optical interconnects for high performance computing · optical interconnects for high performance...
TRANSCRIPT
1
OFC 2011Dec 10, 2010
© 2010 IBM Corporation
Optical Interconnects for High Performance Computing
Marc A. TaubenblattIBM T.J. Watson Research CenterYorktown Heights, NY 10598
M. Taubenblatt OFC 2011Dec 10, 2010 2© 2010 IBM Corporation
Acknowledgements
IBM Colleagues:
– C. Schow, J. Kash, P. Pepeljugoski, D. Kuchta, L. Schares, C. Baks, M. Ritter, L. Shan, K. Gu, D. Kam, Y. Kwark, R. Budd, F. Libsch, C. Tsang, J. Knickerbocker, P. Coteus, A. Gara, Y. Vlasov, S. Assefa, W. Green, B. Offrein, R. Dangel, F. Horst, Y. Taira, Y. Katayama, B. Lee, J. Van Campenhout*, A. V. Rylyakov, M. Yang, J. Rosenberg, S. Nakagawa, A. Benner, D. Stigliani, C. DeCusatis, H. Bagheri, K. Akasofu and many others at IBM.
Government Support:
– IBM’s Terabus , TELL, PERCS and Silicon Photonics programs are partially supported by DARPA.
– Osmosis: This research is supported in part by the University of California under subcontract number B527064.
*now with Interuniversity Micro-Electronics Center (IMEC)
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
©Optical Society of America
2
M. Taubenblatt OFC 2011Dec 10, 2010 3© 2010 IBM Corporation
Outline Why we need Optical Interconnects for High Performance
Computing
Optical Link Basics
Requirements and Trade-offs
– Cost, Power, Density
Technologies of Interest
– VCSELs/fiber
– VCSELS/Optical PCB
– Si Photonics
Architectures: Optical Switching
Conclusions
M. Taubenblatt OFC 2011Dec 10, 2010 4© 2010 IBM Corporation
Evolution of Optical interconnects
WAN, MANmetro,long-haul
LANcampus, enterprise
Systemintra/inter-rack
Boardmodule-module
Modulechip-chip
Chipon-chip buses
1980’s 1990’s 2000’s
Time of Commercial Deployment (Copper Displacement):
Distance Multi-km 100’s m 10’s m < 1 m < 10 cm < 20 mm
> 2012
Increasing integration of Optics with decreasing cost, decreasing power, increasing density
Telecom
Datacom
Computer-com
cards Card edgeCard edge/on card
Module Si C or chipIntegration
BW * Distance: Optics >> Copper
On chip
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
3
M. Taubenblatt OFC 2011Dec 10, 2010 5© 2010 IBM Corporation
Image courtesy of the National Center for Computational Sciences, Oak Ridge National Laboratory
Why High Performance Computing?
• Materials Science• Geophysical Data Processing• Environment and Climate Modeling• Life Sciences / Drug Discovery• Fluid Dynamics / Energy• Industrial Modeling• Financial Modeling• Transportation
Larger scale, more complex, higher resolution, multiscale PhysicsShorter time to solution real time response
More than 50% of Top500 systems are for industry*Growing number of flops are industry (~15% in 1990s to ~30% today)*
*www.top500.orgCourtesy L. Treinish, IBM
M. Taubenblatt OFC 2011Dec 10, 2010 6© 2010 IBM Corporation
Chip ~50% 30% CAGR*
Maintaining the HPC Performance Trend
1.E-01
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
1.E+05
1.E+06
1.E+07
1.E+08
1.E+09
1.E+10
1.E+11
No
v-9
2
No
v-9
4
No
v-9
6
No
v-9
8
No
v-0
0
No
v-0
2
No
v-0
4
No
v-0
6
No
v-0
8
No
v-1
0
No
v-1
2
No
v-1
4
No
v-1
6
No
v-1
8
No
v-2
0
Time
Pe
rfo
rma
nc
e (
Gig
afl
op
s)
Semiconductors & Pkg:~15-20% CAGR, slowing
Systems:85-90% CAGR,continuing**
Accelerators
Increasing Parallelism: System BW at all levels of assembly must scale exponentially
– ~(0.1-1 Byte/Flop) AND/OR new architectures, topologies and algorithms required
* http://www.bigncomputing.org**chart data from www.top500.org
1st500th
Total
10x /3.5-4yrs = 85-90% CAGR
Increasing Parallelism1PF
1EF
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
4
M. Taubenblatt OFC 2011Dec 10, 2010 7© 2010 IBM Corporation
Interconnects in HPC Systems
Core to core (on chip)
CPU to Memory (on card, backplane)
CPU to CPU (on 1st level package)(e.g. SMP)
Storage (rack to rack)
Cluster Fabric(Card, backplane, rack to rack)
LAN/WAN (building and beyond)A. BENNER ET AL., IBM J. RES. & DEV. VOL. 49 NO. 4/5 2005
M. Taubenblatt OFC 2011Dec 10, 2010 8© 2010 IBM Corporation
System BW must scale w/ Performance BW bottlenecks
Increasing chip performance Off chip BW bottleneck
Increasing module performance Off module bottleneck
Increasing card performance On and Off card bottleneck
Increasing rack performance Off rack bottleneck
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
5
M. Taubenblatt OFC 2011Dec 10, 2010 9© 2010 IBM Corporation
Why Optics?
