6, NOVEMBER/DECEMBER 2016
TABLE I
BEST-IN-CLASS PHOTODETECTORS
1
Fig. 2. A typical optoelectronic receiver consisting of a PD, TIA, MA(s) and Total capacitance includes PD, bonding pad, and ESD capacitances.
2
finally a clocked-SA. Reported BW in GHz.
datacenters and other short-to-medium reach (few meters to directly affects the RX sensitivity and impacts the link budget
kilometer) applications. The discussions evaluate the impact on and power consumption.
the overall link performance due to each element in the optical The PD capacitance (CPD ) is the parasitic capacitance at the
link - interconnects, optoelectronic transmitter (TX) and receiver output of the PD (CPD has been incorporated into the total ca-
(RX). Section II of the paper discusses the various elements of a pacitance CT , and not explicitly shown in Fig. 2). CPD can
typical optoelectronic RX, providing an in-depth noise analysis be reduced by making the active region thicker and smaller in
of a transimpedance amplifier (TIA) architecture. The tradeoffs diameter at the expense of increased carrier transit time and
in the design of an RX that set the limit for the RX sensitivity alignment tolerances [6]. In addition, any capacitance between
are analyzed. Section III briefly presents an overview of MRR- the PD output and the TIA input, such as pads, electrostatic
based TX and VCSEL-based TX, and lists the best-in-class discharge (ESD) device, packaging and routing, leads to in-
device performances reported to date. With an understanding creasing the effective PD capacitance and thus degrading the
of the limits of RX sensitivity and TX tradeoffs, and setting RX BW. PDs can either be integrated, i.e. implemented on the
the baseline with the current best-in-class devices, Section IV same CMOS chip as the TIA and the rest of the RX front-end, or
presents the link budget and energy efficiency calculations for discrete, i.e. fabricated on a separate chip and connected to the
MRR-based and VCSEL-based links for various lengths of the RX front-end either using wire bonding or flip-chip packaging.
interconnect. It highlights the benefits of each approach, expos- Monolithically integrated PDs are desirable since they do not
ing the challenges associated with them that limit the overall require additional packaging steps and offer lower parasitic ca-
system performance, and setting the ground for future research. pacitances due to the lack of pads, ESD, and package parasitics
Section V explores the opportunities of MRR-based links in between the PD and the TIA. Although a topic of active ongoing
realizing Tb/s aggregate throughput using wavelength-division research, integrated PDs typically suffer from lower responsiv-
multiplexing (WDM), and discusses the research challenges. ities [9], attributed to their fabrication on CMOS processes that
Finally, the paper is concluded in Section VI. are not optimized for optoelectronic devices. Integrated PDs on
CMOS SOI processes with Ge doping have better responsivities
II. OPTICAL LINK: RECEIVER [7], but introduction of Ge requires high-temperature process-
Fig. 2 shows an optoelectronic RX that converts the received ing and may degrade the transistor performance [2]. Table I lists
modulated optical signal into an electrical signal, amplifies it, some of the state-of-the-art PDs operating up to 25 Gb/s.
and prepares it to be processed by the RX digital core. A photo-
diode (PD) is used to convert the optical signal into an electrical B. TIA and RX Front-End
current. The PD is followed by a TIA and a main amplifier (MA)
to amplify the electrical signal and convert it from a single-ended A TIA is a gain stage used to convert the PD output current
current to a differential voltage. A clocked-sense-amplifier (SA) to a voltage signal that can be further processed by the RX. The
is used after the MA to re-time the signal and provide a rail- TIA is required to have sufficient gain based on the sensitivity
to-rail digital output which can either be processed on the same demands of the SA, appropriate –3 dB BW for the link data
chip, or buffered and sent to another chip. The SA itself could rate, and minimum noise contribution for achieving the BER
be integrated with the optoelectronic RX on the same chip along requirements. These characteristics are strongly coupled to one
with clock recovery, or implemented separately. A replica TIA is another and must be addressed carefully. Several architectures
often used for generating an automatic threshold control voltage have been recently published [16]–[18] to relax the gain-BW
(VATC ), which helps in the single-ended to differential conver- tradeoff in TIAs.
sion. Next, the PD and the TIA are discussed in detail. Fig. 2 shows an RX front-end consisting of an inverter-based
TIA followed by one or more differential MA gain stages. The
inverter-based TIA architecture has been chosen due to its sim-
A. Photodiode
plicity, good performance and suitability to low VDD opera-
The responsivity of a PD, defined as the ratio of the output tion. It consists of a CMOS inverter with a voltage gain, Ainv ,
electrical current to the input optical power, is essentially the given by (1), and a resistance Rf providing the negative feed-
gain of the first stage of the optoelectronic RX, which means it back to reduce the dependency between the gain and BW. CL
3700110 IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS, VOL. 22, NO. 6, NOVEMBER/DECEMBER 2016
includes the output capacitance of the first stage and the loading
effect from the second stage. CF is the parasitic capacitance
between the input and output nodes of the TIA’s forward path
gain stage, which is amplified by the miller effect. The mid-band
transimpedance gain of the inverter-based TIA and the –3 dB
BW at its input, BWi , are given by (2) and (3), respectively.
