concentration of 2 1019 cm3 and nþ - drop-port transmission. Both bandwidth narrowing and
doped region concentration of 1020 cm3 . crosstalk issues are due to the Lorentzian lineshape of the
The switch electrical power consumption scales with single microring resonator. The off-state drop-port trans-
the length of the integrated p-i-n diode [40]. A 6-m- mission at a certain detuning can exceed 10 dB [see the
radius microring resonator that blue-shifts 1.3 nm in pink circle in Fig. 3(a)]. This crosstalk is significantly
1550-nm wavelengths consumes 48<W dc power with a larger than that from the MMI crossing induced cross-port
forward-bias voltage of 0.92 V. This estimation assumes transmission [G 40 dB; see Fig. 2(b)]. Although a larger
the entire microring is surrounded by the diode and detuning can minimize the off-state drop-port transmis-
neglects possible current leakage, which may significantly sion, the detuning is also constrained by the amount of
raise the power consumption. Previously [62], [68], our resonance wavelength tuning available upon practical
research group demonstrated in a p-i-n diode embedded carrier concentration change.
5-m-radius microdisk a 20<W dc power consumption One possible solution to these issues is to adopt high-
for a resonance wavelength blue-shift of only 0.06 nm with order coupled-microring switch designs [22], [45], [79]
a forward-bias voltage of 0.85 V. Xu et al. [25] dem- in order to obtain a box-like filter passband. We adopt third-
onstrated a p-i-n diode integrated microring resonator- order coupled-microring switch element coupled with cross
based modulator that exhibited a dc power consumption of grid. Fig. 7(a) shows the third-order coupled-microring
18–105 W. switch element, illustrating the light routing mechanism.

port transmission loss is also significantly reduced.

Furthermore, the crosstalk from the off-state drop-port
transmission is improved from 11 to 23 dB.
However, the drop-port transmission loss is increased due
to the longer total propagation path length of the three
microring resonators. The total power consumption
ð13 W=ring 3 rings ¼ 39 WÞ is still below that of
single microring switch element, thanks to the reduced
resonance wavelength detuning which requires a lower
driving voltage across each of the three microrings.
The design of a third-order coupled-microring switch
element, however, imposes a larger footprint and also
Fig. 7. Schematic of the 2 2 cross-connect switch element using a more complicated design of the resonators, specifically the
third-order coupled-microring resonator side-coupled to an MMI mutual coupling between the microresonators in order to
waveguide crossing. c : carrier wavelength aligned to the on-state
attain box-like filtering response. It is thus less tolerant to
resonance; 0 : waveguide-to-microring and microring-to-waveguide
coupling coefficient (assumed symmetrical); inter : intercavity coupling fabrication imperfections.
coefficient; A: microring round-trip amplitude transmission factor;
I: input-port; T: throughput-port; D: drop-port. (b) Calculated
throughput-port transmissions and drop-port transmissions at I V. ANALYSIS OF LARGE-SCALE
off- and on-state of a third-order coupled-microring cross-grid INTE GRAT ION
switch element.
Scaling the number of matrix switches for interconnec-
tions in a many-core network is in part limited by the
optical route loss and the optical power budget. Here, we
We assume symmetric coupling between waveguide-to- analyze the feasibility of network-size scaling of the matrix
resonator and resonator-to-waveguide with coupling coeffi- switch. We adopt the loss values based on our analysis in
cient of 0 . The microring mutual couplings are also Section III as follows: i) the off-state throughput-port
assumed to be identical with coupling coefficient of inter . transmission loss ðthr Þ of 0.33 dB; ii) the on-state drop-
Fig. 7(b) shows the calculated throughput- and drop- port transmission loss ðdrp Þ of 1.64 dB; and iii) the
port transmission spectra for both off- and on-state. The MMI crossing loss ðcr Þ of 0.12 dB [see Fig. 2(b)]. The
calculation parameters are as follows: r ¼ 6 m, waveguide propagation loss is neglected. We calculate
neff ¼ 2:65, A ¼ 0:9993, 0 ¼ 0:3, inter ¼ 0:05. With such the optical power loss in a light path within a matrix
a design, we obtain a box-like filtering response with switch with Moff off-state microring resonators and Nc
linewidth of 0.6 nm (75-GHz bandwidth). In order to MMI crossings as follows:
obtain optimized throughput- and drop-port losses, and given
the box-like filtering response, we choose in the analysis a
(dB) ¼ thr Moff þ cr Nc þ drp : (7)
relatively small resonance wavelength detuning of 0.54 nm.
Table 3 compares the calculated performance metrics
between the single microring and the third-order coupled- Table 4 summarizes the expected optical loss for the
microring switch elements. The third-order switch allows a light paths in the 20-microring matrix switch configuration
significantly reduced wavelength detuning. The throughput- [see Fig. 4(b)]. We denote the number of waveguide
Table 3 Comparison of the Switch Element Performances

Table 4 Path-Dependent Optical Loss Across a 20-Ring Matrix Switch
Fig. 8. (a) Schematic layout of an M M tiles chip with a representative edge link connecting two cores at the diagonal corners. Five types of
matrix switch interconnection along the link as described in the text. Calculated (b) transmission spectrum and (c) transmission waveforms of
a 5-Gbit/s NRZ signal for light passing through the edge link in a 4 4 tiles matrix switch array. Calculated (d) link loss and (e) bandwidth as
functions of M. The microring radius R is centered at 6 m (solid symbols, h, , and r) and 12 m (open symbols, g, , and ). We consider microring
resonators with no dimension variation (h and g), with dimension variation of 3
¼ 5:6 nm ( and ) and 3
¼ 20 nm (r and ). We assume a 50 dB
link loss is within the optical power budget.
crossings and off-state switch elements in brackets for trix switch includes 6 off-state throughput transmission
every light path. The optical loss in the longest light paths losses, 9 MMI waveguide crossing losses, and 1 drop-port
(injection-to-Oeast and Isouth -to-ejection) within the ma- transmission loss [see Fig. 4(b)]. This gives the longest