LGA Socket
• Electrical Buses become increasingly difficult at high data rates (physics):• Increasing losses & cross-talk
• Frequency resonant affects
• Optical data transmission is easier: • Much lower loss, esp. at higher data rates
• Additional advantages include:
• Cable bulk, connector size, EMI…
• Potential power savings
• KEY ADVANTAGE: BW * Distance > electrical
• Optics trends for large servers and data centers are following same trends as telecom network:• Longer links first
• Short link optics requires tighter package integration:
• Higher BW closer to signal source
• Changes the supplier/manufacturing paradigm
Card
Card
CPU
Bac
kpla
ne
Cross talk Freq dependent losses
Resonance effects
Copper
Optics
M. Taubenblatt OFC 2011Dec 10, 2010 10© 2010 IBM Corporation
Rack to Rack: Already using optics
IBM Federation Switch for ASCI Purple (LLNL)- Copper for short-distance links (≤10 m)- Optical for longer links (20-40m)
~3000 parallel links 12+12@2Gb/s/channel
Combination of Electrical & Optical Cabling
20052002
NEC Earth Simulator• no optics
• 5Gb/s/ch transceivers• 55 miles of Active
Optical Cables
2008/9: 1PF/sIBM Roadrunner (LANL)
Image courtesy of the National Center for Computational Sciences, Oak Ridge National Laboratory
Cray Jaguar(ORNL)
• Was #2 in June, 2009
• Infiniband • 3 miles of
Optical Cables, longest = 60m
Courtesy of ESC www.jamstec.go.jp
Fiber to the RackFiber to the Rack40,000 optical links40,000 optical links
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
6
M. Taubenblatt OFC 2011Dec 10, 2010 11© 2010 IBM Corporation11
Bandwidth: the Bane of the Multicore Paradigm:
Logic flops continue to scale faster than interconnect BW
• Constant Byte/Flop ratio with N cores (constant ) means:Bandwidth(N-core) = N x Bandwidth(single core)
• 3Di (3D integration) will only exacerbate bottlenecks
Assumptions:• 3 GHz clock• ~ 3 IPC• 10 Gb/s I/O
• 1 B/Flop mem• 0.1 B/Flop data• 0.05 B/Flop I/O
Pins per chip
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
1 2 4 8 16 32 64 128
Number of Cores
Sig
nal
+ R
efer
ence
Pin
sS
igna
l + R
efer
ence
Pin
s
M. Ritter Topical Workshop on Electronics for Particle Physics, Sept 2010
M. Taubenblatt OFC 2011Dec 10, 2010 12© 2010 IBM Corporation12
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
1 2 4 8 16 32 64 128
Number of Cores
Sig
nal
+ R
efer
ence
Pin
s
Implications of BW Scaling:
Module Escape Bottleneck
Card Escape Bottleneck
Chip escape limit, 200m pitch
Module escape, 1mm pitch
Card escape, 8 pair/mm(QCM w/8 Cores…)
Pins per chip
Sig
nal +
Ref
eren
ce P
ins
M. Ritter Topical Workshop on Electronics for Particle Physics, Sept 2010
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
7
M. Taubenblatt OFC 2011Dec 10, 2010 13© 2010 IBM Corporation
High aggregate BW with electrical getting more difficult
Modu
le
Driver
Card-to-cardModule-to-module
Connector
Losses, Crosstalk
Through Vias (Stubs)
Reflections,Crosstalk
Reflections Reflections
Reflections
Losses
Crosstalk
Modu
le
Driver
Backplane
Chip-to-chip
ConnectorConnector
Mod
ule
Re
cvr
Recvr
Recvr
Module
0Hz 10GHz2.0GHz 4.0GHz 6.0GHz 8.0GHz-70
30
-60
-50
-40
-30
-20
-10
0
10
20
30
Frequency Response of 30” line
Frequency
|SD
D21
|
Via stub
No via stub
[M. Ritter, et. al., “The viability of 25 Gb/s on-board signaling”, ECTC 2008]
• Lower losses w/ higher xsection lines?
► BW density drops, cannot route
► Improved Boards (no vias, adv’d mat’ls)?
► Increased cost
• Signal Processing, higher swing?
► Power balloons
• Multi-level signaling?
► NRZ still works better
M. Taubenblatt OFC 2011Dec 10, 2010 14© 2010 IBM Corporation
1000
10000
100000
1000000
10000000
100000000
2004 2006 2008 2010 2012 2014 2016 2018 2020
Year
Nu
mb
er o
f O
pti
cal
Ch
ann
els
HPC driving volume optics Computercom market
MareNostrum
ASCI Purple
Blue Waters*
WW volume in 2008
Single machine volumes similar to today’s WW parallel optics
Roadrunner2.5Gbps
10Gbps
5Gbps
?
* Expected Summer 2011 T. Dunning, NCSA, https://hub.vscse.org/resources/86/download/VSCSE_FutureofHPC_Jul10-2.ppt#265,1,Future of High Performance Computing
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
8
M. Taubenblatt OFC 2011Dec 10, 2010 15© 2010 IBM Corporation
Network topology determines number of links and link lengths required
B. Arimilli et al, 2010 18th IEEE Symposium on High Performance Interconnects
PERCS – All to All w distributed switch - O(n2)
BG/L - 3D Torus - O(n)
A. BENNER ET AL., IBM J. RES. & DEV. VOL. 49 NO. 4/5 2005
ASCI Purple – Centralized Switch (Fat Tree)
And many others e.g. Clos,Hypercube, flattened butterfly, dragonfly…
TRADEOFFS:• Number of links (cost)• Bisection BW (performance)• Hops/Latency (performance)• Link length (cost)• Switch size (cost)
Distribution of Active Cable Lengths in Roadrunner
05
1015202530354045
0 10 20 30 40 50 60 70 80 90 100 110
Length (m)
Per
cen
tag
e o
f L
inks
(%
)
85% of the links are < 20m98% of the links are < 50m
Optics, if cheap enough, enables more connectivity, more BW, more spread out systems
M. Taubenblatt OFC 2011Dec 10, 2010 16© 2010 IBM Corporation
Cost and power of a supercomputer
YearPeak
PerformanceMachine Cost
Total Power Consumption
2008 1PF $150M 2.5MW
2012 10PF $225M 5MW
2016 100PF $340M 10MW
20201000PF(1EF)
$500M 20MW
Assumptions: Based on typical industry trends –(See, e.g., top500.org and green500.org)
– 10X performance / 4yrs (from top500 chart)– 10X performance costs 1.5X more– 10X performance consumes 2X more power
J. Kash Photonics Society Annual Meeting Nov 2010
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
9
M. Taubenblatt OFC 2011Dec 10, 2010 17© 2010 IBM Corporation
Total bandwidth, cost and power for optics in a machine
YearPeak
Performance(Bidi) Optical Bandwidth
Optics Power Consumption
Optics Cost
2008 1PF0.012PB/s
(1.2x105Gb/s)0.012MW $2.4M
2012 10PF1PB/s
(107Gb/s)0.5MW $22M
2016 100PF20PB/sec
(2x108Gb/s)2MW $68M
20201000PF(1EF)
