The Light at the End of the Wire. Dana Vantrease + HP Labs + Mikko Lipasti

Transcription

1 The Light at the End of the Wire Dana Vantrease + HP Labs + Mikko Lipasti 1

2 Goals of This Talk Why should we (architects) be interested in optics? How does on-chip optics work? What can we build with optics? 2

3 Wires 4

4 Wires Are Great, BUT Local Wires Global Wire 5

5 Wires Are Great, BUT Global Wire 6

6 Wires Are Great, BUT Local Wires Global Wire Wire Delay/Gate Delay Local (Scaled Length) Technology (micron) [Ho 2003] 7

7 Solution 1: Give Up: Avoid Global Communication Interconnects are doing just this! (Power4/5, Cell, Tilera ) Repercussion: Locality matters more than ever, affecting: Programming Scheduling 8

8 Solution 2: Don t Give Up: Alternative Technologies Electrical Transmission Lines Low Swing Wires Asynchronous Latched Wires Other Radio Frequency Optics 9

9 Transmission Lines Conventional Global RC Wire Voltage Voltage Driver On-chip Transmission Line Voltage Voltage Vt Receiver Distance Driver [Slide From Beckman and Wood MICRO-03] Vt Receiver Distance 10

10 Transmission Lines Are Great, BUT Low Bandwidth thick wires dielectric spacing 11

11 Radio Frequency data 1 mixer output buffer mixer low pass filter data 1 data 10 f 1. Transmission Line f 1. data 10 f 10 Signal Spectrum X10 Carriers f GHz120GHz 280GHz 300GHz f [Slide from Chang et al. HPCA 2008] 12

12 Radio Frequency Is Great, BUT Components still fairly large Suffers from electromagnetic interference (EMI) 13

13 Optics High Bandwidth Dense Wave Division Multiplexing (DWDM) Multiple (e.g. 64) wavelengths in a 0.5 μm wire Low Power Power is largely independent of length Low Latency ~1/3 the speed of light (~2 cm in one 5GHz clock) Flexible Repeater-less Routing within a Wire Not necessarily point A to point B 14

14 Optics On-Chip CMOS compatible process 3D Stacking Not ideal for short-hops dominated by opticalelectrical-optical (OEO) delay We Are Over Here (~1300 nm) 15

15 Optics Is Great, BUT Still a risky venture (at least 10 years out) Lots of problems to be solved: Thermal Issue Reliability of energy and components Fabrication Cost Professor Sohi: There s Lots of Research To Be Done! 16

16 From the Ground Up Processor Die 17

17 From the Ground Up Analog Die Processor Die 18

18 From the Ground Up Optics Die Analog Die Processor Die 19

19 From the Ground Up Optics Die Analog Die Processor Die 20

20 Communication Checklist Power Source Communication Medium Routing Devices 21

21 Laser Off-chip continuous power source Comb of 64 equally spaced wavelengths Constant power 22

22 Communication Checklist Power Source Communication Medium Routing Devices 23

23 Waveguides Optical Wire Core (Si) Carries light Cladding (SiO 2 ) [Intel] Contains light w/ its lower refractive index Total Internal Reflection Si (index=3.5) μm SiO 2 (index=1.5) 24

24 Communication Checklist Power Source Communication Medium Routing Devices 25

25 Splitter Passive Device Good for power distribution Splits a fraction of light across all wavelengths Splitter 100% 80% 20% 26

26 Ring Resonator Ring-shaped waveguide Wavelength specific filter Functions it can perform: Divert Inject Detect Covert behavior Can signal wavelength 10 Gbits/sec Waveguide Ring resonator Coupler 3-5 μm 27

27 Divert Tune ring s Index of Refraction Bring off-resonance ring into resonance slurp 28

28 Divert Tune ring s Index of Refraction Bring off-resonance ring into resonance slurp 29

29 Divert Tune ring s Index of Refraction Bring off-resonance ring into resonance slurp 30

30 Divert Tune ring s Index of Refraction Bring off-resonance ring into resonance slurp *Note: destructive read 31

31 Inject Couple ring to two waveguides 32

32 Inject Couple ring to two waveguides 33

33 Detect Germanium absorbs light Opto-electric conversion Germanium (Ge) Doped 34

34 Function Summary Ge Doped Splitter Split Off Divert Detect Inject 35

35 Power Estimates Per Ring Resonator: 22 μw for heating 91 μw while on-resonance 474 μw while on-resonance and modulating Per Waveguide: 11 mw for all 64 wavelengths to be detected by 64 detectors μwatts Ring Resonator Power Ring Heater On- Resonance Ring Operating Mode On- Resonance Modulating Ring 36

36 Communication Checklist Power Source Communication Medium Routing Devices 37

37 Let s Lay Some Wire 38

38 Simple Example I want to talk to P3 P0 P3 39

39 Simple Example I want to talk to P3 P0 P3 Detectors 40

40 Simple Example I want to talk to P3 coupling P0 Injectors all wavelengths P3 Detectors 41

41 Simple Example I want to talk to P3 coupling P0 Diverters P3 Detectors 42

42 Extending Simple Example Many-writer single-reader P3 s channel P1 P2 P0 P3 Diverters Detectors 44

43 Finally A Fully Connected Interconnect P0 P1 P2 r... w w w Channel 0 w r w w Channel 1 Pn-1 r w w w Channel w w w Channel n-1... r 45

44 Chip + Interconnect Layout Star Coupler Laser Optical Die [ISCA 2008] 46

45 Elsewhere in Photonic-land Circuit Switched Mesh [Columbia (Schacham, Bergman, Carloni)] Single-writer Multiple-reader Bus [Cornell (Kirman et al.)] North East P0 P1 Pn-1 West South May 18,

46 Beyond Shuffling Bits -- An Optical Barrier No processor may proceed past the barrier until all processors have reached the barrier Barrier Operation Setup Each processor s on-resonance Arrival Processor s ring off-resonance Release Light U-turns and processors detect it (the release) arrive! arrive! arrive! arrive! P0 P1 P2 P3 release! release! release! release! 48

47 How the Technologies Differ (Functionally) Power Wave Transmissions Routing Devices Covert Switches (aka Ring Resonators) Allows variable-point-to-variable-point communication w/o repeaters Wave Transmissions? Covert Switches? RC Wires X X Transmission Lines X RF X Optics 49

48 Goals of This Talk Why should we (architects) be interested in optics? Optics is a promising alternative to wires How does on-chip optics work? Lasers, waveguide, splitters, ring resonators What can we build with optics? Data Interconnects, barriers, 51

49 Thanks! 52