light path optical loss of 4.7 dB. While the shortest light reveals that the largest matrix switch array is 7 7 with
paths (injection-to-Onorth and Iwest -to-ejection) optical loss the link loss within an optical power budget of 50 dB.
comprising only one drop-port transmission loss and one For a more realistic analysis of the matrix switch
crossing loss is 1.76 dB [see Fig. 4(b)]. The optical link network scalability, we must take into account the adverse
loss is in general asymmetrical between the input and effect due to the inevitable fabrication imperfection
output directions (e.g. injection-to-Oeast loss 6¼ Ieast -to- induced nonuniformity in the microring dimension.
ejection loss), except that between Ieast -to-Osouth and Here, we assume a normal distribution of dimensional
Isouth -to-Oeast . We note that the 4.7-dB optical insertion difference ðxÞ with standard deviation ð
Þ is introduced
loss for a single matrix switch, however, still severely to the microrings involved in the light route. The round-
limits the possible number of cascaded matrix switches in a trip length difference from the designed value is
network route. L ¼ 2x, and the associated resonance wavelength shift
Here, we analyze the source-to-destination route loss is given by ¼ L=2r0 , which is inversely proportional
and bandwidth narrowing in an M2 -core M M tiles to the microring radius r0 . The resulting distributed
NoC. Fig. 8(a) depicts a representative optical link along resonance wavelengths along the route thus cause a power
the edge of the NoC interconnecting the corner tile in the penalty in the signal transmission from the ideal case. We
northwest and the corner tile in the southeast. The matrix note that a larger-sized microring is more tolerant to
switches along the edge-link comprise five different types fabrication imperfections. This simple but important fact
of interconnections: should be carefully considered in the tradeoff analysis
A) injection-to-Oeast ; among the switch footprints, cascade scalability, link
B) Iwest -to-Oeast ; bandwidth, and power consumption.
C) Iwest -to-Osouth ; The calculated route losses, accounting for various
D) Inorth -to-Osouth ; degrees of microring round-trip length variations, are
E) Inorth -to-ejection. shown in Fig. 8(d). We assume a lithographic deviation 3
For a four-core 2 2 NoC, the link comprises only single of 5.6 nm (assuming a 65 nm technology node2) and of
interconnections A, C, and E. For M > 2 chips, there are 20 nm (assuming nowadays widely used 248-nm DUV
(M2) number of cascaded B interconnections (4.25 dB) lithography technology [81], [82]) in 6- and 12-m-radius
between interconnections A and C and (M2) number of microring resonators. We average 40 trails of the modeling
cascaded D interconnections (2.66 dB) between inter- results to obtain the trends shown here. We purposely
connections C and E. Thus, as M increases by one, the edge- choose the modeling parameters in order to obtain nearly
link loss increases by 7 dB. identical off-state throughput-port and on-state drop-port
As an example, we consider the edge-link in a 16-core losses for both sized microrings. Thus, the link losses for
4 4 tiles NoC. The calculation parameters are the same as both sized microrings in the ideal case are nearly identical
those in Section III-A. There are in this link seven on-state with each other. The modeling parameters for 12-m-
drop-port losses, 28 off-state throughput-port losses, and radius microring resonator are as follows: neff ¼ 2:65,
43 MMI crossing losses. Fig. 8(b) shows the calculated A ¼ 0:9986, 0 ¼ 1 ¼ 0:4, and ¼ 1:3 nm, result-
transmission spectrum through the link. The cascaded filter ing in 0.30 dB off-state throughput-port loss and
passband shows a sharpened lineshape with linewidth of 1.71 dB on-state drop-port loss.
only 0.29 nm, corresponding to a 3-dB bandwidth of Our analysis reveals that optical routes comprising
only 36 GHz. This is a significant bandwidth narrowing 12-m-radius microring resonators display significantly
from the switch element bandwidth of 90 GHz. The lower power penalties from the ideal case than optical
sharpened passband is potentially undesirable for high- routes comprising 6-m-radius microring resonators. For
data-rate optical signal transmission [80]. The narrow 3
¼ 5:6 nm, the route loss with 12-m-radius microring
bandwidth also allows signal sideband transmission to the resonators results in 9-dB power penalty, and it is
throughput-port and thus causes crosstalk [72]. possible to support up to a 5 5 tiles NoC within a 50-dB
Fig. 8(c) shows the calculated 5-Gbit/s NRZ signal power budget. For the route loss with 6-m-radius
waveform transmission through the link. The waveform microring resonators, the power penalty is as large as
does not show significant distortion. Thus, the link 21 dB in a 5 5 tiles NoC. The core number is only
transmission bandwidth (36 GHz) is still sufficient to practically scalable up to 4 4 within the power budget.
support the 5-Gbit/s NRZ signal. The amplitude of the For 3
¼ 20 nm, the power penalty significantly increases
NRZ pulse, however, suffers a link insertion loss of 28 dB for both microring sizes. The possible core numbers
[see Fig. 8(b)]. enabled by the 12- and 6-m-radius microring resonator-
Fig. 8(d) shows the calculated optical route loss as a based matrix switches are only 3 3. This suggests that
function of tile array dimension M. For the ideal case under the common 248-nm DUV lithography technology
without any fabrication imperfections, we estimate the
route loss simply by summing up the link losses of the 2
See http://www.itrs.net/Links/2007ITRS/2007_Chapters/
different interconnections according to (7). Our analysis 2007_Lithography.pdf.

[56], our proposed 5 5 matrix switch can support 3 3 V. PROOF-OF-CONCEPT EXPERIMENTS

tiles (nine cores) NoC, while preserving a compact
footprint. A. Fabrication
Fig. 8(e) shows the calculated optical link 3-dB We fabricate a microring resonator-based cross-grid
bandwidth as a function of tile array dimension M. The switch element and a 2 2 matrix switch for proof-of-
bandwidth drops with M due to the cascaded filtering concept experiments. The switches are fabricated on an
bandwidth narrowing effect. However, we observe that SOI wafer with 0.34-m silicon device layer on a 1-m
even as M reaches seven, the obtained bandwidth remains buried oxide. Both the waveguide crossing and the
above 20 GHz for both 6- and 12-m-radius microring microring have a designed width of 0.4 m. The gap
resonators and for the ideal and fabrication imperfection spacing between the waveguide and the microring is
cases. We note that 20-GHz link bandwidth is still 0.35 m. We adopt a square-shaped microring resonator
sufficient for 5-Gbit/s NRZ data transmission, albeit with that offers a long interaction length of 12 m for the
tolerable waveform distortion. It is worth mentioning that evanescent coupling. The lateral p-i-n diode surrounds
the calculated bandwidth of the links with size nonunifor- almost the entire microring. The diode intrinsic region
mities can exceed that of the ideal link. We attribute such width is 1.4 m. The pþ -doped outer-ring region has a
bandwidth enhancement to the resonance wavelength boron concentration of 7 1019 cm3 . The nþ -doped
misalignment induced spectral lineshape broadening [21]. inner-ring region has a phosphorous concentration of
Therefore, the scalability is mainly optical-power limited. 1 1020 cm3 . We define the device layouts on a
Fig. 9. (a) Optical micrograph of our fabricated cross-grid switch element using a microring resonator side-coupled to a MMI waveguide
crossing. The gray area highlights the microring and waveguide structures. Metal-pad symbols: G: ground; S: signal. Inset: Scanning-electron
micrograph of MMI crossing. (b) Measured throughput- and drop-port transmission with driving voltages of 0 and 0.95 V at a resonance
around 1544.24 nm. The FWHM of the drop-port passband is 0.45 nm. (c) Measured switch transitions in the throughput-port transmissions
showing nanosecond-speed 10-to-90% switch-on and 90-to-10% switch-off times. (d) Measured throughput- and drop-port data transmissions
of a 5-Gbit/s PRBS data stream upon driving voltages of 0 (throughput) and 0.95 V (drop).

