Approximate Computing Is Dead; Long Live Approximate Computing. Adrian Sampson Cornell

Transcription

1 Approximate Computing Is Dead; Long Live Approximate Computing Adrian Sampson Cornell

2 Hardware Programming Quality Domains

3 Hardware Programming No more approximate functional units. Quality Domains

4

5 Narrower bit widths are just as good or better [Barrois et al., DATE 2017] 1.8 #10-3 approximate adders from the literature better efficiency PDP (pj) ACA ETAIV RCAapx Type 1 RCAapx Type 2 RCAapx Type 3 Fixed-Point trunc. Fixed-Point round MSSIM : Power consumption of DCT in JPEG encoding better accuracy just varying the adder width

6 Hardware Programming No more approximate functional units. No more voltage overscaling. Quality Domains

7 Dual-voltage approximate CPU [ASPLOS 2012] replicated functional units Fetch Decode Reg Read Execute Memory WB Br. Predictor Instruction Cache Decoder Register File Integer FU Data Cache Register File ITLB FP FU DTLB dual-voltage SRAM arrays

8 fft imagefill jmeint lu mc raytracer smm sor zxing together registers multiplier FPU cache ALU (a)

9 Hardware Programming No more approximate functional units. No more voltage overscaling. In general, no more fine-grained approximate operations. Quality Domains

10 The Horowitz imbalance a name I made up for this talk [ISSCC 2014] 25pJ 6pJ Control 70 pj I-Cache Access Register File Access Add

11 Constraint-based programming for spatial architectures [Nowatzki et al., PLDI 2013] nodes (N) links (L) x + y routers (R) z

12 Hardware No more approximate functional units. No more voltage overscaling. Programming No more automatic approximability analysis. In general, no more fine-grained approximate operations. Quality Domains

13 EnerJ type qualifiers [PLDI int a =...; int p =...; p = a; a = p;

14 EnerJ type qualifiers [PLDI int a =...; Let s insert these automatically! int p =...;

15 Hardware No more approximate functional units. No more voltage overscaling. In general, no more fine-grained approximate operations. Programming No more automatic approximability analysis. No more generic unsound compiler transformations. Quality Domains

16 Loop perforation [Sidiroglou-Douskos et al., FSE 2011] for (int i = 0; i < max; i++) { // whatever } i += 2

17 Hardware No more approximate functional units. No more voltage overscaling. In general, no more fine-grained approximate operations. Programming No more automatic approximability analysis. No more generic unsound compiler transformations. Quality Domains No more weak statistical guarantees.

18 Traditional guarantee 8x f(x) is good

19 Statistical guarantee Pr [f(x) is good] T

20 Statistical guarantee, in reality Pr x D [f(x) is good] T anticipated input distribution

21 probability x

22 probability low quality high quality x

23 probability x

24 Adversarial distribution probability x

25

26 Hardware No more approximate functional units. No more voltage overscaling. In general, no more fine-grained approximate operations. Programming No more automatic approximability analysis. No more generic unsound compiler transformations. Quality No more weak statistical guarantees. Domains No more sadness about the imperfection of quality metrics.

27 Application Description Error metric FFT Mean entry difference SOR Scientific kernels from the Mean entry difference MonteCarlo Normalized difference SparseMatMult SciMark2 benchmark Mean normalized difference LU Mean entry difference ZXing Smartphone bar code decoder 1 if incorrect, 0 if correct jmonkeyengine Mobile/desktop game engine Fraction of correct decisions normalized to 0.5 ImageJ Raster image manipulation Mean pixel difference Raytracer 3D image renderer Mean pixel difference

28 Hardware No more approximate functional units. No more voltage overscaling. In general, no more fine-grained approximate operations. Programming No more automatic approximability analysis. No more generic unsound compiler transformations. Quality No more weak statistical guarantees. Domains No more sadness about the imperfection of quality metrics. No more benchmark-oriented research?

29 Application Description Error metric FFT Mean entry difference SOR Scientific kernels from the Mean entry difference MonteCarlo Normalized difference SparseMatMult SciMark2 benchmark Mean normalized difference LU Mean entry difference ZXing Smartphone bar code decoder 1 if incorrect, 0 if correct jmonkeyengine Mobile/desktop game engine Fraction of correct decisions normalized to 0.5 ImageJ Raster image manipulation Mean pixel difference Raytracer 3D image renderer Mean pixel difference

30 ImageNet annual competition 30% Winning Classification Top-1 Error 25% 20% 15% 10% 5% 0%

31 Real-time graphics

32 Hardware No more approximate functional units. No more voltage overscaling. In general, no more fine-grained approximate operations. Programming No more automatic approximability analysis. No more generic unsound compiler transformations. Quality No more weak statistical guarantees. Domains No more sadness about the imperfection of quality metrics. No more benchmark-oriented research?

33 Notes and links: