DSENT A Tool Connecting Emerging Photonics with Electronics for Opto- Electronic Networks-on-Chip Modeling Chen Sun

Transcription

1 A Tool Connecting Emerging Photonics with Electronics for Opto- Electronic Networks-on-Chip Modeling Chen Sun In collaboration with: Chia-Hsin Owen Chen George Kurian Lan Wei Jason Miller Jurgen Michel Dimitri Antoniadis Lionel Kimerling Anant Agarwal Li-Shiuan Peh Vladimir Stojanovic MIT Microphotonics Center Fall 2012 Meeting 10/15/2012 1

2 Multi-Core Connectivity Challenge Every choice has a cost! 10/15/2012 2

3 Huge Potential for Photonics Photonics can tackle the multi-core connectivity problem Photonics core-to-core [Vantrease 08, Kurian 10] Photonics core-to-dram [Beamer 10, Udipi 11] Photonics-inspired architectures can improve performance over fully-electrical designs Device assumptions vary wildly from work to work Architectures presented in a 1-D fashion: No clear way for device/circuit designers to get feedback on their devices 10/15/2012 3

4 What is? Design Space Exploration of Networks Tool Software developed for power/area evaluation of interconnection networks Network Description Core Network Models Network Components Library Network Cost Metrics Technology Primitives Library Technology Parameters Technology Characterization Explore impact of photonics in the context of a bigger system 10/15/2012 4

5 as a Modeling Tool A cost evaluator, NOT a performance simulator A just give me a reasonable number mode Assumes some reasonable network activity Outputs network metrics like power, area, energy/bit, etc. Designed for easy integration with performance simulators that exist Technology Parameters Network Activity (optional) Technology node, optical parameters, etc. User input or from a performance simulator Network Power Network Network parameters, architecture Network Area 10/15/2012 5

6 Modeling Electronics Most of the network is routers and electrical links Rigorous electrical models in an ASIC-driven approach Models hierarchically made from basic technology parameters Technology Characterization Leakages Currents Capacitances Components Library Mesh Network Network Models Constructs Used in Technology Primitives Library Standard Cells Custom Cells Used in Delay/Timing Model Transition Density Model 10/15/2012 6

7 PAR PAR PAR PAR PAR PAR PAR Power (mw) PAR PAR PAR PAR PAR PAR Validating Electrical Models Default Parameters: 5x5 Mesh Router 2 Virtual Channels 64b Flit Width Vary # Bufs Vary # VCs Vary # Ports Vary Flit Width (bit) Input Buffer (Dynamic) Input Buffer (Leakage) Crossbar (Dynamic) Crossbar (Leakage) SA (Dynamic) SA (Leakage) Other (Dynamic) Other (Leakage) Power estimate within 20% of placed and routed router designs Accurate across a wide range of designs Successor to Orion 2.0 [Kahng, DATE 2009] for on-chip network cost modeling 10/15/2012 7

8 Modeling Photonics in Receivers/Modulator driver circuits optimization Based on previous work in [Georgas, CICC 2011] Backend circuitry leverages electrical framework Thermal tuning backend Serializers/Deserializers Clocking 10/15/2012 8

9 Link Optimization Example Designing a 128 Gb/s WDM link Data-rate is a degree of freedom 4 32 Gb/s, or 64 2 Gb/s Thermal tuning, SerDes, laser power are modeled 10/15/2012 9

10 Photonics Integration in Optical link, circuit, device models Based on previous work in [Georgas, CICC 2011] Plugs seamlessly into evaluation interface Enables evaluations of Electro-Optic Networks Technology Characterization Losses Laser Efficiency Electronic Components Routers Used in Network Models Constructs Technology Primitives Library Modulators Receivers Used in Optical Components Used in 10/15/

11 Applying to a Full System ATAC+ Architecture Appears in [Kurian, IPDPS 2012] 1024 Cores Divided into 64 optically connected clusters 16 electrically-connected cores per cluster, with an I/O optical hub Optical ring bus overlaid on top of an electrical mesh Supports efficient low-latency broadcasts Representative of how people think optics will be used in an on-chip system models the full network Performance simulation done using Graphite [Miller, HPCA 2010] 11

12 Bad Energy (?) Energy = Total energy consumed over the course of the application Averaged across all benchmark applications EMesh-Bcast electrical mesh network with optimized broadcast support ATAC+ network uses more energy to finish a benchmark Though with 30% average lower runtime than the best fully electrical network Ring tuning and laser power are significant 12