wafer by i-line (365 nm) photolithography and etch the for all of the four switch elements as above except that the
wafer by CF4 -based reactive ion plasma etching. The microring interaction length is increased to 15 m. Each
fabrication process is CMOS compatible and essentially microring is independently switchable by forward-biasing
follows that detailed in [62]. its integrated lateral p-i-n diode. For characterization
purposes using the microprobe, we design the four sets of
B. Switch Element Characterization aluminum pads to occupy a large chip area, which
Fig. 9(a) shows the top-view optical micrograph of the significantly enlarges the device footprint. A compact
fabricated switch element. The gray lines highlight the footprint can be realized by wire-bonding.
microring and the MMI crossing. The square-shaped Fig. 10(b) shows the schematic of the 2 2 matrix
microring corner radius is 10 m. The inset shows the switch. Like the 5 5 matrix switch, the light that is
scanning electron micrograph of the MMI crossing. We launched from port I1 propagates through to the through-
adopt the optimized MMI crossing design discussed in put-port T1 in the case that both microring resonators A
Section II-A [52]. The aluminum pads’ layout is configured and B are at electrical off-state. Switching microring
to spatially match with a ground-signal-ground radio- resonator A or B routes the light to drop-port D1A or D1B.
frequency microprobe for injecting carriers along essen- Likewise, the light that is launched from port I2 can
tially the entire microring, yet the detailed pad layout selectively exit from throughput-port T2 or drop-ports
design is not optimized for impedance matching. The D2C or D2D depending on the switch states of microring
experimental setup is detailed in our previous work [72]. resonators C and D.
Fig. 9(b) shows the measured TE-polarized (electric Fig. 10(c) and (d) shows the measured TE-polarized
field parallel to the chip) throughput- and drop-port transmission spectra with the light separately launched
transmission spectra in the vicinity of a resonance under from ports I1 and I2 upon zero bias. We measure four
0- and 0.95-V forward bias. The transmission is normalized different sets of microring resonances despite the micror-
to the light output from a lensed fiber firing into the input ing resonators’ being identical in design. The typical
bus waveguide. We remark that the relatively high measured Q is 4500 [full-width at half-maximum
insertion loss (20 dB) is dominated by the large (FWHM) 0:35 nm, bandwidth 44 GHz], which is
fiber-to-waveguide coupling loss, which is estimated to reasonably broadband for 5-Gbit/s data transmission. The
exceed 13 dB [72]. The measured resonance Q value of fabrication imperfection causes a resonance wavelength
3500 (linewidth of 0.45 nm, bandwidth of 56 GHz) is misalignment as large as 0.9 nm between microring
applicable for 5-Gbit/s data transmission with negligible resonators A and B, while the resonance wavelength
distortion. We obtain a resonance wavelength blue-shift of misalignment is 0.3 nm between microring resonators C
0.56 nm (exceeding the linewidth) upon 0.95-V forward- and D. We attribute this relatively large fabrication
bias. We measure a 610-A current across the diode, imperfection to our i-line photolithography, which is not
suggesting a dc power consumption of 600 W for our favorable for defining the 400-nm-width waveguides and
110-m-perimeter microring (5 W=m dc power gap spacing. We also note the adverse effects of the
consumption). fabrication imperfection as seen from the nonuniform and
We measure the switch transition times at the drop- relatively low drop-port transmissions (30 to 40 dB
port transmission at a probe wavelength of 1543.68 nm normalized transmissions) across the four drop-port
(the blue-shifted resonance wavelength) by supplying a transmissions. In order to obtain desirable signal transmis-
0.95-V square-wave pulse (5-ns width) across the p-i-n sions, advanced high-resolution nanoscale lithography
diode. Fig. 9(c) shows the measured 10%-to-90% switch- with high fidelity is of the essence.
on time and 90%-to-10% switch-off time are 1.3 and We demonstrate the 2 2 matrix switch data routing
1.2 ns, which are consistent with our semiconductor by launching 5-Gbit/s PRBS optical NRZ signals from both
device calculations using MEDICI. input-ports I1 and I2. We choose two different probe
Moreover, we measure the data transmission by wavelengths (1550.02 nm for I1 input, 1549.26 nm for
sending a 5-Gbit/s pseudorandom bit stream (PRBS) I2 input). Fig. 10(e) and (f) shows the measured data
optical signal (215 1) in NRZ data format into the switch transmissions at the six output-ports. For data input at
element and measuring at the probe wavelength the port I1, the data directly transmit to throughput-port T1
throughput-port signal transmission upon 0-V bias and the upon zero bias. With a 1.0 V (0.9 V) dc bias applied to
drop-port signal transmission upon 0.95-V forward bias. microring A (B) only, the data are routed to drop-port
Fig. 9(d) shows the transmitted data stream with good D1A (D1B), showing no significant waveform distortion
waveform quality at the throughput-port (switch off-state) compared with that transmitted to throughput-port T1.
and at the drop-port (switch on-state). However, the drop-port signal transmission intensities
are significantly lower than the throughput-port signal
C. 2 2 Matrix Switch Characterization transmission intensity, as expected from the spectra
Fig. 10(a) shows the top-view optical micrograph of the measurements [Fig. 9(c) and (d)]. We remark that the
fabricated 2 2 matrix switch. We adopt the same design pronounced noise fluctuations appearing in the drop-port

Fig. 10. (a) Optical micrograph of our fabricated 2 2 microring resonator-based matrix switch. (b) Schematic of the 2 2 matrix switch and its
light paths. (c) Measured throughput- and drop-port transmissions at port T1, D1A, and D1B when light is launched from port I1. (d) Measured
throughput- and drop-port transmissions at port T2, D2C, and D2D when light is launched from port I2. (e) and (f) Signal routing from the six
different output ports with 5-Gbit/s NRZ optical signal waveforms upon switching on the corresponding microring resonator-based switches
when light is launched from port I1 (c1 ¼ 1550:02 nm) and port I2 (c2 ¼ 1549:26 nm).
signals are due to the amplified spontaneous emission The resonance wavelength misalignments among the
noise from the two erbium-doped fiber amplifiers in our microring resonators impose different switching voltages,
experimental setup. Likewise, we demonstrate signal and thereby different dc power consumptions among the
transmissions between port I2 and ports T2/D2C/D2D by switch elements. From the measured current–voltage (I–V)
selectively switching microrings C (0.95 V)/D (0.9 V). curves, we estimate the dc power consumptions for the four