Ainv = ( gmn + gmp ) (rdsn || rdsp ) = Gm ro (1)
Rf Ainv
GainTIA = (2)
1 + Ainv
1
BWi = (3)
Rf
2π CT 1+A i n v
TABLE II
BEST-IN-CLASS IMS
1
Only includes the dynamic power consumed in the 50 Ω termination.
Tuning WPE λ (nm) ER(dB) Data Rate Output Power Bias Current
(Gb/s) (mW) (mA)
ER IL Driver Voltage Data Rate Method Range Power
(dB) (dB) (Vpp) (Gb/s) (nm) (mW) [2] 17% 850 7.3 18 3 9
[41] 11%1 850 5.1 15-25 0.731 4.2
[28] 11 0.9 4 5 Electrical 0.28 0.34
[29] 7 - 2 10 Thermal 1.6 1.25 1
Average power.
[13] 7-91 - 1.95 20 Thermal 1.5 7.1/nm
[30] 7 5 4.4 25 Thermal 0.8 -
1
are typically wideband, there is no need for tuning circuits to
7 for a WDM system, and 9 for a single wavelength system.
adjust the VCSEL’s output wavelength. However, the loss and
TABLE IV dispersion in MMFs have limited their applications to distances
BEST-IN-CLASS CW LASER < 100 m in warehouse-size datacenters. Another disadvantage
of using MM VCSELs is the lack of support for WDM. How-
Optical λ (nm) Linewidth WPE Peak Optical ever, as an alternative to WDM, arrays of MM VCSEL are put
Power (mW) Power (mW)
together to obtain higher data-rates [40].
[34] 10 ∼1550 10 MHz 1.3%1 20 Table V lists some of the best-in-class MM VCSELs reported,
[35] 9 1550 0.22 pm 12.2%2 10 comparing them in terms of WPE, data rate, and the output
[36] 6 ∼1556 0.46 nm 9.5% -
[37] 15 ∼1555 - 7.6%2 -
power. In Section IV, a 17% peak WPE is assumed for the
[38] 17.3 ∼1566 0.22 pm 7.8% 20 VCSEL for link budget analysis.
1
Laser’s efficiency is 16.7% but drops severely with stabilization.
2
Waveguide-coupled WPE.
IV. LINK BUDGETS AND SYSTEMS EVALUATION
This section presents link budget analyses for both the MRR-
efficiencies. To function properly, wavelength stabilization and VCSEL-based links in datacenter applications using the best
systems [31]–[33] are added to the laser to compensate for efficiency figures reported for both CW lasers and VCSELs.
wavelength drifts that are caused by various effects such as In an MRR-based system an external CW laser source con-
temperature variations and aging. verts electrical power into optical power, PTX , with a WPE,
When the power consumed by the wavelength stabilization W P ECW . The optical power is then coupled to the photonics
systems is included, the wall-plug efficiency (WPE) of com- chip, with coupling losses, PCW −CPL and PSM −CPL , to be mod-
mercially available lasers is typically in the range of 1% [34], ulated by an MRR with an IL, PM RR−IL . On the other hand, in
where WPE is the ratio of laser output power to its total elec- a VCSEL-based link the laser diode is directly modulated by a
trical power consumption. Since CW lasers are power-hungry CMOS driver, which is wirebonded or flip-chip connected to the
and low efficiency devices, their effect on the overall system VCSEL chip. The VCSEL converts electrical power into opti-
power efficiency is significant. Accordingly, it is clear that for cal power, PTX , with an efficiency, W P EVCSEL . Afterwards, in
MRR-based links to be practical and energy efficient, research both cases the modulated optical signal is coupled to an optical
should be directed towards enhancing the laser WPE. fiber, with a coupling loss, PSM −CPL /PM M −CPL , that carries it
Hybrid-integrated silicon-photonic based lasers employing to the RX side after introducing an attenuation, POF−att , and in
distributed-Bragg-reflector topologies can improve the WPE. the case of an MMF, a dispersion of POF−disp as well. Finally,
Together with a power-efficient stabilization technique, [35] the optical signal is coupled to the PD, with a coupling loss,
achieved a 12.2% waveguide-coupled WPE (ratio of laser power PSM −CPL /PM M −CPL , which converts it to an electrical signal
available in the silicon waveguide to its total electrical power with a specific responsivity. The link budget of an MRR-based
consumption). In Section IV, a 12.2% waveguide-coupled WPE link and a VCSEL-based link are calculated based on equations
is assumed for the CW laser for link budget analysis. Table IV (9) and (10), respectively.