400PB/sec(4x109Gb/s)
8MW $200M
Require >0.2Byte/FLOP I/O bandwidth, >0.2Byte/FLOP memory bandwidth– 2008 optics replaces electrical cables (0.012Byte/FLOP, 40mW/Gb/s)– 2012 optics replaces electrical backplane (0.1Byte/FLOP, 10% of power/cost)– 2016 optics replaces electrical PCB (0.2Byte/FLOP, 20% of power/cost)– 2020 optics on-chip (or to memory) (0.4Byte/FLOP, 40% of power/cost)
J. Kash Photonics Society Annual Meeting Nov 2010
M. Taubenblatt OFC 2011Dec 10, 2010 18© 2010 IBM Corporation
Cost and Power per optically-transmitted bit
YearPeak
Performance
Number of unidirectional
optical channels
Optics Power Consumption(Unidirectional)
Optics Cost (Unidirectional)
2008 1PF48,000
(@5Gb/s)50mW/Gb/s
(50pJ/bit)$10/Gb/s
2012 10PF2x106
(@10Gb/s)25mW/Gb/s $1.1/Gb/s
2016 100PF4x107
(@10Gb/s)5mW/Gb/s $0.17/Gb/s
20201000PF(1EF)
8x108
(@10Gb/s)1mW/Gb/s $0.025/Gb/s
To get optics to millions of units in 2012, need ~$1/Gb/s unidirectional– Cost targets continue to decrease with time below that
Power is OK for 2012, then sharp reductions will be needed– < 5mW/Gb/s likely difficult for with VCSEL-based links J. Kash Photonics Society
Annual Meeting Nov 2010
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
10
M. Taubenblatt OFC 2011Dec 10, 2010 19© 2010 IBM Corporation
Outline Why we need Optical Interconnects for High Performance
Computing
Optical Link Basics
Requirements and Trade-offs
– Cost, Power, Density
Technologies of Interest
– VCSELs/fiber
– VCSELS/Optical PCB
– Si Photonics
Architectures: Optical Switching
Conclusions
M. Taubenblatt OFC 2011Dec 10, 2010 20© 2010 IBM Corporation
What does an optics link consist of?Logical view:
Serializer,Coding, & Clock Laser
Driver
Deserializer,Decoding &
CDR
TIA& LA
Vb2
Optical fiber and/or waveguides, optical connectors…
LVb1
III-VVCSEL
PD
OE Module
OE Module
Pre-amp
Output Driver
CPU or Switch chip
CPU, Switch chip
Physical view:
On MCM optics
On Card optics
e.g. CPU
OE Module
PCB wiring & connector
Fiber & connector
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
11
M. Taubenblatt OFC 2011Dec 10, 2010 21© 2010 IBM Corporation
Signal Integrity: Eye Diagrams
Key Parameters:
– Extinction Ratio, ER = P1/P0• Want ER as large as possible
maximize modulated power
– Optical Modulation Amplitude, • OMA = P1 – P0• OMA = 2Pavg(ER-1)/(ER+1)
5 Gb/s, Excellent Eye
15 Gb/s, Marginal EyeEye Diagrams = snapshot of performance:
amplitude, speed, noise, jitter, distortion…
tftr
P1
P0
0 = no light Courtesy Clint Schow, IBM
jitter
Amplitude noise
ISI: Inter-symbol interference, eye depends on bit history
RIN: Random intensity noise
M. Taubenblatt OFC 2011Dec 10, 2010 22© 2010 IBM Corporation
Signal Impairment in Multimode Fiber (MMF)
A multimode fiber (or waveguide) can be represented as a waveguide propagating each of the modes in a separate “tube”
Which modes will be excited and what will be the power distribution between them depends on the launch conditions
The impulse response of the MMF is a weighted superposition of all modal pulses from each “tube”
Net: MM optics distorts the signal, leading to BW*distance limitations
Input pulse
Launching conditions
Output pulse
Diameters of core and cladding
Cladded parabola refractive index profile
Multimode optical fiber
Beam NA
Courtesy Petar Pepeljugoski, IBM
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
12
M. Taubenblatt OFC 2011Dec 10, 2010 23© 2010 IBM Corporation
P. Pepeljugoski, OFC 2005500470050um2009OM4
2002
<1995
<1995
Date of introduction
500200050umOM3
50050050umOM2
50020062.5umOM1
BW*D@ 1300nm (MHz*K)
BW*D @ 850nm (MHz*K)
Core Diam
FiberOM1
OM2
Single mode10km40Gbps2011802.3bg
~80-100m
150m
OM4
40km
SMF
Single mode10x10Gbps2010802.3ba
1300nm w equalization
220m220m220m102006802.aq
850nm100m10x10Gbps2010802.3ba
64B/66B4x25Gbps>2013100GbE
2002
1997
Year (Std)
8B/10B 850nm
550m500m1802.3z
64B/66B 850nm
300m82m33m10802.3ae
CommentOM3OM2OM1DR GbpsEthernet
Data courtesy of P. Pepeljugoski & A. Benner, IBM
Multimode Fiber & Standards
Fiber Channel: 4.25G (2005), 8.5G (2008), 14.025G (2011) 28.05G (2012) Infiniband DDR-5G (2006), QDR-10G (2006, product ~2009), FDR-14G (mid-2011), EDR-20G (mid-2011)
M. Taubenblatt OFC 2011Dec 10, 2010 24© 2010 IBM Corporation
Receiver Sensitivity Measurement
Measure bit error ratio (BER) as a function of input optical power, eye sampled at center
Receiver Sensitivity = minimum optical power required to achieve a specified BER (often 10-12)
Degrades at higher data rates due to bandwidth limitations
-18 -17 -16 -15 -14 -13 -12 -11 -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0
-5
-6
-7
-8
-9
-10
-11
-12
OMA (dBm)
log1
0[B
ER
]
R1AR1BR1CR1DR2AR2BR2CR2DR3AR3BR3CR3DR4AR4BR4CR4D
5 Gb/s
10 Gb/s
12.5 Gb/s
15 Gb/s
Example:985-nm Terabustransceivers
Optical Power
Bit
Err
or
Rat
io
Courtesy Clint Schow, IBM
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
13
M. Taubenblatt OFC 2011Dec 10, 2010 25© 2010 IBM Corporation
Example Link Power Budget (optical waveguides)
0.80.05Margin
-9.8
8.25
3.5
2
<15 Gb/s
-9.8RX sensitivity
8.25Signal Losses
(0.3m att. + CE)
2.75Power penalties
(including Xtalk)
2 dBmTX power OMA
10 Gb/s
Xtalk penalty=2 dBCE=2dB TX, 2dB RXISI penalty 1 dB
Courtesy Petar Pepeljugoski, IBM
M. Taubenblatt OFC 2011Dec 10, 2010 26© 2010 IBM Corporation
Outline Why we need Optical Interconnects for High Performance
Computing
Optical Link Basics
Requirements and Trade-offs
– Cost, Power, Density
Technologies of Interest
– VCSELs/fiber
– VCSELS/Optical PCB
– Si Photonics
Architectures: Optical Switching
Conclusions
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
14
M. Taubenblatt OFC 2011Dec 10, 2010 27© 2010 IBM Corporation
Optimized solutions will require detailed analysis of trade-offs
Power
CostDensity
Margin, Packaging Integration, Data-rate…
Packaging Integration, Channel Integration, Margin, Cooling, Data-rate…
Base Manufacturing Cost, Yield, Channel Integration, Data-rate, Reliability…