microring resonators A, B, C, and D are 2220 W, 300 W, Our initial devices, however, suffer high optical power loss,
960 W, and 290 W, respectively. Considering the resonance wavelength misalignments, and high power
122-m-perimeter microring structure, we see that consumption largely due to fabrication imperfections.
these correspond to, respectively, 18.1, 2.44, 7.8, and
2.36 W=m dc power consumption. These are, however, B. Design Guidelines
larger than our MEDICI-simulated values under such voltages Based on the study described in this paper, we
[see Fig. 6(b)]. We attribute the discrepancy to i) possible summarize some design guidelines of cascaded
imperfect contact and leakage current and ii) possible extra microresonator-based matrix switches targeting the gen-
carriers injected in order to compensate the thermal-induced eral performance metrics of bandwidth (capacity), power
resonance wavelength red-shift upon ohmic heating and consumption, footprint, and scalability for on-chip optical
carriers recombination. The higher voltages and power interconnection.
consumptions for resonators A and C are due to the higher i) The matrix switch bandwidth must be sufficiently
wavelength detuning to the probe wavelength. From our wide to support the optical signal bandwidth for
measurements, we project that a 6-m-radius microring the designed number of cascaded switches along a
resonator-based switch consumes around 100-W dc power. network route. We have shown that an optical
This suggests that our proposed 5 5 router with all five link in a 4 4 tile NoC comprising seven on-state
connections established (i.e., five on-state microring and 28 off-state switch elements (90-GHz
switches) consumes around 0.5 mW. microresonator bandwidth) can still transmit a
5-Gbit/s NRZ data stream with acceptable wave-
form distortion and route loss.
VI . SUMMARY AND OUTLOOK ii) The signal wavelength with a proper detuning
from the off-state resonance wavelength is needed
A. Key Results to optimize the on-state drop-port transmission
In summary, we reviewed in this paper the recent loss and the off-state throughput-port transmis-
development of a cascaded microresonator-based matrix sion loss. Moreover, the detuning also governs the
switch on a silicon chip for nascent optical interconnection carrier concentration change needed to turn-on
applications in many-core computing chips. Specifically, the switch element for routing, and thus deter-
we emphasized our earlier proposed structure of a 5 5 mines the on-state electrical power consumption.
matrix switch, comprising cascaded microring resonator- iii) The optical link loss within a matrix switch sets the
based carrier-injection-type switches side-coupled to a grid upper limit on the number of switches allowed in a
of MMI-based waveguide crossings. The MMI waveguide route. The link loss comprises not only the
crossings are critical building blocks in a cross-grid insertion loss of the switch element but also the
network, as they enable low-loss low-crosstalk broadband power penalty due to the fabrication imperfection
waveguide intersections that can be fabricated with induced passband variations. In light of this, it is
reasonably high tolerance. Our analysis assuming com- perhaps preferable to trade off the small footprint
pact-sized silicon microring resonator-based switches with for larger sized microresonators in order to enable
integrated lateral p-i-n diodes indicated that such matrix higher tolerance to fabrication imperfections and
switch is favorable for: better large-scale integration. The microresonator
i) optical interconnections among multiple input- size along with fabrication tolerance and the
ports and multiple output-ports in a compact tens implied optical link loss should be regarded as one
to hundreds of micrometers-scale footprint; critical optimization in the design.
ii) 5 Gbit/s-scale data transmission on a single Regarding the choice of the signal wavelength, we should
wavelength channel using a switch element with point out that we can also adopt the signal wavelength at the
filtering bandwidth of a few tens of gigahertz (a electrical off-state resonance instead of at the on-state
microresonator Q of a few thousands); resonance emphasized in this work. This alternative scheme
iii) nanosecond-speed circuit-switching upon a routes the signal wavelength to the drop-port (throughput-
CMOS-compatible driving voltage (1 V); port) when the switch element is at the electrical off-state
iv) few tens of microwatts scale dc power consump- (on-state). One potential merit of this scheme is that the off-
tion per switch element per link; state drop-port signal transmission is free from the FCA,
v) up to 7 7 tile chip for 6- and 12-m-radius while the on-state throughput-port signal transmission also
microring resonators (assuming no fabrication essentially stops the light from passing through the FCA-
imperfections). degraded microresonator. Thus, it offers potentially a better
Our proof-of-concept experiments demonstrated a insertion loss for routing. Another potential merit is its
single cross-grid switch element and a 2 2 matrix switch better tolerance to the driving voltage variation. Upon the
array, with 1 ns-speed switch on/off times and 5-Gbit/s electrical on-state, the signal wavelength in the throughput-
data transmission without significant waveform distortion. port transmission sees a relatively flat band outside the edge

of the inverted Lorentzian lineshape. A small variation in back integrated lateral p-i-n diodes [89]. The proposed
the driving voltage only slightly varies the throughput- device structure promises high switching speed and
port transmission. In contrast, with the present scheme, ultralow power consumption (fJ/bit). However, to our
the signal wavelength in the drop-port transmission sees the knowledge, this structure has not been experimentally
peak of the Lorentzian lineshape. A small variation in the demonstrated. We remark that such p-i-n-i-p diodes’
driving voltage can significantly attenuate the drop-port integrated microresonator structure can also be adopted
transmission. for low-power high-speed matrix switch applications.
Nonetheless, the shortcoming of this alternative The most technologically challenging question facing
scheme is that it imposes potentially larger switching the cascaded microresonator-based matrix switch is the
power consumption. For example, in order to establish an inevitable fabrication-induced nonuniformities. The re-
injection-to-Oeast connection, six switches need to be in sulting resonance wavelength misalignment can lead to
the electrical on-state for the throughput-port transmis- severe optical power penalty and also significantly increase
sions. The on-state switch number using this alternative the electrical power consumption, as shown in this paper.
scheme therefore far exceeds the present scheme which Although the fabrication technology has been maturing
only imposes one on-state switch for each optical route. into the nanometer resolution regime and promising
trimming methods have been proposed [10], [90], it is still
C. Future Challenges a formidable challenge to fabricate a large number of
Many engineering questions remain to be answered micrometer-scale resonator structures on a chip within
before we may use cascaded microresonator-based matrix nanometer accuracy. Substantial research effort on fabri-
switch in a silicon photonic network. The challenges span cation and trimming techniques are called for. Specifically,
from material to device to system levels, and we believe we need more advanced nanoscale lithography that
that many of them are relevant to silicon photonic NoC enables the large array of microresonators to be defined
research at large. We highlight and briefly discuss only a down to few-nanometer precision.2 Perhaps nanoimprint
few specific ones as follows. lithography [83] may answer the challenge.
• What is the optimal electrooptic switching mech- Finally, WDM technology may be one possible solution
anism or structure to realize low-power high-speed to enhance the aggregated data transmission rate via the
circuit switching? matrix switch, with the data multiplexed into multiple
• How to improve the scalability of the matrix switch wavelength channels. The microresonator-based switches
for high-performance many-core computation? acting as a comb switch/filter can pass multiple periodic
• Will WDM technology help? channels [40], [71], [91]. With each of the wavelength
Silicon electrooptic switches using either MOS capaci- channels’ carrying only a moderate data rate, the
tors [84]–[86] or reverse-biased p-n diodes [87], [88] are bandwidth for the resonator may remain relatively narrow
also applicable to matrix switches. The switching is then which is desirable to reduce the switch power consump-
based on electric-field induced majority carrier dynamics. tion. At the system level, however, the WDM approach
The response is electrically limited by the resistance– requires detailed tradeoff analysis of the additional
capacitance (RC) time delay that is ideally of the order of electrical power consumption and the accompanied
picosecond. However, for both MOS capacitor approach thermal loading (heating) imposed by possibly integrating
and depletion-based p-n diode approach, the carrier- more on-chip lasers, modulators, and receivers. h
induced refractive index changes are relatively small due
to the small amount of carriers that can be accumulated or
depleted. Therefore, relatively high driving voltages are Acknowledgment
required in order to obtain sufficient resonance wavelength The authors gratefully acknowledge many fruitful
shifts [84], [87], [88]. In order to minimize the resonance discussions with Prof. C. Y. Tsui and Prof. J. Xu from
wavelength shift imposed, it is likely that a high-Q HKUST. The nanofabrication on silicon chips by HKUST
resonator (Q 104 or above) is needed, which is, however, Nanoelectronics Fabrication Facility is gratefully acknowl-
unfavorable for high-data-rate signal transmission. edged. X. Luo would like to acknowledge the fellowship
Recently, Lipson’s group proposed a novel silicon support from the NANO program of HKUST. Assistance
microring resonator-based modulator using two back-to- from S. Feng is also gratefully acknowledged.
REFERENCES nanometre regime,[ Integration-VLSI J., [4] R. G. Beausoleil, M. Fiorentino, J. Ahn,

vol. 38, no. 1, pp. 3–17, Oct. 2004. N. Binkert, A. Davis, D. Fattal, N. P. Jouppi,
[1] L. Benini and G. De Micheli, BNetworks on M. McLaren, C. M. Santori, R. S. Schreiber,
chips: A new SoC paradigm,[ Computer, [3] U. G. Nawathe, M. Hassan, L. Warriner,
K. Yen, B. Upputuri, D. Greenhill, A. Kumar, S. M. Spillane, D. Vantrease, and Q. Xu,
vol. 35, no. 1, pp. 70–78, Jan. 2002. BA nanophotonic interconnect for
and H. Park, BAn 8-core 64-thread 64b
[2] D. Pamunuwa, J. Oberg, L. R. Zheng, power-efficient SPARC SoC,[ in Proc. 2007 high-performance many-core computation,’’
M. Millberg, A. Jantsch, and H. Tenhunen, IEEE Int. Solid-State Circuits Conf., Feb. 2007, in Proc. IEEE 5th Int. Conf. Group IV Photon.,
BA study on the implementation of pp. 108–110. Sep. 17–19, 2008.
2-D mesh-based networks-on-chip in the

