Ultra-dense monolithic integration of optical and electrical functions on silicon for optical interconnects

Abstract

CMOS Integrated Nanophotonics which allows dense monolithic integration of optical and electrical functions on the same chip can enable future Exaflops supercomputers by connecting racks, modules, and chips together with ultra-low power massively parallel optical interconnects.

Full text

Page 1

Ultra-dense Monolithic Integration of Optical and Electrical
Functions on Silicon for Optical Interconnects
Solomon Assefa, William M. J. Green, Alexander Rylyakov, Clint Schow, Folkert Horst* and Yurii A. Vlasov
IBM Thomas J. Watson Research Center, Yorktown Heights, NY 10598, USA; *IBM Zurich GMBH, Rueshlikon, Switzerland
Abstract: CMOS Integrated Nanophotonics which allows dense monolithic integration of optical and electrical
functions on the same chip can enable future Exaflops supercomputers by connecting racks, modules, and chips
together with ultra-low power massively parallel optical interconnects.

High-performance computing (HPC) systems capable of delivering Exaflops performance are envisioned to become a
reality by the end of this decade. In order to provide the enormous communication bandwidth that is necessary,
hundreds of millions of optical interconnects will have to be deployed to connect together racks, modules and chips. To
achieve such massive level of parallelism for HPC systems, monolithic integration of deeply scaled silicon optical
circuits into the front-end of standard CMOS process have been demonstrated as a promising technology [1-3].

The monolithic integration of electronic and nanophotonic components was performed at the IBM using 200 mm SOI
wafers (SOITEC) having a 220 nm silicon device layer on top of a 2 ?m BOX [1-4]. Several processing modules have
been added to a standard CMOS processing flow at the front-end of the line. These modules require a minimal number
of additional unique masks and processing steps, while sharing most mask levels and processing steps with the rest of
CMOS. For example passive waveguides and electro-optical and thermo-optical modulators share the same silicon
device layer with CMOS PFETS and NFETS. The Ge waveguide photodetectors were fabricated by utilizing a rapid
melt growth (RMG) technique wherein the Ge was melted and crystallized during the source-drain anneal step [5-6]. As
opposed to traditional approaches where Ge is typically grown by CVD after the source-drain anneal step, the RMG
approach enables the sharing of many CMOS steps and mask levels, thus minimizing cost, while yielding a very thin
defect-free Ge layer.

After completion of the monolithic integration including the metallization levels, the waveguides (200 nm x 500 nm
cross-section) exhibit propagation loss below 3dB/cm and bending losses less than 0.02dB per 90o-bend with 6 ?m
radius. Also, a cascaded Mach-Zehnder (CMZ) 8-channel WDM filter provided 8 flat-top 2 nm-wide pass-bands with
in-band ripple less than 1 dB and cross-talk less than -10 dB [7]. Furthermore, reverse-biased modulators were
demonstrated to have flat RF response with a 3 dB cutoff exceeding 20GHz [8], while the Germanium photodetectors
showed 3dB cutoff beyond 40 GHz at 1.5-2.0 V bias with some devices producing up to 10 dB avalanche gain at the
same low bias conditions [9].

The CMOS ring oscillator (RO) testsite is taken from a standard IBM digital library and consists of a 65-stage ring
oscillator, 10-stage ripple counter, and 4-stage divider (Fig. 1a)). After integration of all nanophotonics processing
modules the RO exhibits 12 ps delay per stage at saturation, under a 1.5 V bias. This performance is on-target for 130
nm bulk technology.

The receiver amplifier (Fig. 1b)) consists of a DC-coupled common-gate transimpedance amplifier (TIA) occupying
170 ?m x 40 ?m, followed by a cascade of seven current-mode logic (CML) buffer limiting amplifier (LA) stages,
followed by a single-ended open-drain output driver. The total area occupied by the multi-stage LA and output driver is
160 ?m x 50 ?m. The total circuit area is dominated by decoupling capacitors used to minimize the supply voltage
noise. After incorporation of all nanophotonic processing modules, the receiver amplifier exhibits an open eye diagram
at 5 Gbps with total power dissipation of about 28 mW (5.6 pJ/bit).

