Y Soft Technology Hour - Ultimate Commodity Supercomputer

Presentation by Jiri Matela and Martin Jirman - Comprimato.

Source code to this presentation can be found at https://github.com/ysoftdevs/th/tree/master/2014_06_19_Comprimato_Ultimate_Commodity_supercomputer

1. GPU:  The  Ul*mate  Commodity   Supercomputer     Jiří  Matela   &   Mar-n  Jirman  
3. The  Evolu-on  of  Compu-ng   Intel  ASCI  Red     1  TFLOPS     7904  CPUs     850  KW     150  m2  
4. The  Evolu-on  of  Compu-ng   NVIDIA  GeForce     5.1  TFLOPS     250W     296  cm2  
5. The  Evolu-on  of  Compu-ng   Mobile  GPU     0.36  TFLOPS     5W     0.14  cm2  
6. Ar-fical  Brain  –  Neural  Network  –  Deep  learning     *  Source  NVIDIA  
7. Where  GPUs  Shine   Neural  networks  (Ne_lix)     GPU  accelerated  database  query  (PgOpenCL)     Physics  (Games)     Ray  tracing  (FurryBall,  NVIDIA  Op-X)     Linear  Algebra  (CUBLAS)     Video  Compression      
8. General-­‐purpose  compu-ng  on   graphics  processing  units  (GPU)   •  Video  Controller  -­‐>  GPU  (nvidia)   -­‐>  GPGPU   •  Shaders  Languages  -­‐>  CUDA    /  OpenCL   •  CUDA   – Compu-ng  architecture   – Programming  language  
9. CUDA  Quick  Start  SLIDE   •  NVIDIA  GPU  –  GeForce  (mobile),  Tesla,  Quadro     •  Win  /  Lin  /  Mac   •  NVIDIA  Driver   •  NVIDIA  Installer*   –  Toolkit   –  Samples   –  Tools   *  developer.nvidia.com/cuda-­‐downloads  
10. CUDA  Architecture  and  Programming   model     1st  example  –  Vector  Addi-on  
11. Thread  Hierarchy  
12. GPU  as  Mul-core  SIMD  
13. Memory  Hierarchy  
14. Gaussian  Image  Blur  
15. CPU  Basic  Single  Thread  Implementa-on       GPU  Basic  Implementa-on   (1pixel  =  1thread)  
16.   CPU  Parallel  Using  OpenMP  
17.   GPU  using  Shared  Memory  
18.   GPU  –  Overlapping  Transfers   and  Computa-ons  
19. GPU  –  Final     1  thread  =  mul-ple  pixels   private  array  (registers)   #pragma  unroll  
20. Conclusion   •  Gaussian  blur   •  CPU  160  -­‐>  100  ms   – Core  i5  –  4  Cores   •  GPU  16  -­‐>  3ms   – GeFroce  740m   – 2  SM   – 368  cuda  cores  
21. Thank  you!   Jiří  Matela  &  Mar-n  Jirman