[5] J. Appenzeller, BCarbon nanotubes for Express, vol. 15, no. 12, pp. 7489–7498, [37] N. Kirman, M. Kirman, R. K. Dokania,
high-performance electronicsVProgress and Jun. 2007. J. F. Martinez, A. B. Apsel, M. A. Watkins,
prospect,[ Proc. IEEE, vol. 96, pp. 201–211, [24] S. Xiao, M. H. Khan, H. Shen, and M. H. Qi, and D. H. Albonesi, BOn-chip optical
Feb. 2008. BA highly compact third-order silicon technology in future bus-based multicore
[6] M. C. F. Chang, V. P. Roychodhury, microring add-drop filter with a very large designs,[ IEEE Micro, vol. 27, pp. 56–66,
L. Y. Zhang, H. C. Shin, and Y. X. Qian, free spectral range, a flat passband and a low Jan./Feb. 2007.
BRF/wireless interconnect for inter- and delay dispersion,[ Opt. Express, vol. 15, no. 22, [38] T. Fukazawa, T. Hirano, F. Ohno, and
intra-chip communications,[ Proc. IEEE, pp. 14765–14771, Oct. 2007. T. Baba, BLow loss intersection of Si photonic
vol. 89, pp. 456–466, Apr. 2001. [25] Q. Xu, B. Schmidt, S. Pradhan, and M. Lipson, wire waveguides,[ Jpn. J. Appl. Phys., vol. 43,
[7] K. L. Wang, K. Galatsis, R. Ostroumov, BMicrometre-scale silicon electro-optic no. 2, pt. 1, pp. 646–647, Feb. 2004.
A. Khitun, Z. Zhao, and S. Han, modulator,[ Nature, vol. 435, no. 7040, [39] N. Sherwood-Droz, H. Wang, L. Chen,
BNanoarchitectonics for heterogeneous pp. 325–327, May 2005. B. G. Lee, A. Biberman, K. Bergman, and
integrated nanosystems,[ Proc. IEEE, [26] S. Manipatruni, Q. Xu, B. S. Schmidt, M. Lipson, BOptical 4 4 hitless silicon
vol. 96, pp. 212–229, Feb. 2008. J. Shakya, and M. Lipson, BHigh speed carrier router for opitcal networks-on-chip (NoC),[
[8] D. A. B. Miller, BOptical interconnects to injection 18 Gb/s silicon micro-ring Opt. Express, vol. 16, no. 20,
silicon,[ IEEE J. Sel. Topics Quantum Electron., electro-optic modulator,[ in Proc. Conf. Lasers pp. 15 915–15 922, Sep. 2008.
vol. 6, pp. 1312–1317, Nov.–Dec. 2000. Electro-Opt., Oct. 21–25, 2007. [40] B. A. Small, B. G. Lee, K. Bergman, Q. Xu, and
[9] D. A. B. Miller, BRationale and challenges for [27] A. Huang, C. Gunn, G. Li, Y. Liang, M. Lipson, BMultiple-wavelength integrated
optical interconnects to electronic chips,[ S. Mirsaidi, A. Narasimha, and T. Pinguet, BA photonic networks based on microring
Proc. IEEE, vol. 88, pp. 728–749, Jun. 2000. 10 Gb/s photonic modulator and WDM resonator devices,[ J. Opt. Netw., vol. 6, no. 2,
MUX/DEMUX integrated with electronics in pp. 112–120, Feb. 2007.
[10] R. G. Beausoleil, J. Ahn, N. Binkert, A. Davis,
D. Fattal, M. Fiorentino, N. P. Jouppi, 0.13 m SOI CMOS,[ in Proc. 2006 IEEE Int. [41] A. Shacham, K. Bergman, and L. P. Carloni,
M. McLaren, C. M. Santori, R. S. Schreiber, Solid-State Circuits Conf., San Francisco, CA, BOn the design of a photonic
S. M. Spillane, D. Vantrease, and Q. Xu, 2006. network-on-chip,[ in Proc. 1st Int. Symp.
BA nanophotonic interconnect for [28] Q. Xu, S. Sandhu, M. L. Povinelli, J. Shakya, Networks-on-Chip (NOCS’07), 2007.
high-performance many-core computation,[ S. H. Fan, and M. Lipson, BExperimental [42] Y. P. Li, J. Z. Yu, and S. W. Chen,
IEEE LEOS News., vol. 22, pp. 15–23, realization of an on-chip all-optical analogue BRearrangeable nonblocking SOI waveguide
Jun. 2008. to electromagnetically induced thermooptic 4 4 switch matrix with low
[11] R. G. Beausoleil, P. J. Kuekes, G. S. Snider, transparency,[ Phys. Rev. Lett., vol. 96, insertion loss and fast response,[ IEEE Photon.
S. Y. Wang, and R. S. Williams, BNanoelectronic no. 12, p. 123901, Mar. 2006. Technol. Lett., vol. 17, pp. 1641–1643,
and nanophotonic interconnect,[ Proc. IEEE, [29] H. S. Rong, S. B. Xu, Y. H. Kuo, V. Sih, Aug. 2005.
vol. 96, pp. 230–247, Feb. 2008. O. Cohen, O. Raday, and M. Paniccia, [43] R. A. Soref and B. E. Little, BProposed
[12] M. Haurylau, G. Q. Chen, H. Chen, BLow-threshold continuous-wave Raman N-wavelength M-fiber WDM crossconnect
J. D. Zhang, N. A. Nelson, D. H. Albonesi, silicon laser,[ Nature Photon., vol. 1, switch using active microring resonators,[
E. G. Friedman, and P. M. Fauchet, BOn-chip no. 4, pp. 232–237, Apr. 2007. IEEE Photon. Technol. Lett., vol. 10,
optical interconnect roadmap: Challenges and [30] H. Park, A. W. Fang, R. Jones, O. Cohen, pp. 1121–1123, Aug. 1998.
critical directions,[ IEEE J. Sel. Topics O. Raday, M. N. Sysak, M. J. Paniccia, and [44] B. E. Little, S. T. Chu, W. Pan, and
Quantum Electron., vol. 12, pp. 1699–1705, J. E. Bowers, BA hybrid AlGaInAs-silicon Y. Kokubun, BMicroring resonator arrays for
Nov.–Dec. 2006. evanescent waveguide photodetector,[ Opt. VLSI photonics,[ IEEE Photon. Technol. Lett.,