M. Taubenblatt OFC 2011Dec 10, 2010 28© 2010 IBM Corporation
Density of optical links
Courtesy of Avago Technologies
Optics Module and electrical connectorCard edge
~8X8mm, ~0.75mm pitch
~18x41mm 1.27mm pitch
Active cable, electrical at card edge
Courtesy of AvagoTechnologies
4x4 VCSELArray
4x4 PD
Array
4x4 VCSELArray
4x4 PD
Array
Optical at card edge
~60K fibers/ RACK
ROADRUNNER ~1PF)
~40K fibers / SYSTEM
PERCS >10PF
8@5Gbps/cable
48@10Gbps/cable
~5x3mm 0.2mm pitch
~10x25mm
~5X15mm
> 40x denser
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
15
M. Taubenblatt OFC 2011Dec 10, 2010 29© 2010 IBM Corporation
Packaging of optical interconnects is critical
Optical bulkheadconnector
Ceramic Organic card
Opto module
1cm FlexNIC
Laser+driver ICfiber1.7cm traces
Optical bulkheadconnector
Ceramic Organic card
Opto module
1cm FlexNICNIC
Laser+driver ICfiber1.7cm traces
Optics on-MCM
Good: Optics on-card
• Better to put optics close to logic rather than at the card edgeAvoids power, distortion & cost of electrical link on each end of optical linkBreaks through pin-count limitation of multi-chip modules (MCMs)
Operation at 10 Gb/s:equalization required
Ceramic Organic card
Opto module
Up to ~1m, 50
NIC Laser+driver IC
fiber1cm Flex
1.7cm traces
Ceramic Organic card
Opto moduleNICNIC Laser+driver IC
fiber1cm Flex
~2cm 50 traces
Optics on-card
Operation to >15 Gb/s:no equalization required ~2cm 50 traces
Bandwidth limited by # of pins
Implemented, P7-IH
Improved electrical power and off module BW
No internal jumper/bulkhead connectorSimpler standardized packaging
M. Taubenblatt OFC 2011Dec 10, 2010 30© 2010 IBM Corporation
Integrated packaging is more complex, requires close relationship with suppliers (IBM PERCS/Blue Waters)
Heat Spreader for Optical DevicesCooling / Load Saddle for Optical Devices
Optical Transmitter/Receiver Devices 12 channel x 10 Gb/s28 pairs per Hub - (2,800+2,800) Gb/s of optical I/O BW
Heat Spreader over HUB ASIC
Strain Relief for Optical RibbonsTotal of 672 Fiber I/Os per Hub, 10 Gb/s each
Hub ASIC (Under Heat Spreader)
A.Benner, Future Directions in Packaging (FDIP) Workshop, EPEP Oct 2010
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
16
M. Taubenblatt OFC 2011Dec 10, 2010 31© 2010 IBM Corporation
Power consumption of an optical Link
LaserDriver
TIA& LA
Vb2
LVb1
III-VVCSEL
PD
OE Module
OE Module
Pre-amp
Output Driver
See C. Lai et al – OFC 2011
Power distribution for a nominal 10mw/Gbps link (at 10-20Gbps)
Power vs Speed in a given technology
Tx PA: 2.0-2.5 mW/Gbps
Tx LD: 1.0-2.0 mW/Gbps
Tx VCSEL: 1.0 mW/Gbps
Tx Total: 4.0-5.5 mW/Gbps Rx TIA: 1.0-1.5 mw/Gbps
Rx LA: 3.0-3.5 mW/Gbps
Rx OD: 0.5-2.5 mW/Gbps
Rx Total: 4.5-7.5 mW/Gbps
• VCSEL PA/LD dependent on VCSEL quality
• Rx OD dependent on electrical channel length and quality
Low power data rate sweet spot for given CMOS technology generation and design
M. Taubenblatt OFC 2011Dec 10, 2010 32© 2010 IBM Corporation
Costs associated with an optics linkBill of Materials, e.g:• Substrates• Lenses• Laser & PD arrays• u-controller• Driver/receiver chips• Connectors
• optical and electrical• Fiber cabling• Heat sinks…
Test & Yield• Tester cost• Tester time• Built in self-test• Bit Error Rate
requirements…
Assembly & Fabrication• Assembly rate & Equipment costs• Output rate• Active vs passive alignment• Manual vs automated• Tolerances vs yield• Assembly (w/ server)…
Courtesy of Avago Technologies
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
17
M. Taubenblatt OFC 2011Dec 10, 2010 33© 2010 IBM Corporation
What will Help Optics Compete with shorter Copper links for Computer-com?
Increased Channel Rate
– Key challenges: VCSELs, electrical part of link
More parallelism
– Key challenges: packaging density, standardization
Move closer to the signal source (e.g. processor, switch)
– Key challenges: packaging integration, standardization
Links from standardized sub-components (no more modules)
– Key challenges: technology, supporting infrastructure (design tools, components, manufacturing and test)
IC-like cost structure using Integrated Photonics
– Key challenges: integration vs yield, packaging, volumes
M. Taubenblatt OFC 2011Dec 10, 2010 34© 2010 IBM Corporation
Outline Why we need Optical Interconnects for High Performance
Computing
Optical Link Basics
Requirements and Trade-offs
– Cost, Power, Density
Technologies of Interest
– VCSELs/fiber
– VCSELs/Optical PCB
– Si Photonics
Architectures: Optical Switching
Conclusions
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
18
M. Taubenblatt OFC 2011Dec 10, 2010 35© 2010 IBM Corporation
Optics Technologies of Interest
VCSELs/ multi-mode fiber
VCSELs/multi-mode optical waveguides on PCB or flex
Si Photonics/single-mode WDM
2D waveguide array
32 parallel channels
35 x 35m62.5m pitch
3.9 mm
3.9 mm
Terabus 160Gb/sTRx (bottom view)
M. Taubenblatt OFC 2011Dec 10, 2010 36© 2010 IBM Corporation
VCSELs & Multi-mode Fiber
Incumbent technology with well established infrastructure
Current low cost technology
Can be more highly integrated, but will always be heterogeneous packaging
Cost can be improved with:
– Higher speeds
– More automated, designed for low cost optics modules
– More BW/fiber
Power can be improved with higher speeds, faster CMOS or well designed SiGe and signal processing
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
19
M. Taubenblatt OFC 2011Dec 10, 2010 37© 2010 IBM Corporation
Development of VCSELs for >25 Gb/s links: collaborations with OE vendors
26Gb/s 30Gb/s
Joint work with Finisar AOC
Joint work with Emcore Corp
20Gb/s 25Gb/s
8mA700mVpp
*R. Johnson and D. M. Kuchta, “30Gb/s directly modulated VCSELs,” CLEO 2008
*N. Y. Li et al., “High-Performance 850 nm VCSELs and Photodetector Arrays for 25 Gb/s Parallel Optical Interconnects,” OFC 2010.*N. Y. Li et al., “Development of High-Speed VCSELs Beyond 10 Gb/s at Emcore,” Photonics West 2010.
