John Bowers - 2013 UCSB Faculty Research Lecture
John Bowers - 2013 UCSB Faculty Research Lecture
John Bowers - 2013 UCSB Faculty Research Lecture
2013 Faculty Research Lecture 1 John E. Bowers, Director, Institute for Energy Efficiency Kavli Professor of Nanotechnology
Why is it important? What is the future? What is the problem? What is the economic driver?
Big Data
Internet Users: 2 billion (>20% of World population) Google Search: More than a billion searches daily YouTube adds 24 hours of video every minute, 1+ billion views a day
Tbytes/month
Today
Source: Network Traffic Forecast and Analysis 2013-2017 by Information Gatekeepers 2013.
100x
Kilper et, al., Power Trends in Communication Networks, J. Selected Topics in Quantum Electronics, March/April 2011
C. Lange et al., Energy Consumption of Telecommunication Networks and Related Improvement Options, J. Selected Topics in Quantum Electronics, March/April 2011
Data Servers
Source: Energy Cost, the Key Challenge of Todays Data Centers, Michael Poess (Oracle) and Raghunath O. Nambiar (HP) 2010
10
11
12
13
From Donn Lee (Facebook)
Servers
14
15
16
Follow the sun data center strategy: Global load balancing. Use cheaper electricity at night.
18
100 Million
TAT-9
TAT-11
TAT-10 TAT-12/13
WDM terrestrial
20
21
Wireless Telegraphy
1902 Marconi Invention of wireless telegraphy (radio) Wireless transmission across the Atlantic
1912 Word of the Titanic sinking sent to shore. Marconi had been scheduled on that trip, but had left on the Lusitania three days earlier.
22
24
Multiplexing is Crucial
Time multiplexing was invented in 1910 by George O. Squier. It allowed multiple telephone calls to be transmitted over one wire.
Coax Cable
1931 Lloyd Espenschied and Herman A. Affel invented coaxial cable
Cable with 22 coaxial cables that can carry 90,000 telephone calls
26
Transistor
The transistor was demonstrated in 1947 and Shockley, Bardeen, Brattain received the Nobel Prize in 1956. This enabled solid state amplifiers.
Microwave cable and microwave antennas were dominant through the 1960s
6 cm diameter circular waveguides with 40 to 120 GHz for 230,000 voice circuits.
31
Robert Mauer, Donald Keck, and Peter Schultz, from Corning purified silica and reduced optical loss to 16 dB/km in 1970.
32
33
400x
TAT-9
38
Hypernia 10 Tbps
1860 1880 1900 1920 1940 1960 1980 2000 2020 2040 Year
EDFA WDM
1860 1880 1900 1920 1940 1960 1980 2000 2020 2040 Year
100 Million
EDFA WDM
1860 1880 1900 1920 1940 1960 1980 2000 2020 2040 Year
100
Polarization Division Multiplexing started
Tb/s
10
WDM started
6 Million In 32 years
1990
1994
1998
2002
2006
2010
1977: 50 Mbit/s
Long Haul:
6 million increase in capacity in 32 years. 100 million decrease in energy/bit in 150 years.
43
44
Silicon MEMS Mirror Plane 384 Mirrors Electrostatic Control Moves in X & Y 10 nA x 100 V = 1 microW/channel 2D Gymbal Mirror Silicon Photonic integration is key! Impossible to build large (>100 port) switches without integration. Large wafer size essential. CMOS process uniformity essential (hinges).
>10,000x
Many electronic servers and switches require significant power (1 MW for some routers)
47
48
49
50
Moving to Interconnects
Optical
Metro & Long Haul
0.1 80 km
Copper
Chip to Chip
1 50 cm
Billions
Board to Board
50 100 cm
Rack to Rack
1 to 100 m
Millions
Volumes
51
52
53
Reduce cost by sharing VLSI facility with electronics Improve yield by taking advantage of silicon process development Volume driver: Solve IC interconnect bottleneck (from 4 Tbps to 1 Pbps). Optical transmitters/receivers on processors, memories, switches.
54
High volume, low cost Highly integrated Silicon Photonic Integrated circuits
2007 2010 2013 2016 65 43 32 22 300 Courtesy: 300 300 450 Rattner (Intel)
Six Generations
Scalability
2001 130 200 2004 90 200
Silicon
Optical gain from III-V Material Efficient coupling to silicon passive photonic devices No bonding alignment necessary: suitable for high volume CMOS All back end processing low temperature (<350 C)
56
58
19 dB/mm amplifier
(Kurczveil, SPIE PW, 2011)
ring
AMP
detector
Microring laser
Photonic Integration
62
1. Intel is working with Facebook and Quanta to define a new class of server architectures 2. First architecture is disaggregation 3. Intel has sampled its 100G photonic modules
To Spine Switches
Optical Rack
Compute Network
I/ O Appliance
RRC
SiPh
Server
Network
Storage
Compute Server Server Server Mem Mem Mem DDR CPU PCIe DDR CPU PCIe DDR CPU PCIe
SiPh
Xeon: PCIe Atom: Enet Server Server Server Compute Remote Storage HDDs SSDs CPU PCIe RRC Mem Mem Mem DDR CPU DDR CPU DDR CPU RRC
SiPh SiPh
Silicon PIC
interconnect
xR
xT x
Now
1 Year
3 Years
5 Years
7 Years
DWDM
10 Years
CWDM
>.1 pJ/bit
Modulators: 23 GHz, >15 dB ER, 1.3 Vpi >36 GHz, 0.8 A/W Photodetectors
0 5 10 15 20 25 30 35 40
-2
-4
-6
Frequency (GHz)
69
Fibers have low loss: 0.0002 dB/m (0.2 dB/km) Optical waveguides on InP, GaAs, Si have high loss: 30 dB/km How can we get fiber losses on chips?
