Optical Interconnect to Chips

Size: px

Start display at page:

Download "Optical Interconnect to Chips"

Katrina McKenzie
6 years ago
Views:

Optical Interconnect to Chips David A. B. Miller Stanford University http://ee.

1 Optical Interconnect to Chips David A. B. Miller Stanford University 2/29/08 Computing of the Future, David A. B. Miller, Stanford 1

2 Summary Wiring, optics, and communicating information Electronics is remarkable, but wiring does not scale well Requirements on devices for optical interconnects Energies particularly challenging Silicon photonics technology Missing optical output device Germanium quantum well physics and devices New strong modulator mechanism for silicon systems Nanometallic antennas Concentrating light into deeply subwavelength devices Very small wavelength splitters? Fundamental limits to optical components Conclusions 2/29/08 Computing of the Future, David A. B. Miller, Stanford 2

3 Collaborators and Funding Collaborators Funding Stanford students Elizabeth Edwards, Onur Fidaner, Y. Ge, Martina Gerken, Yang Jiao, Ekin Kocabas, Yu-Hsuan Kuo, Salman Latif, Y.-K. Lee, Dany Ly- Gagnon, Ali Okyay, Joe Matteo, Bianca Nelson, Jon Roth, Rebecca Schaevitz, Shen Ren, Luke Tang, Y. Yuen Stanford faculty Others Intel Mark Brongersma, Shanhui Fan, Jim Harris, Bert Hesselink, Krishna Saraswat Ted Kamins (HP) FCRP/DARPA Interconnect Focus Center DARPA EPIC, UNIC, and Optocenters programs AFOSR Nanometallics MURI 2/29/08 Computing of the Future, David A. B. Miller, Stanford 3

4 Levels of interconnection Telecommunications Campus networks LANs 10,000 km 1000 km 1 km 100 m interconnect distance Optics currently dominates for long distance interconnects Increasingly, optics is used in local area network applications 2/29/08 Computing of the Future, David A. B. Miller, Stanford 4

is encountering severe limitations itations -- can optics help at these

5 Levels of interconnection Backplanes & board-to-board Chip-to-chip On-chip 1 m 10 cm 1 mm interconnect distance Electrical signaling within computers is encountering severe limitations itations -- can optics help at these length scales? 2/29/08 Computing of the Future, David A. B. Miller, Stanford 5

6 Wiring, Optics, and Communicating Information - Architectural Aspect Ratio Limit to Capacity l A this wire carries the same number of bits per second as this wire Get universal form of scaling for simple digital connections (no repeaters, no multilevel modem techniques) bit rate B A / l 2 B ~ A / l 2 bits/s for LC lines B ~ A / l 2 bits/s for RC lines B ~ A / l 2 bits/s for equalized LC lines Miller and Ozaktas (1997) Once the wiring fills all space, the capacity cannot be increased either by making the system smaller or making it larger Optics completely avoids this scaling limitation 2/29/08 Computing of the Future, David A. B. Miller, Stanford 6

7 Electrical Communication Energy Wires always have large capacitance per unit length and/or low impedancei (Lines on chips are anyway nearly always RC) Simple logic-level level signaling on-chip chip results in specific dissipation E.g., at 2pF/cm and a 2 cm chip, at 1 V on-off off signaling Dissipate at least ~ ½CV 2 per bit sent across chip ~ 2pJ Basic impedance mismatch between small logic devices and low impedance/high capacitance of wires electrical connection low impedance and/or high capacitance / unit length small, high-impedance devices 2/29/08 Computing of the Future, David A. B. Miller, Stanford 7

8 Optical and Electronic Physics - Differences Very short wavelength nm nm (electronics 3 cm cm m) m) Optics Very high frequency THz THz (electronics MHz MHz GHz) Large photon energy 2 ev ev (electronics nev nev µev) Miller /29/08 Computing of the Future, David A. B. Miller, Stanford 8

9 Features of Optics for Interconnection High frequency of optics Can carry a very large amount of information Can also have very short pulses (less than 1 picosecond) Short wavelength Allows the use of very low-loss loss optical fibers use dielectric guides larger than lossy metals Possibility of imaging interconnects 10,000 connections with one lens Both fibers and imaging enable very dense interconnects Large photon energy Quantum mechanical reception of signals means No pick-up of electrical noise Optical isolation of different voltage levels May allow lower power connections quantum impedance conversion Miller /29/08 Computing of the Future, David A. B. Miller, Stanford 9

