New landscape of compu)ng Personalized and targeted compu)ng

Transcription

1 Approxima)on Neal Anderson, Tom Conte, Hadi Esmaeilzadeh, Jennifer Hasler, Rakesh Kumar, Dick Lipton, Yung- Hsiang Lu, Kathryn Mckinley, Ravi Nair, Peter Petre, Abbas Rahimi, Mar)n Rinard 1

2 Why Approxima)on? Performance growth hits the energy wall Conven)onal compu)ng technologies are an)cipated to fall significantly short of historical trends and the projected demand Radical departures from conven)onal approaches are necessary New technologies are emerging that are variable There are emerging applica)on that can tolerate inaccuracy 2

3 New landscape of compu)ng Personalized and targeted compu)ng 3

4 Perpetual Compu)ng Mobile Compu)ng Cloud Compu)ng 4

5 REPEATABILITY Nondeterminis)c Determinis)c Exact Approximate PRECISION

6 Why approxima)on is beneficial? Lower layers work hard to expose a general, reliable, "precise" and (mostly) determiniscc interface. Application! Language! Compiler! ISA/Architecutre! Circuits! At a big cost in efficiency! Physics! 6

7 Avoiding overkill design Approximate Compu)ng Generality Applica)on Efficiency Performance Programming Language Compiler Architecutre Microarchitecture Cost Precision Reliability Determinism Cost Circuit Physical Device 7

8 Opportuni)es Contribute to prolonging CMOS scaling Embrace variability Improve Yield Expose parallelsim Enable new technologies that are intrinsically variable Small feature sizes (5nm, ) Memristors, PCM Magne)cs Chemical Photonic Analog 8

9 Opportuni)es Bridge non- von Neumann models with von Neumann models Allow interoperability Neuromorphic Increasingly, emerging applica)ons of interest are error resilient/ inherently approximate Op)miza)on (e.g., approximate SAT, analy)cs) Randomized algorithms Machine learning Pacern recogni)on Decision Making Sensory applica)ons Financial applica)ons Web search Robo)cs Augmented reality Cyber- physical systems 9

10 NPU accelera)on Source Codes Code 1 Code 2 Code 3 Code 4 Code 5 Code 6 Common Intermediate Representa)on Neural RepresentaCon + Accelera)on CPU NPU 10

11 NPU design alterna)ves CPU NPU CPU GPU FPGA Digital ASIC FPAA Analog ASIC (Speed: 2.8, (Speed: 2.3, Energy: 3.8, Energy: 3.0, Quality: 10% ) Quality: 10% ) (Speed: 3.7, Energy: 6.3, Quality: 10% ) 11

12 Challenges Abstrac)ons for algorithm design and programming Exposing the low level error in an abstract way to the higher level Exposing the knobs of the low level Algorithm to hardware transla)on Automa)on and design tools Conver)ng component- based approxima)on to end- to- end solu)on 12

13 Challenges Quality of results: understanding how to measure quality for each applica)on and what level is acceptable Modeling low- level component behavior and composing into overall behavior (compounding) Introspec)ve feedback loop for self- adap)ng systems: need both knobs for measurements and tuning Fallback strategies 13

14 Challenges Full- stack effort, but no layer wants to be the first to change but somebody has to be the first Start with a killer app that mo)vates effort at all layers simultaneously Start with a sufficiently universal program like SAT Evolu)onary path with augmen)ng current prac)ces Incorpora)ng imprecision into hardware design and synthesis process e.g., Design tools 14

15 Challenges Adop)on by a large body of programmers and designers Common problem with other energy efficient technique Users should be oblivious to the approxima)on 15

16 Next Steps Exploring sopware and algorithmic approxima)on Start collabora)ons between academia and semiconductor industry to begin prototyping approximate hardware Modeling variability and error Prize compe))ons that come with university funding, like Grand Challenges, e.g., Scene reconstruc)on and object recogni)on on the shroud Face/voice recogni)on on the shroud Energy, network, privacy 16

17 Informa)on Gaps Marke)ng input on the key applica)ons to start User requirements: what quality level is good enough for these applica)ons? Solu)ons for reusability across different opera)ng condi)ons How much tes)ng is enough valida)on that you ve met those requirements? How do errors from different components compound? Composability and interoperability 17

18 What is the new technical idea; why can we succeed now? Embracing approxima)on, holis)cally, in general- purpose compu)ng Relaxing the abstrac)on of near- perfect accuracy in general- purpose compu)ng/communica)on/storage Increasingly, emerging applica)ons of interest are error resilient Conven)onal CMOS does not seem to outpace other approaches The difference in cost between precise and approximate output is increasing 18

19 A fer)le ground for innova)on Processor Pareto.Fron0er Energy Performance 19