6mA375mVpp
M. Taubenblatt OFC 2011Dec 10, 2010 38© 2010 IBM Corporation
90-nm IBM CMOS-Driven VCSEL Transmitters and Compatible Receivers
– Power and Speed Records
2.8 pJ/bit0.7 pJ/bit
20Gb/s
1.9 pJ/bit
29Gb/s 32Gb/s20Gb/s
1.75 pJ/bit
VCSEL Transmitters
Compatible Receivers
2.9 pJ/bit 3.5 pJ/bit 6.7 pJ/bit
15Gb/s 17.5Gb/s 20Gb/s
Progress in VCSEL based CMOS Transmitters and Receivers
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
20
M. Taubenblatt OFC 2011Dec 10, 2010 39© 2010 IBM Corporation
High volumes can take advantage of mass manufacturing
Mitch Fields, OFC 2010Courtesy of Avago Technologies
Manufacture of MicroPOD – paradigm shiftProduction requirement 30,000 pairs per monthSolution Simple vertical stack design
Invest in manufacturing technology for 100% automationManufacture parallel optics in panel form
Panel with MicroPOD modules
100% automated process.Assemble in panel form.
Singulate
M. Taubenblatt OFC 2011Dec 10, 2010 40© 2010 IBM Corporation
10 Gb/s10 Gb/sTXTX RXRX
7.5 Gb/s
10 Gb/s
Higher Density: 24-Channel Fiber-Coupled Optical Transceiver
[A. Rylyakov et al., OFC 2010 Post-Deadline]
RX Sensitivity
Merits afforded by flip-chip integrated “holey” IC package• Channel Count (24 TX + 24 RX)• Bandwidth (300 Gb/s)• Power (8.2 pJ/bit at 12.5 Gb/s)• Density (10 Gb/s/mm2)
[F. Doany et al., ECTC 2010]
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
21
M. Taubenblatt OFC 2011Dec 10, 2010 41© 2010 IBM Corporation
More BW/fiber: 4-Fiber 24-Core Optical Transceiver
silicon carriersilicon carrier
TXTXICIC
RXRXICIC
VCSEL arrayVCSEL array PD arrayPD array
backsidebackside
PD arrayPD array
VCSEL arrayVCSEL array
PCB
PDVCSELTX IC RX IC
Silicon Carrier
MCF
Custom VCSEL/PD Arrays Matched to 4 Multicore Fibers
Custom OE chips designed to fit into existing configuration of Terabus project—Match silicon carrier designed for 24-channel polymer optical waveguide transmitter.
Fabricated by Emcore Corporation
B.G. Lee OFC 2009Fiber fabricated by OFS Laboratories
M. Taubenblatt OFC 2011Dec 10, 2010 42© 2010 IBM Corporation
Single-Channel MCF-TX/RX PerformanceTX at 15 Gb/s TX at 15 Gb/s (~ 6 dB ER)
MCFMCF--TX aligned to 100TX aligned to 100--m MCF core.m MCF core.Output coupled to reference RX.Output coupled to reference RX.
RX at 12.5 Gb/sRX at 12.5 Gb/s
Ref TX coupled to 100Ref TX coupled to 100--m MCF core. m MCF core. Output aligned to MCFOutput aligned to MCF--RX.RX.[B. G. Lee et al., IEEE Phot. Soc. Mtg. 2010]
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
22
M. Taubenblatt OFC 2011Dec 10, 2010 43© 2010 IBM Corporation
Vcsels & Multi-mode Fiber : Challenges for the future
• VCSELs:• ~30Gbps likely• Continued focus on reliability lower temperatures can help• Continued focus on high performance VCSELs and Trx circuits for
lower power• Change of wavelength? 900-1100nm band or 850nm? no reliability
data, could slow acceptance
• Receiver performance at higher data rates:
• Higher data rates will require smaller photodiode size to achieve higher bandwidth• Smaller photodiode size will impact the ability to collect all the light from
the fiber – may lead to modal noise penalty• Smaller core (e.g. 30um) multi-mode fiber may be a good trade-off
• Receiver sensitivity will be one of the biggest challenges• becomes worse with increased data rate• impacts the jitter budget
• Reducing receiver power consumption becoming increasingly more difficult
M. Taubenblatt OFC 2011Dec 10, 2010 44© 2010 IBM Corporation
Optical Printed Circuit Boards - based on polymer waveguide technology
Vision
Optical
MCM
Optical
MCM
Optical
DIMM
Optical
DIMM
Optical
MCM
Optical
MCM
Optical
MCM
Optical
MCM
Waveguides
Waveguides
Low cost PCB card for control signals, power, ground
All off-MCM links are optical
Technology Progression
Optical Flex
OSALGA
Optical connector
Polymer WG’s
WG on PCB
OE elements
ProcessorSocketOptical PCB
CPU ProcessorSocketOptical PCB
CPU
WG in PCB
Technology Elements
8 waveguide flex sheets, 192 waveguides, 8 connectors
4x12 MT Waveguide connector
TRX1:
16TX + 16RX
TRX2:
16TX + 16RX
4x4 VCSELArray
4x4 PD
Array
TRX2:
16TX + 16RX
4x4 VCSELArray
4x4 PD
Array
48-channel Waveguide mirror array
waveguide cores on 62.5um pitch
48-channel Waveguide mirror array
waveguide cores on 62.5um pitch
Waveguide processing on large panels, e.g., 305 mm x 460 mm
Ink-jet / UV laser writing system
TIR mirrors with laser ablation
16+16 OSA, incl circuits >12.5 Gbps/ch
• Low cost pick and place asm’bly
• Addresses fiber mgt, including shuffles, spits
• Optics close to chips to alleviate electrical power to optics cost
Advantages
16+16 WG channels on PCB
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
23
M. Taubenblatt OFC 2011Dec 10, 2010 45© 2010 IBM Corporation
RXTXLDD RX
Si CarrierVCSEL
Lens Arrays
PD
Organic Carrier
PCBPolymer Wavegu
Optical Printed Circuit Boards (Terabus)Optochip: Based on a Si carrier platform for hybrid integration of
multiple chips Dense high-speed wiring layers for inter-chip connections Dense Thru Silicon Vias (TSV) for electrical I/O Optical vias for optical I/O Allows use of Industry standard 850nm vcsels (also lower WG loss than
980nm – 0.03dB/cm)
850-nm Optochip demonstrated300Gb/s bidirectional aggregate data rate
F. Doany ECTC 2010
6.4mm x 10.4mm
2x12LDD IC3
.9m
m 2x12RX IC
1.6mm
24-channel VCSEL Array
0.9 x 3.5mm
24-channel PD Array
0.9 x 3.5mm
Full link is operational, to be presented at ECTC 2011 (F. Doany)
TX: 20 Gb/sRX: 15 Gb/s
M. Taubenblatt OFC 2011Dec 10, 2010 46© 2010 IBM Corporation
-50 -40 -30 -20 -10 0 10 20 30 40 50
-4.5
-4