10 Wiring, Optics, and Communicating Information Optics as a Solution at Many Levels? Requirements on optical technology For shorter distances, low power more important integration and efficient devices essential Electrical system energies ~ 100 pj per bit for backplane connections ~ 10 pj per bit for on-module (chip chip) interconnections ~ 1pJ per bit ( ( 1 mw/(gb/s)) for global interconnections on chip Optical system energies should be ~ 10 times lower for sufficient advantage Optical output device energies should be x3 x10 lower still than these system energies 2/29/08 Computing of the Future, David A. B. Miller, Stanford 10

11 What are the technical challenges for optics on chip? Device performance Want output device efficient at 10fJ optical transmitted energies Given ~ 1 V drive, capacitance of output device therefore < 10 ff device must work at injected charge levels ~ 10 fc ( ~ 10 5 electrons) equivalent to inverting a few square microns of one quantum well nanocavities likely essential for lasers Want detector and receiver that work at 1fJ received optical energy Therefore < or ~ 1fF photodetector Note 1 micron cube of semiconductor has a capacitance of ~ F (100aF) 1 fj will give ~ 1V swing in detector receiverless operation 2/29/08 Computing of the Future, David A. B. Miller, Stanford 11

12 What are the technical challenges for optics on chip? Systems and integration requirements Need to work at CMOS voltages < or ~ 1V Need to work at very high speeds Need technology with >> 10GHz speed possibilities Need dense optical channels At least as dense as current global wiring channels Otherwise cannot make a significant impact on the global interconnect traffic Hence need ~ 1000 s s of channels Must be manufacturable in integrated platform with CMOS 2/29/08 Computing of the Future, David A. B. Miller, Stanford 12

13 Universal Electronics/Optics/Optoelectronics Platform for All Levels? Use same physical technology for interconnect connecting to longer distance optical networks? E.g., interface WDM optical fiber directly to silicon? WDM splitters/combiners Modulator and detector arrays hybridized to, or made in silicon? Key problem optical output device integrable with silicon, with good enough performance i.e., comparable to that in III-Vs Input WDM fiber Optical power from multiwavelength source Modulator array WDM splitter CMOS chip Detector array WDM combiner Output WDM fiber 2/29/08 Computing of the Future, David A. B. Miller, Stanford 13

14 Silicon Photonics Technology Current status broad range of demonstrated optical waveguide, detector, and modulator technologies with see silicon, silicon dioxide, silicon nitride, SiGe and Ge all of these materials can be grown compatible with silicon CMOS fabrication technology Special Issue on Silicon Photonics, IEEE J. Selected Topics in Quantum Electronics Vol 12 Issue:6 Part: 2 (Nov.-Dec. 2006) 2/29/08 Computing of the Future, David A. B. Miller, Stanford 14

Silicon Photonics Devices waveguides e.g., silicon ridge on SiO 2 (SOI), silicon wire detectors emitters Ge detectors on silicon see, e.g., recent representative work Huang et al.

15 Silicon Photonics Devices waveguides e.g., silicon ridge on SiO 2 (SOI), silicon wire detectors emitters Ge detectors on silicon see, e.g., recent representative work Huang et al. (2006), Koester et al. (2006), Colace et al. (2006), Luxtera Corp. apparently no electrically pumped room temperature silicon laser many mechanisms would anyway not give lasers that could be modulated at high frequency even for III-Vs, not clear that direct modulation is viable for low power and high data rates 10um fiber core waveguides Luxtera Corp. (courtesy Cary Gunn) 2/29/08 Computing of the Future, David A. B. Miller, Stanford 15

$Silicon Photonics Devices modulators refractive modulators based on carrier density index change technically very impressive use of a very weak mechanism Mach-Zehnder modulators see, e.g., Liao et al.$