13 Why are Laser and Tuning Bad? Static vs dynamic power Non-data-dependent (NDD) vs data-dependent (DD) power Leakage and ungated clocks are NDD Transistor switching is DD Off-chip laser, cannot be throttled quickly, NDD thermal tuning, cannot tune rings in and out quickly, NDD Network Peak Saturation Throughput = 0.15 flits/cycle/core Network sized to handle peak throughput Average utilization <20% 80% of the network is idle! Benchmark Applications 10/15/

14 Energy (Normalized) Energy (Normalized) Wish List Item #1: Throttled Lasers Laser Other Laser Other Standard Laser Throttled Laser Idle and unicast modes through laser throttling Without throttling, laser must be at full broadcast-capable strength all the time Laser energy drops to 2% of total (cache + network) Throttle-able on-chip lasers are extremely valuable 14

15 Energy (Normalized) Energy (Normalized) Wish List Item #2: No Ring Tuning Laser Ring Tuning Other Laser Other Thermally Tuned Athermal Rings Thermal tuning 35% of total energy Athermal rings that do not require thermal tuning are extremely important 15

16 Comparing Energy ATAC+ (Cons) Standard laser Needs tuning ATAC+ Throttled Laser, Athermal rings ATAC+ (Ideal) 100% efficient laser lossless devices Significant energy savings if wish-list items are available Wishlist ATAC+ achieves energy efficiency that is close to the ideal Using E-D product metric, ATAC+ wins by >80% 16

17 Conclusion We created to bridge photonics and electronics Generalized methodology for digital components Allows architectural explorations of photonic interconnects We showed a -enabled study of a representative photonic network architecture Problem of non-data-dependent laser and thermal tuning power due to low average utilizations Identified throttle-able on-chip lasers and athermal ring resonators as key for energy-efficient photonics Website (With Download Link): 10/15/

18 Backups 10/15/

19 Framework After initial modeling of implementation, design can be optimized and evaluated 10/15/2012 [Georgas, CICC 2011] 19

20 Effect of Utilization Non-data-dependent energy dominant Low Throughput Data-Dependent energy dominant Max Throughput 10/15/

21 Evaluation Infrastructure Cache Models Core Models Benchmark Network Models Inputs Cache & Directory Counters Electrical Technology Parameters Graphite NM Optical Link Counters Electrical Router & Link Counters Optical Technology Parameters McPAT Tools Completion Time Cache Energy & Area Electrical Router & Link Energy & Area Optical Link Energy & Area 10/15/ Outputs

22 Great Performance! Benchmarks ATAC+ demonstrates significant performance advantage 40% over broadcast-assisted electrical baseline (EMesh-Bcast) 130% over standard electrical baseline (EMesh-Pure) EMesh-BCast and EMesh-Pure differ according to the frequency of read-write sharing 22

23 Comparing E-D Product ATAC+(Ideal) ATAC+ ATAC+(ThrottledOnly) ATAC+(Cons) EMesh-BCast EMesh-Pure ATAC+(Cons) = Throttled laser, Needs tuning ATAC+(ThrottleOnly) = Throttled laser, Needs tuning ATAC+ = Throttled Laser, Athermal rings ATAC+(Ideal) = 100% efficient laser, lossless devices With the wishlist items, ATAC+ achieves % improvement over best electrical baseline (EMesh-BCast) Even ATAC+(Cons) can win over best electrical baseline in some benchmarks 23

24 Energy (Normalized) Core NDD Energy Depends On Network Performance Core DD Core NDD Caches Network ATAC+ EMesh -BCast 0 radix fmm ocean ocean average contig non-contig Higher performance network with higher energy may give better system energy benefits Correct trade-off only apparent with a full-system evaluation 24

25 Energy (Normalized) Energy (Normalized) System Energy Shows Different Trends Than Network Energy ATAC+ conservative variants worse ATAC+ conservative variants better Network Network Caches Core NETWORK SYSTEM 25

26 Technology Sensitivity Tuned Rings, Standard Laser, 30% Laser Efficiency, 0.2 db Waveguide Loss ATAC+ (Cons) 100% Laser Efficiency, 0dB Waveguide Loss ATAC+ (Ideal) ATAC+ (ThrottledOnly) ATAC+ Laser Throttling Athermal Rings 26