Power (dBm)
FSR=25.3nm 6 7 8 9 10 11 12 13 14 15 16 1
al. Optics 1527et 1531 1535 1539Express 1543 1547 (2011) 1551 1555 silicon substrate Bauters High Q Resonators (80Million Wavelength (nm)
SOA
Interconnects
Assumes each generation will require a doubling of port speed to match improvements in processor performance Total device power must be less than 130W for normal forced air cooling with heatsink, 200W possible with special cooling. Assume max power budget of 50% (65W) for IO.
Optical IO on CMOS
als
optica
Photonic Plane
On-chip o
l si g n
ptical tr affic
Memory Plane Logic Plane
Kash, Photonics in Supercomputing: the Road to Exascale, IPNRA, 2009
Off-ch ip
Computational Throughput
10 TFLOPS
1 GFLOPS/W
Power Consumption
Haney et al, OFC 2010, paper OMV2 Haney OFC (2011)
Performance enabled by a ~10 TFLOPS many-core chip with integrated chip I/O at ~.4 pJ/b:
~100 GFLOPS/W
Peta-FLOPS in a rack
1 PetaFLOPS
Computational Throughput
100 TFLOPS
10 TFLOPS
Assuming: 4 pJ/b chip I/O link power (for 10-100 cm links) 200 W/chip with 40 W allocated to I/O maintaining ~.5 Bps/FLOPS I/O Performance barrier at ~10 GFLOPS/W limit
1 MW 100 kW 10 kW 1 kW
Power Consumption
Haney et al, OFC 2010, paper OMV2 Haney OFC (2011)
Lower threshold lasers Higher slope efficiency Di Liang Chong Zhang Alan Liu
Stranski Krastanov growth High cw temperature (115C) High cw power (60 mW)
80
Oscillators
The quietest oscillators are now optical (NIST)
Ultra-stable CW Laser Femtosecond Laser Comb
81
Oscillators
The quietest oscillators are now optical (NIST) Integrate on a chip! Integrated mode locked lasers (Sudha Srinivasan, Mike Davenport) High Q resonators (Daryl Spencer) (90 million)
82
Spectroscopy-on-chip
Theogarajan, Meinhart, Moskovits
Absorp=on
spectroscopy
Broadband
probe
spectro
meter
Broadband
Filter
probe
D D D D
Fluorescence
spectroscopy
Laser
probe
spectro
meter
Raman
spectroscopy
D
Laser
probe
Filter
spectro
meter
D D
Courtesy: Baets
43
papillary dermis
reticular dermis
Courtesy: Baets
G. Yurtsever et al 36
Acknowledgements
UCSB: Former students and postdocs: Alex Fang, Jared Bauters, Hui Wen Chen, Daoxin Dai, Jon Doylend, Martijn Heck, Sid Jain, Geza Kurzveil, Brian Koch, Di Liang, Hyundai Park, Molly Piels, Paolo Pintus, Matt Sysak, Yongbo Tang, Jason Tien Present students: Jock Bovington, Mike Davenport, Jared Hulme, Alan Liu, Jon Peters, Daryl Spencer, Alex Spott, Eric Stanton, Sudha Srinivasan, Chong Zhang Colleagues: Rod Alferness, Dave Auston, Dan Blumenthal, Larry Coldren, Nadir Dagli, Steve Denbaars, Art Gossard, Herb Kroemer, Chris Palmstrom, Mark Rodwell, Adel Saleh, Luke Theogarajan Intel : Richard Jones, Yimin Kang, Mario Paniccia, Matt Sysak Aurrion: Alex Fang, Greg Fish, Rob Guzzon, Eric Hall, Brian Koch, Erik Norberg, Anand Ramaswamy, John Roth, Dan Sparacin Hewlett Packard: Di Liang, Geza Kurzveil, Ray Beausoleil Cleanroom, MBE, MOCVD staff My Family: Ariel, Eric, Steve, Chris and Laetitia
86
Summary
CMOS compatible silicon photonics combines two great inventions:
The transistor microprocessor The laser optical communications
and enable >100x lower power and lower power density. Silicon photonics makes sense for high volume markets like data centers, supercomputers, biosensors, and telecommunications and for highly integrated, high performance markets like low phase noise oscillators, gyroscopes, sensors.
87