16 Silicon Photonics Devices modulators refractive modulators based on carrier density index change technically very impressive use of a very weak mechanism Mach-Zehnder modulators see, e.g., Liao et al. (2006), Luxtera Corp. But high operating energies (10 s of pj per bit) ring-resonator resonator modulators see, e.g., review by Lipson (2006), Luxtera Corp. but high-q Q devices needing very precise tuning thermal tuning requires additional power may also have too much chirp for long-haul courtesy Michal Lipson (Cornell) 2/29/08 Computing of the Future, David A. B. Miller, Stanford 16

17 Refractive Mechanisms for Modulators in Silicon and Germanium (1) Carrier density index change used for modulators so far Δ = Δ + Δ ( ) n N P Soref and Bennett (1987) ΔN (ΔP) electron (hole) density in cm -3 typical operating energies to make the necessary change in carrier density for a Mach-Zehnder modulator ~ 10 s s of pj (Liao et al. (2006)) weak mechanism leading to very high powers without resonators 2/29/08 Computing of the Future, David A. B. Miller, Stanford 17

18 The Missing Modulation Mechanism Want mechanism Compatible with silicon CMOS manufacture for low cost Works at telecommunications wavelengths and speeds Low energy and strong, for small, dense devices so can address all application areas Hence need performance as good as good III-V V devices But integration of III-V s s on silicon still challenging Group III and Group V elements are dopants for silicon However, some hope for future use of III-Vs III-V V modulators can successfully be grown on silicon (Goossen et al. 1989) silicon manufacturers are researching possible introduction of III-V s s for better transistors Ideally want a mechanism as strong as those in III-V s s but in a Group IV system 2/29/08 Computing of the Future, David A. B. Miller, Stanford 18

19 New optical modulator mechanism for silicon chips Quantum-confined Stark effect (QCSE) Strongest high-speed optical modulation mechanism Used today for high-speed, low power telecommunications optical modulators but in III-V V semiconductors QCSE in germanium quantum wells on silicon substrates Fully compatible with CMOS fabrication processes Can work over C-band C at 1.55 µm Surprises Works in indirect gap semiconductor Uses Ge direct gap absorption Actually better than in III-Vs Clearer peaks, stronger absorption Y.-H. Kuo, Y.-K. Lee, Y. Ge, S. Ren, J. E. Roth, T. I. Kamins, D. A. B. Miller & J. S. Harris, Nature 437, (2005) Funded by DARPA EPIC Program, Intel, MARCO/DARPA Interconnect Focused Research Center J. Harris and D. Miller groups, Stanford University 2/29/08 Computing of the Future, David A. B. Miller, Stanford 19

20 Quantum-Confined Stark Effect Energy Speed Electrostatic to charge or discharge device ( ) CV ( ) 2 2 1/2 1/2 εε r o E dv 13 volume Need field E ~ 10 5 V/cm for good modulation Need only a few microns optical path (1 10) for good absorption contrast Hence, per square micron of device cross-sectional sectional area, need ~ 6 60 fj of energy if no resonator/slow light 2 3 orders of magnitude less than, e.g., silicon carrier density refractive index modulators Speed QCSE works as fast as we can get the electric field onto the devices Fundamental limit to speed << 1 ps ε r 2/29/08 Computing of the Future, David A. B. Miller, Stanford 20

21 Strain-Balanced Structure n+ SiGe cap layer growth direction Undoped SiGe buffer layer Ge/SiGe quantum wells Y. H. Kuo Undoped SiGe buffer layer p+ Relaxed SiGe buffer layer Silicon Substrate Compressive Tensile Strain force ε Average Si concentration in Ge/SiGe quantum wells equals that in SiGe buffer Allows growth of thick structures without exceeding critical thickness 2/29/08 Computing of the Future, David A. B. Miller, Stanford 21

22 Side entry modulator on SOI in C-bandC Resonant cavity Input port Reflectors Si Substrate Optically active material Output port 50 nm buried oxide layer Buried oxide layer increases reflectivity frustrated total internal reflection 3 db modulation possible with 1 V swing (from 3.5 to 4.5 V) Shows Ge quantum wells can be grown on SOI also Operates in C-band C when heated to 100 C J. E. Roth, O. Fidaner, E. H. Edwards, R. K. Schaevitz, Y.-H. Kuo, N. C. Helman, T. I. Kamins, J. S. Harris, and D. A. B. Miller, Electronics Lett. 44, (2008) wavelength nm 2/29/08 Computing of the Future, David A. B. Miller, Stanford 22 Percentage transmission Transmission through SOI Modulator, Green: 1V,2V,...Red: 5V