[13] M. Lipson, BOvercoming the limitations of Express, vol. 15, no. 10, pp. 6044–6052, vol. 12, pp. 323–325, Mar. 2000.
microelectronics using Si nanophotonics: May 2007. [45] S.-J. Chang, C.-Y. Ni, Z. Wang, and Y.-J. Chen,
Solving the coupling, modulation and [31] L. Schares, C. L. Schow, S. J. Koester, BA compact and low power consumption
switching challenges,[ Nanotechnology, G. Dehlinger, R. A. John, and F. E. Doany, optical switch based on microrings,[ IEEE
vol. 15, no. 10, pp. S622–S627, Oct. 2004. BA 17-Gb/s low-power optical receiver using a Photon. Technol. Lett., vol. 20, pp. 1021–1023,
[14] M. Lipson, BGuiding, modulating, and Ge-on-SOI photodiode with a 0.13-m CMOS Jun. 2008.
emitting light on siliconVChallenges and IC,[ in Proc. Opt. Fiber Commun. Conf. (OFC), [46] Y. Goebuchi, M. Hisada, T. Kato, and
opportunities,[ J. Lightw. Technol., vol. 23, 2006. Y. Kokubun, BOptical cross-connect circuit
pp. 4222–4238, Dec. 2005. [32] Y. Tao, R. Cohen, M. M. Morse, G. Sarid, using hitless wavelength selective switch,[
[15] B. Jalali and S. Fathpour, BSilicon photonics,[ Y. Chetrit, D. Rubin, and M. J. Paniccia, Opt. Express, vol. 16, no. 2, pp. 535–548,
J. Lightw. Technol., vol. 24, pp. 4600–4615, B40 Gb/s Ge-on-SOI waveguide photodetectors Jan. 2008.
Dec. 2006. by selective Ge growth,[ in Proc. Optical Fiber [47] Y. Goebuchi, T. Kato, and Y. Kokubun,
Commun./Nat. Fiber Opt. Eng. Conf., 2008. BMultiwavelength and multiport hitless
[16] J.-P. Colinge, Silicon-on-Insulator Technology:
Materials to VLSI. Norwell, MA: Kluwer, [33] J. F. Liu, J. Michel, W. Giziewicz, D. Pan, wavelength-selective switch using
2004. K. Wada, D. D. Cannon, S. Jongthammanurak, series-coupled microring resonators,[ IEEE
D. T. Danielson, L. C. Kimerling, J. Chen, Photon. Technol. Lett., vol. 9, pp. 671–673,
[17] L. Pavesi and D. J. Lockwood, Eds., Silicon
F. O. Ilday, F. X. Kartner, and J. Yasaitis, May–Jun. 2007.
Photonics. New York: Springer, 2004.
BHigh-performance, tensile-strained Ge p-i-n [48] Y. Hatakeyama, T. Hanai, S. Suzuki, and
[18] G. T. Reed and A. P. Knights, Silicon photodetectors on a Si platform,[ Appl. Phys. Y. Kokubun, BLoss-less multilevel crossing of
Photonics: An Introduction. New York: Lett., vol. 87, no. 10, p. 103501, Sep. 2005. busline waveguide in vertically coupled
Wiley, 2004.
[34] A. Shacham, K. Bergman, and L. P. Carloni, microring resonator filter,[ IEEE Photon.
[19] Proc. 1st–5th IEEE Int. Conf. Group IV BPhotonic networks-on-chip for future Technol. Lett., vol. 16, pp. 473–475, Feb. 2004.
Photonics, 2004–2008. generations of chip multiprocessors,[ IEEE [49] S. T. Chu, B. E. Little, W. G. Pan, T. Kaneko,
[20] R. Soref, BThe past, present, and future of Trans. Comput., vol. 57, pp. 1246–1260, and Y. Kokubun, BCascaded microring
silicon photonics,[ IEEE J. Sel. Topics Sep. 2008. resonators for crosstalk reduction and
Quantum Electron., vol. 12, pp. 1678–1687, [35] T. Barwicz, H. Byun, F. Gan, spectrum cleanup in add-drop filters,[ IEEE
Nov.–Dec. 2006. C. W. Holzwarth, M. A. Popovic, P. T. Rakich, Photon. Technol. Lett., vol. 11, pp. 1423–1425,
[21] F. Xia, L. Sekaric, and Y. Vlasov, M. R. Watts, E. P. Ippen, F. X. Kartner, Nov. 1999.
BUltracompact optical buffers on a silicon H. I. Smith, J. S. Orcutt, R. J. Ram, [50] S. M. Zheng, H. Chen, and A. W. Poon,
chip,[ Nature Photon., vol. 1, no. 1, pp. 65–71, V. Stojanovic, O. Olubuyide, J. L. Hoyt, BMicroring-resonator cross-connect filters-in
Jan. 2007. S. Spector, M. Geis, M. Grein, T. Lyszczarz, silicon nitride: Rib waveguide dimensions
[22] F. Xia, M. Rooks, L. Sekaric, and Y. Vlasov, and J. U. Yoon, BSilicon photonics for dependence,[ IEEE J. Sel. Topics Quantum
BUltra-compact high order ring resonator compact, energy-efficient interconnects Electron., vol. 12, pp. 1380–1387,
filters using submicron silicon photonic wires invited,[ J. Opt. Netw., vol. 6, no. 1, pp. 63–73, Nov.–Dec. 2006.
for on-chip optical interconnects,[ Opt. Jan. 2007.
[51] A. Kaźmierczak, M. Briere, E. Drouard,
Express, vol. 15, no. 19, pp. 11 934–11 941, [36] A. Kaźmierczak, E. Drouard, M. Brière, P. Bontoux, P. Rojo-Romeo, I. O. Connor,
Sep. 2007. P. Rojo-Romeo, X. Letartre, I. O. Connor, X. Letartre, F. Gaffiot, R. Orobtchouk, and
[23] S. Xiao, M. H. Khan, H. Shen, and M. Qi, F. Caffiot, and Z. Lisik, BOptimization of an T. Benyattou, BDesign, simulation, and
BMultiple-channel silicon micro-resonator integrated optical crossbar in SOI technology characterization of a passive optical add-drop
based filters for WDM applications,[ Opt. for optical networks on chip,[ J. Telecommun. filter in silicon-on-insulator technology,[
Inf. Technol., vol. 3, pp. 109–114, 2007.