lists some of the best-in-class CW lasers reported. PRX (dBm) = PTX − PM RR−IL − PCW −CPL
Fig. 6. MRR-based link efficiency degradation for CW Laser WPE of 1.3% Fig. 7. VCSEL-based link efficiency degradation for 10 m, 100 m, and 300 m
and 12% (no significant difference in performance between 10 m and 100 m MMF (a dispersion penalty of 0 dB for the 10 m and 100 m links, and 5 dB for
SMF). the 300 m link is assumed, respectively).
TABLE VI
MRR-BASED LINK CHARACTERISTICS
1
Laser to SMF (2 dB).
SMF to MRR, MRR to SMF, and SMF to PD (1.3 dB).
TABLE VII
VCSEL-BASED LINK CHARACTERISTICS
Fig. 8. Energy efficiency versus RX sensitivity for TX, RX and the overall
adding the ER penalty [42], the crosstalk (XT) and ISI penalties, VCSEL-based and MRR-based links (100 m), excluding any CDR or clocking.
and the relative intensity noise (RIN) penalty [43]. For a link
with 7 dB ER, PPen adds up to about 4.8 dB. Tables VI and VII
summarize the numbers used in the link budget analysis. Recent
reported numbers for PSM −CPL range from 0.5 dB [44], 0.7 dB poses, Fig. 7 also shows the efficiency degradation for a VCSEL
[45], 1.3 dB [46] to 2.8 dB [47]. In this work, a PSM-CPL of link on a 300 m long MMF assuming 5 dB of POF−disp [6].
1.3 dB is assumed [46]. The output of the CW laser is coupled With a 3 dB margin assumed for each link [2], [4], the overall
to an SMF with a coupling loss, PCW −CPL , of 2 dB [48]. efficiency of a VCSEL-based link with a 17% W P EVCSEL and
In MRR-based link, an SMF with 0.5 dB/Km attenuation 10 m MMF is calculated to be 1.7%. On the other hand, if it
introduces 0.005 dB and 0.05 dB loss for 10 m and 100 m is replaced by a 12% W P ECW MRR-based link, the overall
length, respectively. These losses are negligible, and therefore, efficiency is calculated to be 0.6%. The corresponding numbers
the curves in Fig. 6 are almost overlapping. However, for a for a 100 m long fiber are 1.6% and 0.6%, respectively (assuming
VCSEL-based link using an OM4 MMF with 3.5 dB/Km at- no MMF dispersion penalty).
tenuation, the optical fiber introduces attenuation of 0.035 dB Fig. 8 shows the RX, TX, and the overall energy efficiency
and 0.35 dB for 10 m and 100 m long fibers, respectively. Fur- versus the RX sensitivity for 25 Gb/s MRR-based and VCSEL-
thermore, for longer distances an MMF will cause significant based links with 100 m long channel and an inverter-based TIA
modal and chromatic dispersion to the transmitted signal, and RX as discussed in Section II. Beyond the optical devices and
the associated power penalty along with mode partition noise interconnects, the energy efficiency includes power consump-
can severely degrade the overall efficiency. For illustration pur- tion in the TX driver, MRR and CW laser tuning, and RX
3700110 IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS, VOL. 22, NO. 6, NOVEMBER/DECEMBER 2016
Fig. 9. Energy efficiency breakdown versus data rate for VCSEL-based and
MRR-based links (100 m), excluding any CDR or clocking.
[54] Y. Chen et al., “A 25Gb/s hybrid integrated silicon photonic transceiver Bryan Casper (M’96) is the Director of the PHY Research Lab and Senior
in 28nm CMOS and SOI,” in Proc. IEEE Int. Solid-State Circuits Conf., Principal Engineer with Intel Labs, based in Hillsboro, OR, USA. He currently
Feb. 2015, pp. 402–404. leads the organization responsible for all high-speed PHY research at Intel in
[55] D. Livshits et al., “High efficiency diode comb-laser for DWDM opti- the electrical, optical, and short-distance mmWave domains. In 1998, he joined
cal interconnects,” in Proc. IEEE Opt. Interconnects Conf., May 2014, the Performance Microprocessor Division of Intel Corporation and contributed
pp. 83–84. to the development of the Intel Pentium and Xeon processors. Since 2000, he
[56] C.-H. Chen et al., “A comb laser-driven DWDM silicon photonic trans- has been a Circuit Researcher, contributing to the development of I/O self-test
mitter with microring modulator for optical interconnect,” in Proc. Lasers technology, signaling analysis methods, high-speed I/O circuit architectures,
Electro-Opt. Conf., May 2015, pp. 1–2. and multiple I/O standards.