First waveguide Ge quantum well devices Waveguide mode (in multimode guide) SiO 2 SiGe Silicon MQW Fabricated modulator, interconnect waveguide, and detector CW Light Modulator Interconnect Detector

23 First waveguide Ge quantum well devices Waveguide mode (in multimode guide) SiO 2 SiGe Silicon MQW Fabricated modulator, interconnect waveguide, and detector CW Light Modulator Interconnect Detector Modulator Interconnect First waveguide modulator detector interconnecting waveguide with Ge quantum wells on Si substrates Transmission (db) transmission (d Preliminary modulator transmission with bias voltage transmission try2 0 V 6 4 V 6 V wavelength 1490 (nm) Responsivity (A/W) 0.35 responsivity (A/W) Detector responsivity Wavelength (nm) wavelength 1480 (nm) Wavelength (nm) O. Fidaner et al., IEEE Photonics Technol. Lett. 19, (2007) 2/29/08 Computing of the Future, David A. B. Miller, Stanford V 2.5 V 5.0 V 7.5 V 10.0 V 0.5 V 2.5 V 5 V 7.5 V 10 V Detector eye diagram at 2.5 Gb/s

24 Mask and layout for nanoantenna 2/29/08 Computing of the Future, David A. B. Miller, Stanford 24

25 Dipole Antenna Enhanced Photodetector for 1.3 micron Wavelength Antenna arms Collecting electrodes 500 nm Ge detector element Y X Goal couple into deeply subwavelength detectors Antenna reduce capacitance to very low (<< 1 ff ) levels match micron optical wavelengths to deeply sub-micron devices 160 nm long dipole antenna arms Ge detector element: 150nm long (between collection electrode), 50nm wide (between antenna arms) and 80nm thick Antenna enhancement 24 times greater photocurrent for y polarization (antenna direction) than x polarization (@ 30 mv bias) evidence for strong antenna effects and field enhancement Liang Tang, S. Ekin Kocabas, Salman Latif, Ali K. Okyay, Dany Ly-Gagnon, Krishna C. Saraswat and David A. B. Miller (to be published, Nature Photonics) 2/29/08 Computing of the Future, David A. B. Miller, Stanford 25

26 Compact WDM splitters? Nanophotonic superprism array wavelength splitter Could we make very compact wavelength spliters? Superprism effects in photonic nanostructures very large separation of wavelengths in compact structure form of dispersion can be engineered by design e.g., 66 layer non-periodic structure separates 4 wavelengths In Out 1 Out 2 Out 3 Out 4 Mirror Substrate Focussing lens x y Mirror z Dielectric Stack Bianca E. Nelson, Martina Gerken, David Miller et al. Opt. Lett. 25, (2000) 2/29/08 Computing of the Future, David A. B. Miller, Stanford 26

27 Periodic vs. Non-Periodic Structure layer periodic (experiment) 200-layer periodic (theory) 66-layer non-periodic (experiment) 66-layer non-periodic (theory) Shift in μm Shift in um Wavelength in nm 66-layer non-periodic structure has larger, more linear shift than 200 layer periodic structure Martina Gerken and David A. B. Miller, IEEE Photonics Technol. Lett. 15, (2003); Experimental results are scaled to 1550 nm for comparison. Applied Optics 42, (2003) 2/29/08 Computing of the Future, David A. B. Miller, Stanford 27

Two dimensional design compact mode splitter Essentially randomly add and/or subtract cylinders within a region

After ~10000 search steps (48 hours on a Pentium III computer) We have no idea why it works!

make it? Multimode input Engineer precise mode splitting with positioning of dielectric columns Y. Jiao, S.