IEEE Photon. Technol. Lett., vol. 17, [64] R. A. Soref and B. R. Bennett, BElectrooptical [78] C. A. Barrios and M. Lipson, BModeling and
pp. 1447–1449, Jul. 2005. effects in silicon,[ IEEE J. Quantum Electron., analysis of high-speed electro-optic
[52] F. Xu and A. W. Poon, BSilicon cross-connect vol. QE-23, pp. 123–129, Jan. 1987. modulation in high confinement silicon
filters using microring resonator coupled [65] Q. Xu, S. Manipatruni, B. Schmidt, waveguides using metal-oxide-semiconductor
multimode-interference-based waveguide J. Shakya, and M. Lipson, configuration,[ J. Appl. Phys., vol. 96, no. 11,
crossings,[ Opt. Express, vol. 16, no. 12, B12.5 Gbit/s carrier-injection-based silicon pp. 6008–6015, Dec. 2004.
pp. 8649–8657, Jun. 2008. micro-ring silicon modulators,[ Opt. Express, [79] Y. Yanagase, S. Suzuki, Y. Kokubun, and
[53] A. Shacham, B. G. Lee, A. Biberman, vol. 15, no. 2, pp. 430–436, Jan. 2007. S. T. Chu, BBox-like filter response and
K. Bergman, and L. P. Carloni, BPhotonic [66] C. Li, L. J. Zhou, and A. W. Poon, BSilicon expansion of FSR by a vertically triple coupled
NoC for DMA communications in chip microring carrier-injection-based microring resonator filter,[ J. Lightw. Technol.,
multiprocessors,[ in Proc. 15th Annu. IEEE modulators/switches with tunable extinction vol. 20, pp. 1525–1529, Aug. 2002.
Symp. High-Perform. Interconnects ratios and OR-logic switching by using [80] B. G. Lee, B. A. Small, K. Bergman, Q. Xu, and
(HOTI 2007), 2007, pp. 29–38. waveguide cross-coupling,[ Opt. Express, M. Lipson, BTransmission of high-data-rate
[54] H. Wang, M. Petracca, A. Biberman, vol. 15, no. 8, pp. 5069–5076, Apr. 2007. optical signals through a micrometer-scale
B. G. Lee, L. P. Carloni, and K. Bergman, [67] M. R. Watts, D. C. Trotter, and R. W. Young, silicon ring resonator,[ Opt. Lett., vol. 31
BNanophotonic optical interconnection BMaximally confined high-speed no. 18, pp. 2701–2703, Sep. 2006.
network architecture for on-chip and off-chip second-order silicon microdisk switches,[ in [81] J. P. Kuijten, F. Duray, and T. D. Kinderen,
communications,[ in Proc. Conf. Optical Fiber Proc. Optical Fiber Commun./Nat. Fiber Opt. BAnalysis of reticle contributions to CD
Commun./Nat. Fiber Opt. Eng. Conf., Eng. Conf. (OFC/NFOEC’08), uniformity for 0.25 m DUV lithography,[ in
Feb. 2008, pp. 1–3. Feb. 24–28, 2008. Proc. SPIE, 1998, vol. 3348, pp. 620–662.
[55] J. Takahashi, T. Tsuchizawa, T. Watanabe, and [68] L. Zhou, BSilicon microring and microdisk-based [82] M. S. Krishna, E. Gurer, T. X. Zhong,
S. Itabashi, BOxidation-induced improvement active devices using integrated p-i-n diodes,[ E. C. Lee, J. W. Salois, and R. M. Reynolds,
in the sidewall morphology and Hong Kong Univ. of Sci. and Technol., 2007. BIntegrated yield and CD enhancement for
cross-sectional profile of silicon wire [69] A. W. Poon, F. Xu, and X. Luo, BCascaded advanced DUV lithography,[ in Proc. SPIE,
waveguides,[ J. Vac. Sci. Technol. B, vol. 22, active silicon microresonator array 1999, vol. 3884, pp. 56–64.
no. 5, pp. 2522–2525, Sep.–Oct. 2004. cross-connect circuits for WDM [83] R. F. Pease and S. Y. Chou, BLithography and
[56] P. Dumon, W. Bogaerts, V. Wiaux, J. Wouters, networks-on-chip,[ in Proc. SPIE Photon. West other patterning techniques for future
S. Beckx, J. Van Campenhout, D. Taillaert, 2008, San Jose, CA, Jan. 19–24, 2008. electronics,[ Proc. IEEE, vol. 96, pp. 248–270,
B. Luyssaert, P. Bienstman, [70] B. G. Lee, B. A. Small, K. Bergman, Q. Xu, and Feb. 2008.
D. Van Thourhout, and R. Baets, BLow-loss M. Lipson, BTransmission of high-data-rate [84] L. Liao, D. Samara-Rubio, M. Morse,
SOI photonic wires and ring resonators optical signals through a micrometer-scale A. S. Liu, D. Hodge, D. Rubin, U. D. Keil, and
fabricated with deep UV lithography,[ IEEE silicon ring resonator,[ Opt. Lett., vol. 31, T. Franck, BHigh speed silicon Mach-Zehnder
Photon. Technol. Lett., vol. 16, pp. 1328–1330, no. 18, pp. 2701–2703, Sep. 2006. modulator,[ Opt. Express, vol. 13, no. 8,
May 2004. pp. 3129–3135, Apr. 2005.
[71] B. G. Lee, B. A. Small, Q. Xu, M. Lipson, and
[57] T. Tsuchizawa, K. Yamada, H. Fukuda, K. Bergman, BCharacterization of a 4 4 Gb/s [85] C. Li and A. W. Poon, BActive silicon
T. Watanabe, J. Takahashi, M. Takahashi, parallel electronic bus to WDM optical link microring resonators using
T. Shoji, E. Tamechika, S. Itabashi, and silicon photonic translator,[ IEEE Photon. metal-oxide-semiconductor capacitors,[ in
H. Morita, BMicrophotonics devices based on Technol. Lett., vol. 19, pp. 456–458, Proc. IEEE 1st Int. Conf. Group IV Photon.,
silicon microfabrication technology,[ IEEE J. Mar.–Apr. 2007. Hong Kong, China, 2004.
Sel. Topics Quantum Electron., vol. 11,
[72] L. Zhou, H. Chen, and A. W. Poon, BOn-chip [86] C. Li and A. W. Poon, BActive silicon
pp. 232–240, Jan.–Feb. 2005.
NRZ-to-PRZ format conversion using octagonal micropillar resonator modulators
[58] W. Bogaerts, P. Dumon, D. Van Thourhout, narrowband silicon microring resonator-based using metal-oxide-semiconductor capacitors,[
and R. Baets, BLow-loss, low-cross-talk notch filters,[ J. Lightw. Technol., vol. 26, in Proc. Conf. Lasers Electro-Opt., Baltimore,
crossings for silicon-on-insulator pp. 1950–1955, Jul. 2008. MA, 2005.
nanophotonic waveguides,[ Opt. Lett., vol. 32,
[73] A. Yariv, BUniversal relations for coupling of [87] A. S. Liu, L. Liao, D. Rubin, H. Nguyen,
no. 19, pp. 2801–2803, Oct. 2007.
optical power between microresonators and B. Ciftcioglu, Y. Chetrit, N. Izhaky, and
[59] H. L. Liu, H. Tam, P. K. A. Wai, and E. Pun, dielectric waveguides,[ Electron. Lett., vol. 36, M. Paniccia, BHigh-speed optical modulation
BLow-loss waveguide crossing using a no. 4, pp. 321–322, Feb. 2000. based on carrier depletion in a silicon
multimode interference structure,[ Opt. waveguide,[ Opt. Express, vol. 15, no. 2,
[74] S. J. Emelett and R. A. Soref, BSynthesis of
Commun., vol. 241, no. 1–3, pp. 99–104, pp. 660–668, Jan. 2007.
dual-microring-resonator cross-connect
Nov. 2004.
filters,[ Opt. Express, vol. 13, no. 12, [88] D. Marris-Morini, L. Vivien, J. M. Fedeli,
[60] S. G. Johnson, C. Manolatou, S. H. Fan, pp. 4439–4456, Jun. 2005. E. Cassan, P. Lyan, and S. Laval, BLow loss and
P. R. Villeneuve, J. D. Joannopoulos, and high speed silicon optical modulator based on
[75] S. R. Vangal, J. Howard, G. Ruhl, S. Dighe,
H. A. Haus, BElimination of cross talk in a lateral carrier depletion structure,[ Opt.
H. Wilson, J. Tschanz, D. Finan, A. Singh,
waveguide intersections,[ Opt. Lett., vol. 23, Express, vol. 16, no. 1, pp. 334–339, Jan. 2008.
T. Jacob, S. Jain, V. Erraguntla, C. Roberts,
no. 23, pp. 1855–1857, Dec. 1998.
Y. Hoskote, N. Borkar, and S. Borkar, [89] S. Manipatruni, Q. Xu, and M. Lipson, BPINIP
[61] H. Chen and A. W. Poon, BLow-loss BAn 80-tile sub-100-W TeraFLOPS processor based high-speed high-extinction ratio
multimode-interference-based crossings for in 65-nm CMOS,[ IEEE J. Solid-State Circuits, micron-size silicon electro-optic modulator,[
silicon wire waveguides,[ IEEE Photon. vol. 43, pp. 29–41, Jan. 2008. Opt. Express, vol. 15, no. 20, pp. 13035–13042,
Technol. Lett., vol. 18, pp. 2260–2262, Oct. 2007.
[76] A. Vorckel, M. Monster, W. Henschel,
Nov. 2006.
P. H. Bolivar, and H. Kurz, BAsymmetrically [90] J. Schrauwen, D. Van Thourhout, and
[62] L. Zhou and A. W. Poon, BSilicon coupled silicon-on-insulator microring R. Baets, BTrimming of silicon ring resonator
electro-optic modulators using p-i-n diodes resonators for compact add-drop by electron beam induced compaction and
embedded 10-micron-diameter microdisk multiplexers,[ IEEE Photon. Technol. Lett., strain,[ Opt. Express, vol. 16, no. 6,
resonators,[ Opt. Express, vol. 14, no. 15, vol. 15, pp. 921–923, Jul. 2003. pp. 3738–3743, Mar. 2008.
pp. 6851–6857, Jul. 2006.
[77] C. A. Barrios, V. R. Almeida, R. Panepucci, [91] B. G. Lee, A. Biberman, D. Po, M. Lipson, and
[63] C. Li, L. Zhou, S. Zheng, and A. W. Poon, and M. Lipson, BElectrooptic modulation of K. Bergman, BAll-optical comb switch for
BSilicon polygonal microdisk resonators,[ silicon-on-insulator submicrometer-size multiwavelength message routing in silicon
IEEE J. Sel. Topics Quantum Electron., vol. 12, waveguide devices,[ J. Lightw. Technol., photonic networks,[ IEEE Photon. Technol.
pp. 1438–1449, Nov.–Dec. 2006. vol. 21, pp. 2332–2339, Oct. 2003. Lett., vol. 20, pp. 767–769, May 2008.