-3.5
-3
-2.5
-2
-1.5
-1
-0.5
0
Offset (um)
Coupl
ing
Eff
icie
ncy
(dB
)
-50 -40 -30 -20 -10 0 10 20 30 40 50
-4.5
-4
-3.5
-3
-2.5
-2
-1.5
-1
-0.5
0
Offset (um)
Coupl
ing
Eff
icie
ncy
(dB
)
Tx: ±35 µm
-80 -60 -40 -20 0 20 40 60 80
-3.5
-3
-2.5
-2
-1.5
-1
-0.5
0
Offset (um)
Coup
ling
Eff
icie
ncy
(dB
)
-80 -60 -40 -20 0 20 40 60 80
-3.5
-3
-2.5
-2
-1.5
-1
-0.5
0
-80 -60 -40 -20 0 20 40 60 80
-3.5
-3
-2.5
-2
-1.5
-1
-0.5
0
Offset (um)
Coup
ling
Eff
icie
ncy
(dB
)
Rx > ±65 µm
Design Study: One lens vs two lens coupling Single lens coupling design
– Beam focused to a point at the WG cores
– Tight OE module to WG board alignment tolerances
Si Carrier
SLC
Driver/TIA
OESi Carrier
Base Substrate
Driver/TIA :
OE
Lens Array SLC
Dual lens coupling design– Collimated beam between module and board
– Relaxed OE module alignment tolerances
Nearly constant coupling efficiency for:
– VCSEL : ±12µm
– Photodiode: ±10µm
-3.0
-2.5
-2.0
-1.5
-1.0
-0.5
0.0
0.5
-20 -15 -10 -5 0 5 10 15 20
X (microns)
Rel
ati
ve
Co
up
lin
g L
oss
(d
B)
SimulationMeasured
Flat efficiency withinmeasurement noise for
+/- 12 µm
VCSEL0
0.2
0.4
0.6
0.8
1
-50 -40 -30 -20 -10 0 10 20 30 40 50
Y (microns)
Re
lati
ve R
esp
on
se
SimulationMeasured
Photodiode
OE
YX
Relaxed module to board alignment tolerance of > ± 35 µm realized
VCSEL
Photodiode
Measured Alignment Tolerances
Measured Alignment Tolerances
after R. Budd, IBM Lower BOM cost
Higher Yield
or
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
24
M. Taubenblatt OFC 2011Dec 10, 2010 47© 2010 IBM Corporation
Advantages of Polymer Waveguide Technologycompared to Parallel Fiber Optics Integrated mass manufacturing
– Board, sheet, film level processing of optical interconnects– Lower assembly, waveguide jumper costs– Costs for wide busses should scale better
Simple assembly
– Avoid fiber handling– Similar assembly procedures as for electrical components and boards– Establish electrical and optical connections simultaneously– Avoid separate optical layer (if integrated with board)
New or compact functionality supporting new architectures
– Shuffles, Crossings, splitters, …– Enabler for multi-drop splitting, & complex re-routing that is expensive in fiber
Higher density, waveguide pitch < 125 um (best future fiber pitch)
– Higher bandwidth density, less signal layers
Cost
– Reduced Optics module cost AND jumper/connector costs both important– Possible Lower Maintenance costs
But
– Fiber / waveguide counts remain high for single optics– Still need lots of rack-to-rack fiber interconnects– Technology ecosystem needs to mature
M. Taubenblatt OFC 2011Dec 10, 2010 48© 2010 IBM Corporation
Si Photonics Vision/Goal
– Ultra low cost CMOS fabrication
– WDM/single mode fiber for high BW/fiber and distance at high data rate
– Nanophotonics for high BW density and integration
– Wafer level testing
– Ultra low power with low load devices and tight integration
Op
tical
I/O
Logic Plane
Off
-ch
ip o
pti
cal
sig
nal
s
On
-ch
ip o
pti
cal
tra
ffic
Photonic PlaneMemory Plane
Op
tical
I/O
Op
tical
I/O
Logic Plane
Off
-ch
ip o
pti
cal
sig
nal
s
On
-ch
ip o
pti
cal
tra
ffic
Photonic PlaneMemory Plane
3D Integrated Chip
F. Horst et al OFC 2010F. Horst et al PTL 21 (2009)
http://www.research.ibm.com/photonics
Also see Y. Vlasov OFC 2011
waveguideSi
Ge PD
waveguideSi
Ge PD
MZ modulator Ge WG detector
W.Green et al Optics Express, 2007J. Van Campenhout et al, Optics Express Dec 2010S. Assefa et al, Nature March 2010 & OE, April 2010
Y.VlasovSEMICON Dec 2010
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
25
M. Taubenblatt OFC 2011Dec 10, 2010 49© 2010 IBM Corporation
Goal: Integrate Ultra-dense Photonic Circuits with electronicsApplications: High Bandwidth On-chip optical network
Opt
ical
I/O
Logic Plane
Off
-ch
ip o
pti
cal s
ign
als
On
-ch
ip o
pti
cal t
raff
icPhotonic Plane
Memory Plane
3D Integrated Chip
WD
M W
DM
Switch fabric
WDM bit-parallel message
Message1Tbps aggregate BW
Core N
Deserializer
N parallel channels
Detector
Opt
ical
-to-
Ele
ctri
cal
Serializer
Message1Tbps aggregate BW
N parallel channels
Modulator
N modulators at different wavelengths
Ele
ctri
cal-
to-O
ptic
al
Core 1
Ultimate Vision, circa 2020:Optically connected 3-D Supercomputer Chip
Y.Vlasov, ECOC 2008, and Semicon, 2010Also, J. Kash, Photonics in Switching, SF, CA, pp 55-56, 2007
M. Taubenblatt OFC 2011Dec 10, 2010 50© 2010 IBM Corporation
Example: Performance Trade-Offs of Two Fabricated Photonic Switches
1810series resistance ()
0.0010.0010.02footprint (mm2)
90200diode length (µm)
1×22×2port configuration
yesnodigital response
< 34ON power (mW)
< 1> 100> 100optical bandwidth (nm)
21no. stages
Ring Resonator (2RRS)
Wavelength-Insensitive Mach Zehnder (WIMZ)1
WIMZWIMZ
2RRS2RRS
50-µm scale bars
1) J. Van Campenhout, Optics Express, 17 (26) 24020.B. Lee IEEE Photonics Society Nov 2010
• Mach-Zehnder: less T sensitivity, more power, larger area• Ring Resonator: high Q for lower power but T sensitivity,
also much smaller
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
26
M. Taubenblatt OFC 2011Dec 10, 2010 51© 2010 IBM Corporation
A. On-board:
No, already given up power and density
B. On-multi-chip-carrier:
Yes, could provide all high-speed off chip I/O at low-power
C. On-chip:
Yes, even better than B provided the integration comes for “free”: no added power, minimal added cost in processing or packaging
D. CMOS integrated:
No, digital real-estate too valuable in the substrate, back-end metal too crowded, too hard to optimize
What is the optimal level of integration for Si Photonics?