28 Two dimensional design compact mode splitter Essentially randomly add and/or subtract cylinders within a region to try to get desired function, iteratively Successfully designed possibly smallest mode splitter ever designed After ~10000 search steps (48 hours on a Pentium III computer) We have no idea why it works! Negligible intuition to guide the rod pattern If we have no idea how it works, how can we know how good we could make it? Multimode input Engineer precise mode splitting with positioning of dielectric columns Y. Jiao, S. Fan, and D. A. B. Miller., Optics Lett. 30, (2005) Single mode outputs 2/29/08 Computing of the Future, David A. B. Miller, Stanford 28

29 Predicted performance of designed linear dispersers Wavelength Normalized Wavelength Range / Range Shift-Model [a.u.] [nm/μm] Model 10 % < relative error < 25 % relative error < 10 % Chirped Designs Coupled-Cavity Design Optimized Design Empirical shift model for a good design Δ s= c ( Δλ ) Δn = 16 L sin n 2 avg M. Gerken and D. A. B. Miller, Applied Optics 44, (2005) Dispersion [um/nm] [μm/nm] Results from 623 different designs of multilayer linear dispersive elements Plotted as ratio of wavelength range Δλ / total spatial shift Δs,, against dispersion c Disp (spatial shift per unit wavelength) Strongly indicates underlying limit fundamental limit in performance ( θ ) 2/29/08 Computing of the Future, David A. B. Miller, Stanford 29 Disp max

Probably there is a limit Somewhere between And 1 cubic micron A room full of optics Or at least an arrayed waveguide grating D. A. B.

30 Is There a Fundamental Limit to Nanophotonic Devices? Suppose we want, e.g., to split 32 wavelengths to different positions Using, e.g., glass and air Intuitively Is there a limit to how little material we need If so, what is the limit? Probably there is a limit Somewhere between And 1 cubic micron A room full of optics Or at least an arrayed waveguide grating D. A. B. Miller, "Fundamental limit for optical components," J. Opt. Soc. Am. B 24, A1-A18 (2007) 2/29/08 Computing of the Future, David A. B. Miller, Stanford 30

Example 1-D D Pulse Dispersion scattering volume receiving volume Pulse entering from the left containing identical pulses at four different center frequencies is dispersed into separate pulses in

31 Example 1-D D Pulse Dispersion scattering volume receiving volume Pulse entering from the left containing identical pulses at four different center frequencies is dispersed into separate pulses in the receiving volume Want to deduce a limit to such a device by counting modes i.e., orthogonal waves that can be generated in the receiving space as a result of scattering Δω input pulses δω D. A. B. Miller, "Fundamental limit for optical components," J. Opt. Soc. Am. B 24, A1-A18 (2007) 2/29/08 Computing of the Future, David A. B. Miller, Stanford 31

32 Example Result for 1D Pulse Separator For a transmissive device 3 π NB + NS λ η max Length of scattering volume in wavelengths maximum variation of dielectric constant relative to background For example, to separate pulses of 32 different equally spaced center wavelengths, near 1.55 microns using glass ( (ε r ~ 2.25) and air, then ηmax 1.25 then the length has to be greater than ~ 41.7 microns completely independent of the details of the design! D. A. B. Miller, "Fundamental limit for optical components," J. Opt. Soc. Am. B 24, A1-A18 (2007) D. A. B. Miller, Fundamental Limit to Linear One-Dimensional Slow Light Structures, Phys. Rev. Lett. 99, (2007) 2/29/08 Computing of the Future, David A. B. Miller, Stanford 32

33 Conclusions Optics has many attractive features for interconnects Silicon photonics has been advancing rapidly Ge quantum wells on silicon may give missing output device mechanism in Group IV semiconductors as good as the best in III-V s 2 3 orders of magnitude stronger mechanism than carrier density index change in silicon Very promising for devices Integrated with CMOS With high speed and low power Possible basis for one technology platform from interconnects to long haul Nanometallics offer strong light localization for improved devices es Nanophotonic limits now being understood suggest very small splitter devices 2/29/08 Computing of the Future, David A. B. Miller, Stanford 33

OThX1.pdf OSA / OFC/NFOEC /10/$ IEEE

OThX1.pdf OSA / OFC/NFOEC /10/$ IEEE 978-1-55752-884-1/10/$26.00 2010 IEEE Optical Interconnects David A. B. Miller Stanford University http://ee.stanford.edu/~dabm David A. B. Miller, Stanford 1 Scales of connections Telecommunications Campus