The modulator driver (Mod DRV) circuitry (Fig. 1c)) consists of the input pre-driver (105 ?m x 175 ?m) made of a 6-
stage differential CML amplifier, and an output stage (105 ?m x 70 ?m) consisting of a cascoded differential driver
with a dedicated power supply. The total circuit area is similarly dominated by the decoupling capacitors. High-speed
operation of the output stage is enabled by a differential pair of low-threshold thin oxide NFETs. High-voltage, high-
output swing capability of the output stage is enabled by a pair of long-channel thick oxide devices in a cascode
configuration, protecting the thin oxide devices from the output voltages. After incorporation of all the nanophotonic
processing modules, the transmitter amplifier exhibits an open eye diagram at 5 Gbps with total power dissipation of
about 36 mW (7.2 pJ/bit).

Figure 2 contains die photographs of a 6-channel transmitter (Fig. 2a)) and 6-channel receiver (Fig. 3b)). The chip
footprint of the transmitter in Fig. 3a) (counted as shown by the dotted outlines in Fig. 1b)) is 3700 ?m x 350 ?m. The
0.21 mm2 area per channel is limited by the pad frame and decoupling capacitors. The area of the 6-channel receiver

Page 2

(Fig. 2b) counted the same way is 3700 ?m x 500 ?m, which totals to an area of 0.31 mm2 per channel.
Correspondingly, the area per single transceiver channel can be estimated as 0.52 mm2, which is approximately 10
times smaller than previous demonstrations [10].


Fig. 1. Schematics, die photos, and performance measurements of CMOS digital circuits (ring oscillator in a)), CMOS analog circuits
(transimpedance amplifier in b), modulator driver amplifier in c)), and nanophotonic circuits (cascaded Mach-Zehnder WDM in d)),
taken after completion of the full integrated process flow.


Fig. 2. Micrographs of an integrated silicon nanophotonic a) 6-channel transmitter and b) 6-channel receiver.

To conclude, a proof-of-principle demonstration of dense monolithic integration of silicon nanophotonic components
with analog and digital CMOS circuitry has been described. The integration density offered by the technology is at least
10 times higher than previous reports. The technology offers a scalable solution for massively parallel Terabit/sec-class
optical transceivers, with 50 channels supporting serial lines rates of 20 Gbps each, occupying only 4 mm x 4 mm on a
single CMOS die.

Acknowledgements
The contributions of many of our colleagues across various organizations at IBM Research are acknowledged. We are
particularly grateful to Fengnian Xia, Leathen Shi, Jeffrey Sleight, Young-Hee Kim, Chris Jahnes, and the staff at the
Microelectronics Research Laboratory.

References
1. W. Green, S. Assefa, A. Rylyakov, C. Schow, F. Horst, Y. Vlasov, SEMICON 2010.
2. S. Assefa, W. Green, A. Rylyakov, C. Schow, F. Horst, Y. Vlasov, OMM6, OFC 2010.
3. J. Van Campenhout, et al., Optics Letters 35(7), 1013-1015 (2010).
4. J. Van Campenhout, et al., Optics Express 17(26), 24020-24029 (2009).
5. S. Assefa, et al., IEEE Journal of Selected Topics in Quantum Electronics, 16(5) 1376-1385 (2010).
6. S. Assefa, et al., Optics Express 18(5), 4986-4999 (2010).
7. C.G.H. Roeloffzen et al., IEEE Photonics Technology Letters 12, 1201-1203 (2000).
8. W. M. J. Green, et al., Optics Express, 15(25), 17106-17113, (2007).
9. S. Assefa, et. al, Nature 464, 80 (2010).
10. B. Analui et al., IEEE Journal of Solid-State Circuits 41(12), 2945-2955 (2006).