1. Computational Sprinting on a Hardware/Software Testbed Arun Raghavan *, Laurel Emurian *, Lei Shao #, Marios Papaefthymiou +, Kevin P. Pipe +#, Thomas F. Wenisch +, Milo M. K. Martin * University of Pennsylvania, Computer and Information Science * University of Michigan, Electrical Eng. and Computer Science + University of Michigan, Mechanical Engineering #

2. This work licensed under the Creative Commons Attribution-Share Alike 3.0 United States License You are free: to Share — to copy, distribute, display, and perform the work to Remix — to make derivative works Under the following conditions: Attribution. You must attribute the work in the manner specified by the author or licensor (but not in any way that suggests that they endorse you or your use of the work). Share Alike. If you alter, transform, or build upon this work, you may distribute the resulting work only under the same, similar or a compatible license. For any reuse or distribution, you must make clear to others the license terms of this work. The best way to do this is with a link to: http://creativecommons.org/licenses/by-sa/3.0/us/ Any of the above conditions can be waived if you get permission from the copyright holder. Apart from the remix rights granted under this license, nothing in this license impairs or restricts the author's moral rights. 2

3. Overview 3 Computational sprinting [HPCA’12] Targets responsiveness in thermally constrained environments Far exceed sustainable power for short bursts of computation Simulation based feasibility study This work: what can we learn with today’s hardware? Engineer hardware/software testbed for sprinting Reduce heat venting capacity Sustain only lowest power mode Can sprint on today’s system Longer with phase-change material Sprinting improves energy-efficiency Even for sustained computations

4. 4 T max power temperature Computational Sprinting Using Dark Silicon [HPCA’12]

5. 5 T max power temperature Effect of thermal capacitance Computational Sprinting Using Dark Silicon [HPCA’12]

6. 6 T max power temperature Effect of thermal capacitance Computational Sprinting Using Dark Silicon [HPCA’12]

7. 7 T max power temperature Effect of thermal capacitance Computational Sprinting Using Dark Silicon [HPCA’12]

8. 8 T max power temperature State of the art: Turbo Boost 2.0 exceeds sustainable power with DVFS (~25%) Our goal: 10x Effect of thermal capacitance Computational Sprinting Using Dark Silicon [HPCA’12]

9. Evaluating Sprinting Simulation-based feasibility study [HPCA’12] Thermal models: buffer heat using thermal capacitance Electrical models: stabilize voltage with gradual core activation Architectural models: Large responsiveness improvements Little dynamic energy overheads Next steps: understanding sprinting on a real system Build a real chip? Sprint on today’s mobile chips? 9 Our approach: study sprinting on hardware available today

10. This Work: Testbed for Computational Sprinting How long can the testbed sprint? How to select sprint intensity? How can we extend sprint duration? How does sprinting impact energy? 10

11. Designing a testbed for sprinting 11

12. sprinting sustainable Quad-core Intel i7-2600 With heatsink and fan: 95W 12 Remove heatsink, slow fan; 10W thermal design (TDP) CoresFreq.Power Normalized Power Peak Speedup 1 core1.6 GHz10 W1x 4 cores1.6 GHz20 W~2x4x 4 cores3.2 GHz50 W~5x8x 3 operating modes:

13. Sprinting Performance sobel disparity segment kmeans feature texture Cores + Frequency (3.2GHz): 6.3x speedup Cores only (1.6 GHz): 3.5x speedup Max 4 core, 3.2 GHz Max 4 core, 1.6 GHz Baseline (no sprint) 3.2GHz 1.6GHz 13

14. How long can the testbed sprint? 14

15. Testbed Thermal Response 40 50 60 70 0 20 40 60 Power (W) time (s) Temp (°C) time (s) T max 15 0 20 40 60 40 50 60 70 sustained

16. Testbed Thermal Response 40 50 60 70 0 20 40 60 Power (W) time (s) Temp (°C) time (s) 16 5x 3s 0 20 40 60 40 50 60 70 sustained sprint (3.2 GHz) sustained sprint (3.2 GHz) 20g copper  Δ 25 o C, ~188J  3.5s @ 50W Heat spreader

17. Testbed Thermal Response 40 50 60 70 0 20 40 60 Power (W) time (s) Temp (°C) time (s) 17 5x 3s 0 20 40 60 40 50 60 70 21s sustained sprint (3.2 GHz) sprint (1.6 GHz) sustained sprint (3.2 GHz) sprint (1.6 GHz) 2x 20g copper  Δ 25 o C, ~188J  3.5s @ 50W Heat spreader

18. What if computation doesn’t complete during the sprint? 18

19. Truncated Sprint Performance 19 computation length normalized speedup Little benefit Near-peak performance for short tasks

20. Truncated Sprint Performance computation length normalized speedup 20 Little benefit Near-peak performance for short tasks Lower peak performance; benefits longer tasks Best sprint intensity depends on task size How to sprint when task size is unknown? Sprint pacing Max intensity sprint for half thermal capacitance Cores-only sprinting for other half

21. Truncated Sprint Performance computation length normalized speedup 21 Best sprint intensity depends on task size How to sprint when task size is unknown? Sprint pacing Max intensity sprint for half thermal capacitance Cores-only sprinting for other half

22. Truncated Sprint Performance 22 Best sprint intensity depends on task size How to sprint when task size is unknown? Sprint pacing Max intensity sprint for half thermal capacitance Cores-only sprinting for other half computation length normalized speedup

23. Increasing sprint duration 23

24. Two Ways of Adding Thermal Capacitance Specific heat capacity: introduce thermal mass Latent heat: absorb heat to change phase (e.g. melting) 24 temperature time (s) Phase-change absorbs heat while holding temperature constant Baseline sprinting More specific heat  takes longer to heat 20g copper, Δ 25 o C  ~188J 1g of wax  ~200J

25. 25 Computational Sprinting on a Hardware- Software Testbed Extending Sprints with Phase Change Material 4g of wax, 1g of aluminum foam Copper shim

26. Impact of Phase Change temperature (°C) 26 air time (s)

27. temperature (°C) 27 Small extension from heat capacity of encasement… foam air time (s) Impact of Phase Change

28. temperature (°C) 28 Small extension from heat capacity of encasement… …6x increase in sprint duration with phase change foam wax air time (s) Impact of Phase Change

29. temperature (°C) 29 Small extension from heat capacity of encasement… …6x increase in sprint duration with phase change due to phase change foam air water time (s) Impact of Phase Change wax

30. 30 Phase Change Material in Action Time lapse 15x

31. How does sprinting impact energy? 31

32. Energy Impact of Sprinting sobel disparity segment kmeans feature texture 32 3.2GHz 1.6GHz Race-to-idle: 7% energy savings! Sprint 3.2GHz Sprint 1.6GHz Idle

33. If sprinting is more energy efficient, why not sprint all the time… 33

34. Sprint-and-Rest seconds power (W) seconds temperature ( o C) seconds cumulative work sprint-and-rest 34 35% faster 5s @ 20W sprint, 12s @ 5W rest < 10W average

35. Conclusions Testbed confirms sprinting improves responsiveness Sprint pacing can extend benefits of sprinting Exploiting phase change allows longer sprints Sprinting can save energy in thermally limited systems 35

36. 36 www.cis.upenn.edu/acg/sprinting/