[57] H. Jayatilleka et al., “Automatic wavelength tuning of series-coupled
Vernier racetrack resonators on SOI,” presented at the Optical Fiber Com-
munication Conf., Anaheim, CA, USA, Mar. 2016.
Shahriar Mirabbasi (M’02) received the B.Sc. degree in electrical engineering
[58] H. Jayatilleka et al., “Crosstalk in SOI microring resonator-based filters,”
J. Lightw. Technol., vol. 34, no. 12, pp. 2886–2896, Jun. 2016. from the Sharif University of Technology, Tehran, Iran, in 1990, and the M.A.Sc.
and Ph.D. degrees in electrical and computer engineering from the University
[59] D. M. Kuchta et al., “A 55 Gb/s directly modulated 850nm VCSEL-based
of Toronto, Toronto, ON, Canada, in 1997 and 2002, respectively. Since 2002,
optical link,” in Proc. IEEE Photon. Conf., Sep. 2012, pp. 1–2.
he has been with the Department of Electrical and Computer Engineering,
[60] C. Xiong et al., “A monolithic 56 Gb/s CMOS integrated nanophotonic
PAM-4 transmitter,” in Proc. IEEE Opt. Interconnects Conf., Apr. 2015, University of British Columbia, Vancouver, BC, Canada, where he is currently
a Professor. His current research interests include analog, mixed-signal, RF, and
pp. 16–17.
mm-wave integrated circuit and system design with a particular emphasis on
[61] G. Denoyer et al., “Hybrid silicon photonic circuits and transceiver for
communication, sensor interface, and biomedical applications.
56 Gb/s NRZ 2.2 km transmission over single mode fiber,” in Proc. IEEE
Eur. Conf. Opt. Commun., Sep. 2014, pp. 1–3.
[62] A. Roshan-Zamir et al., “A 40 Gb/s PAM4 silicon microring resonator
Sudip Shekhar (S’00–M’10–SM’14) received the B.Tech. degree (Hons.) in
modulator transmitter in 65nm CMOS,” in Proc. IEEE Opt. Interconnects
Conf., May 2016, pp. 1–2. electrical and computer engineering from the Indian Institute of Technology,
Kharagpur, India, in 2003, and the M.S. and Ph.D. degree in electrical engineer-
ing from the University of Washington, Seattle, WA, USA, in 2005 and 2008,
respectively.
Abdelrahman H. Ahmed (S’14) received the B.Sc. degree in electronics and From 2008 to 2013, he was with Circuits Research Lab, Intel Corporation,
communication engineering (Hons.) from Alexandria University, Alexandria, Hillsboro, OR, USA, where he worked on high-speed I/O architectures. He is
Egypt, in 2012, and the M.Sc. degree in electronics engineering from the currently an Assistant Professor of electrical and computer engineering at the
American University in Cairo, Cairo, Egypt, in 2014. He is currently working University of British Columbia, Vancouver, BC, Canada. His research interests
toward the Ph.D. degree in electrical and computer engineering at the University include circuits for high-speed electrical and optical I/O interfaces, frequency
of British Columbia, Vancouver, BC, Canada. synthesizers, and wireless transceivers.
Between 2012 and 2014, he was a Research Assistant with the Center for Dr. Shekhar received the IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS
Nanoelectronics and Devices, Zewail City for Science and Technology, Cairo. Darlington Best Paper Award in 2010 and the IEEE Radio-Frequency IC Sym-
His research interests include circuits for high-speed electrical and optical I/O posium Best Student Paper Award in 2015.
interfaces.
Ahmad Sharkia (S’15) received the B.A.Sc. and the M.A.Sc. degrees in elec-
trical and computer engineering from the University of British Columbia,
Vancouver, BC, Canada, in 2013 and 2015, respectively, and is currently
working toward the Ph.D. degree as a Graduate Research Assistant.
He was an Engineering Intern at Qualcomm/Pixtronix in San Jose, CA, USA,
where he developed a stroboscope-based system for characterizing digital mi-
croshutter (DMS) displays in 2014. He was a Research Assistant at the UBC
Microsystems and Nanotechnology Group, where he developed and tested high
sensitivity capacitive readout circuits in 2012.