CMOS: CPUOptical Layer
Long electrical link is bad, costs power
Board packagingnot dense enough
CMOS: TX + RX
OR
Circuit Board
CMOS:CPU + Photonics
Circuit Board
CMOS: CPUPhotonic Layer
CMOS: TX +RX
Circuit Board
CMOS:CPU
Integrated Photonics
MemoryMemory
Circuit Board
CMOS: CPU Integrated Photonics
after C. Schow OFC 2009 Workshop
M. Taubenblatt OFC 2011Dec 10, 2010 52© 2010 IBM Corporation
Robust packaging of Si Photonics is needed
tapered glass tapered glass waveguides waveguides match pitch, match pitch,
cross section, cross section, and NA of chipand NA of chip
standard 250standard 250--µµm pitch SM/PM m pitch SM/PM fiber array aligned in Vfiber array aligned in V--groove, groove, buttbutt--coupled to glass waveguidescoupled to glass waveguides
B. G. Lee, et al Proceedings of OFC 2010, paper PDPA4.F. Doany et al, J. Lightwave Techn. 99, Oct 2010
Multichannel tapered coupler allows interfacing 250-µm-pitch PM fiber array with 20-µm-pitch silicon waveguide array 8-channel coupling demonstrated < 1 dB optical loss at interface Uniform insertion loss and crosstalk …but TBD: reliability, cost…
In collaboration with
Chiral Photonics
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
27
M. Taubenblatt OFC 2011Dec 10, 2010 53© 2010 IBM Corporation
Si Photonics
High potential for future HPC systems
– Ultra low cost and power
– Highly integrated
Focus for further progress
– Device choices and design for best power/speed/temperature tradeoffs
– Low cost packaging, including light source
M. Taubenblatt OFC 2011Dec 10, 2010 54© 2010 IBM Corporation
Optical Link Reliability for PetaScale+
•Effective FIT rate for 11ch ‘link’ with typical VCSEL Wearout and Random FIT=10/device•Sparing (12th device) effectively reduces FIT to low levels.
No SparingWearout
Random
12Khrs20hrs3.8 fails/day1M
39Khrs200hrs2.68 fails/week100K
174kHrs2Khrs1 fail/month10K
252Khrs20Khrs1.5 fails/year1K
Spare VCSEL only [time to 1st fail]
Spare VCSEL + 50 FIT unspared[time to 1st fail]
No Spare VCSEL + 50 FIT unspared
# of Links
ExaScale will need sparing + ultra-reliable components
11+1 spare
D. Kuchta IEEE Systems Packaging Japan Workshop, Jan 2010
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
28
M. Taubenblatt OFC 2011Dec 10, 2010 55© 2010 IBM Corporation
Outline Why we need Optical Interconnects for High Performance Computing
Optical Link Basics
Requirements and Trade-offs
– Cost, Power, Density
Technologies of Interest
– VCSELs/fiber
– VCSELS/Optical PCB
– Si Photonics
Architectures: Optical Switching
Conclusions
M. Taubenblatt OFC 2011Dec 10, 2010 56© 2010 IBM Corporation
What about new Architectures?
Rule #1: Cannot simultaneously use new Architecture AND new optics technology in the same generation…too much risk
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
29
M. Taubenblatt OFC 2011Dec 10, 2010 57© 2010 IBM Corporation
Typical architecture: Electronic Packet Switching
Typical architecture (electronic switch chips, interconnected by electrical or optical links, in multi-stage networks)
works well now---
– Scalable BW & application-optimized cost
• Multiple switches in parallel
– Modular building blocks • many identical switch chips & links)
-- but challenging in the future
– Switch chip throughput stresses the hardest aspects of chip design
• I/O & packaging
– Multi-stage networks will require multiple E-O-E conversions
• N-stage Exabyte/s network = N*Exabytes/s of costN*Exabytes/s of power
Central switch racks
J. Kash OFC tutorial 2008
By courtesy of Barcelona Supercomputing Center - www.bsc.es
M. Taubenblatt OFC 2011Dec 10, 2010 58© 2010 IBM Corporation
Scalable Optical Circuit Switch (OCS)
MEMS-based OCS HW is commercially available (Calient, Glimmerglass,..)
• 20 ms switching time• <100 Watts
Possible new architecture: Optical Circuit Switching
Optical Packet Switches (all-optical network, one E-O-E conversion from original source to final destination, <10 ns switching time)
are hard.– Not commercially available now
– Probably not cost-competitive against electronic packet switches, even in 2015-2020
But Optical Circuit Switches (>1millisecond switching time) are available
– Several technologies (MEMS, piezo-, thermo-,..)
– Low power, no extra O-E-O– But require single-mode optics
If multiple circuit-switch planes are provisioned, performance for parallel apps can be good
OCS
Input fiber(one channel
shown)
Output fibers
OCS Concept
2-axis MEMSMirror
(one channelshown)J. Kash OFC tutorial 2008 L.Schares OFC 2009
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
30
M. Taubenblatt OFC 2011Dec 10, 2010 59© 2010 IBM Corporation
EQ
control
2 Rx
central scheduler(bipartite graph matching algorithm)
VOQs
Tx
control
64 Ingress Adapters
Optical Switch
64 Egress Adapters
EQ
control
2 Rx
control links
8 Broadcast Units128 Select Units
8x11x88x1
Com-biner
Fast SOA 1x8FiberSelectorGates
Fast SOA 1x8FiberSelectorGates
Fast SOA 1x8WavelengthSelectorGates
Fast SOA 1x8WavelengthSelectorGates
OpticalAmplifier
WDMMux
StarCoupler
8x1 1x128VOQs
Tx
control
1
8
1
128
1
64
1
64
OSMOSIS Demonstrator System Centralized optical switch with electrical scheduler
64-ports at 40 Gb/s port speed Broadcast-and-select architecture Combination of wavelength and space division multiplexing fast switching based on SOAs R. Luijten, OFC 2009
In collaboration w / Corning Inc.