ABOUT THE AUTHORS

Andrew W. Poon (Member, IEEE) received the Fang Xu received the B.E. degree in information
B.A. degree (Hons.) from the University of Chicago, engineering from Zhejiang University, Zhejiang,
Chicago, IL, in 1995 and the M.Phil. and Ph.D. China, in 2006 and the M.Phil. degree in electronic
degrees from Yale University, New Haven, CT, in and computer engineering from The Hong Kong
1998 and 2001, respectively, all in physics. University of Science and Technology, Hong Kong
In 2001, he joined the Department of Electronic SAR, China, in 2008.
and Computer Engineering, The Hong Kong Uni-
versity of Science and Technology, Hong Kong
SAR, China, as an Assistant Professor. He is
currently an Associate Professor. His research
group focuses on experiments and modeling of optical microresonators
and silicon photonics.
Xianshu Luo (Student Member, IEEE) received Hui Chen received the B.E. degree in telecommu-
the B.E. degree in microelectronics from Jilin nication engineering from Beijing University of
University, Changchun, Jilin, China, in 2003 and Posts and Telecommunications, Beijing, China, in
the M.Phil. degree in microelectronics and solid- 2000 and the M.Phil. degree in electronic and
state physics from the Institute of Semiconduc- computer engineering from The Hong Kong Uni-
tors, Chinese Academy of Sciences, Beijing, in versity of Science and Technology, Hong Kong
2006. He is currently pursuing the Ph.D. degree in SAR, China, in 2003, where he is currently pursuing
the Department of Electronic and Computer the Ph.D. degree in electronic and computer
Engineering, The Hong Kong University of Science engineering.
and Technology, Hong Kong SAR, China.

View publicationAuthorized
stats licensed use limited to: CAMBRIDGE UNIV. Downloaded on July 21, 2009 at 06:22 from IEEE Xplore. Restrictions apply.

Ex JUnj JGve NBe DFN

Uploaded by

Copyright:

Available Formats

Ex JUnj JGve NBe DFN

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

Ex JUnj JGve NBe DFN

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Cascaded Microresonator-Based Matrix Switch for Silicon On-Chip Optical

Article in Proceedings of the IEEE · August 2009

Xianshu Luo Hui Chen

SEE PROFILE SEE PROFILE

modern optics View project

silicon nitride View project

The user has requested enhancement of the downloaded file.

ABSTRACT | This paper reviews developments in cascaded I . INTRODUCTION

modulators [25]–[28], hybrid-integrated evanescent lasers

Vol. 97, No. 7, July 2009 | Proceedings of the IEEE 1217

1218 Proceedings of the IEEE | Vol. 97, No. 7, July 2009

Table 1 Microresonator Arrays for Multiple-Input Multiple-Output Optical Interconnects

Vol. 97, No. 7, July 2009 | Proceedings of the IEEE 1219

1220 Proceedings of the IEEE | Vol. 97, No. 7, July 2009

Table 2 Low-Loss High-Index-Contrast Wire Waveguide Crossings

Vol. 97, No. 7, July 2009 | Proceedings of the IEEE 1221

Lorentzian lineshape filter profile. Upon a change in the

1222 Proceedings of the IEEE | Vol. 97, No. 7, July 2009

Vol. 97, No. 7, July 2009 | Proceedings of the IEEE 1223

Fig. 4(b) illustrates one possible physical layout of the

1224 Proceedings of the IEEE | Vol. 97, No. 7, July 2009

Et t0 t20 t1 A expði Þ þ 20 t1 A expði Þ

Vol. 97, No. 7, July 2009 | Proceedings of the IEEE 1225

1226 Proceedings of the IEEE | Vol. 97, No. 7, July 2009

Fig. 6(b) shows the modeled power consumption

Vol. 97, No. 7, July 2009 | Proceedings of the IEEE 1227

port transmission loss is also significantly reduced.

Table 3 Comparison of the Switch Element Performances

1228 Proceedings of the IEEE | Vol. 97, No. 7, July 2009

Table 4 Path-Dependent Optical Loss Across a 20-Ring Matrix Switch

Vol. 97, No. 7, July 2009 | Proceedings of the IEEE 1229

1230 Proceedings of the IEEE | Vol. 97, No. 7, July 2009

[56], our proposed 5 5 matrix switch can support 3 3 V. PROOF-OF-CONCEPT EXPERIMENTS

Vol. 97, No. 7, July 2009 | Proceedings of the IEEE 1231

1232 Proceedings of the IEEE | Vol. 97, No. 7, July 2009

Vol. 97, No. 7, July 2009 | Proceedings of the IEEE 1233

1234 Proceedings of the IEEE | Vol. 97, No. 7, July 2009

REFERENCES nanometre regime,[ Integration-VLSI J., [4] R. G. Beausoleil, M. Fiorentino, J. Ahn,

Vol. 97, No. 7, July 2009 | Proceedings of the IEEE 1235

1236 Proceedings of the IEEE | Vol. 97, No. 7, July 2009

Vol. 97, No. 7, July 2009 | Proceedings of the IEEE 1237

ABOUT THE AUTHORS

1238 Proceedings of the IEEE | Vol. 97, No. 7, July 2009

You might also like

Et t0 t20 t1 A expði Þ þ 20 t1 A expði Þ