M. Taubenblatt OFC 2011Dec 10, 2010 60© 2010 IBM Corporation
Example of wavelength based packet routing
λF
λH
O/E
O/Econtrollogic
50:50
50:50
70:30
70:30
North
West
East
South
SOA
SOA
transparent routingwavelength-parallel
multi-λ packet
COLUMBIAUNIVERSITY
Lightwave Research LaboratoryO. Liboiron-Ladouceur, JLT, VOL. 26, NO. 13, JULY 1, 2008
Courtesy of K. Bergman, Columbia University
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
31
M. Taubenblatt OFC 2011Dec 10, 2010 61© 2010 IBM Corporation
Optical I/O
3D Chip Stack: (top) optical NoC, (center) memory layer, (bottom) processing cores
PIN-Diode MZI Modulator
[Green et al., Optics Express 2007]
Coupled Ring Resonator Switch
[Vlasov et al., Nature Photonics 2008]Ge MSM Photodetector
[Assefa et al., Optics Express 2010]
Silicon Photonics for Chip-Scale Photonic Communication: Key to low cost switching
Passive Elementso Waveguideso Crossingso WDMso SplittersActive Elementso Modulatorso Switcheso Detectors
Critical Switch Metricso Speedo Insertion Losso Optical Bandwidtho Optical Crosstalko Footprinto Static Power Consumptiono Dynamic Switching Energyo Thermal Insensitivityo Fabrication Tolerance
after B. Lee CLEO 2010
Add’l refs: A.Rylyakov ISSCC Jan 2011W Green et al, Proc Photonics Society Annual Meeting, 2010, pp. 512-513.
M. Taubenblatt OFC 2011Dec 10, 2010 62© 2010 IBM Corporation
Outline Why we need Optical Interconnects for High Performance
Computing
Optical Link Basics
Requirements and Trade-offs
– Cost, Power, Density
Technologies of Interest
– VCSELs/fiber
– VCSELS/Optical PCB
– Si Photonics
Architectures: Optical Switching
Conclusions
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
32
M. Taubenblatt OFC 2011Dec 10, 2010 63© 2010 IBM Corporation
Characteristics of “Computer-com”
Ultra-short distance
– Some 100m, many <10m, backplanes and cards eventually
Very wide busses
– 10’s to 100’s of channels
Very high aggregate BW
– Multi-Tbps busses, multi-Tbps off chip modules
High density
– > 100Gbps/cm2 (unidirectional T+R)
Highly integrated packaging
– Close to electrical signal source
Low power
– 10’s mW/Gbps 1mW/Gbps in the future
Very high volume/low cost per Gbps
– $1’s /Gbps << $1/Gbps in the future
Reliability
– Low BER, ECC, redundancy
M. Taubenblatt OFC 2011Dec 10, 2010 64© 2010 IBM Corporation
Conclusions
Computer-com optics is leaving the “plateau of productivity” and rounding the corner of the hockey stick!
..but there’s a lot of trade-offs to be considered.
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf
33
M. Taubenblatt OFC 2011Dec 10, 2010 65© 2010 IBM Corporation
Selected ReadingIn addition to talks and papers cited in the presentation, the following may be useful:
Computer Architecture and Optics Requirements A. F. Benner et al, “Exploitation of optical interconnects in future server architectures”, IBM J. Res. & Dev. Vol. 49 No. 4/5 July/Sept. 2005 J. Hennessy & D. Patterson, Computer Architecture, Morgan-Kaufman, 2003 P. Kogge et al, “Exascale Computing Study, Technology Challenges in Achieving Exascale systems”, Sept 2008,
http://www.er.doe.gov/ascr/Research/CS/DARPA%20exascale%20-%20hardware%20(2008).pdf
High Speed Electrical Interconnects: D.G. Kam, et al, “Is 25 Gb/s On-Board Signaling Viable?”, IEEE Transactions on Advanced Packaging, May 2009 Vol: 32 Issue:2 328 - 344
Optical Interconnects P. Pepeljugoski et al, “Modeling and Simulation of the next generation multimode fiber, IEEE J. Lightwave Technol., vol. 21, 5 1242-, 2003 P. Pepeljugoski et al, "Development of System Specification for Laser Optimized 50-um Multimode Fiber for Multigigabit Short-Wavelength
LANs", IEEE J. Lightwave Technology 21(5), pp.1256-75, May 2003 D. Kuchta “Progress in VCSEL based Parallel Links”, in VCSELs --- Fundamentals, Technology and Applications of Vertical-Cavity Surface-Emitting Lasers, Eds. R. Michalzik and F. Koyama, to be published, Springer-Verlag Berlin-Heidelberg-New York in 2011 D.A.B Miller, Optical Interconnects OFC Tutorial 2010 http://ee.stanford.edu/~dabm
Optical PCB F. E. Doany et al, "Terabit/s-Class 24-Channel Bidirectional Optical Transceiver Module Based on TSV Si Carrier for Board-Level
Interconnects," Electronic Components and Technology Conference (ECTC) 2010, pp. 58–65, Jun 2010 D. Jubin et al, Polymer waveguide-based multilayer optical connector, Proc. SPIE, Vol. 7607, 76070K (2010); doi:10.1117/12.841904 F. D. Doany et al,"160 Gb/s Bidirectional Polymer-Waveguide Board-Level Optical Interconnects using CMOS-Based Transceivers," IEEE
Trans. Adv. Packag., Vol. 32, No. 2, pp. 345-359, May 2009. R. Dangel et al, "Polymer-Waveguide-Based Board-Level Optical Interconnect Technology for Datacom Applications," IEEE Transactions on
Advanced Packaging, vol. 41, pp. 759-767, Nov 2008
Si Photonics: Y. Vlasov et al, multiple papers can be downloaded at: http://www.ibm.research/photonics Dries Van Thourhout, Si Photonics OFC Tutorial 2010 http://photonics.intec.ugent.be/download/ D.A.B Miller, Optical Interconnects OFC Tutorial 2010 http://ee.stanford.edu/~dabm
OThH3.pdf
OSA/OFC/NFOEC 2011 OThH3.pdf