Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

VTA 試してみた

887 views

Published on

This slide was presented in Halide 勉強会 at Fixstars Corporation, 2018/7/28.

Published in: Technology
  • Be the first to comment

VTA 試してみた

  1. 1. VTA: An Open, Customizable Deep Learning Acceleration Stack 試してみた TAKATO YAMADA © 2018 Halide 勉強会
  2. 2. ? FPGA © 2018 Halide 勉強会 @tkclimb
  3. 3. © 2018 Halide 勉強会 …. VTA https://tvm.ai/2018/07/12/vta-release-announcement.html TVM FPGA HW
  4. 4. VTA Deep learning DL (FPGA ) © 2018 Halide 勉強会 https://arxiv.org/abs/1807.04188 VTA
  5. 5. VTA JIT TVM RPC NN HW Д © 2018 Halide 勉強会 https://arxiv.org/abs/1807.04188 https://twitter.com/tqchenml/status/1017469205011054592
  6. 6. © 2018 Halide 勉強会 https://arxiv.org/abs/1807.04188 TVM VTA HW
  7. 7. Resnet18 © 2018 Halide 勉強会 https://docs.tvm.ai/vta/install.html
  8. 8. © 2018 Halide 勉強会 1) Pynq RPC 2) PC FPGA 3) I. PC II. Pynq VTA
  9. 9. 1. RPC 2. FPGA 3. VTA BN FPGA 4. FPGA (max 27GOPS) © 2018 Halide 勉強会 https://arxiv.org/abs/1807.04188
  10. 10. VTA Pynq FPGA © 2018 Halide 勉強会 https://arxiv.org/abs/1807.04188 https://tvm.ai/2018/07/12/vta-release-announcement.html
  11. 11. © 2018 Halide 勉強会 https://arxiv.org/abs/1712.08934
  12. 12. © 2018 Halide 勉強会 https://arxiv.org/abs/1712.08934 Resnet18 51GOPS ( 27GOPS) VTA TVM VTA DNN FPGA
  13. 13. VTA 4 • Fetch • Load • Store • Compute 2 • Tensor ALU • GEMM Core © 2018 Halide 勉強会 https://arxiv.org/abs/1807.04188
  14. 14. VTA © 2018 Halide 勉強会 https://arxiv.org/abs/1807.04188 • LOAD DRAM 2D • STORE DRAM 2D • GEMM - • ALU
  15. 15. VTA © 2018 Halide 勉強会 https://arxiv.org/abs/1807.04188 • LOAD DRAM 2D • STORE DRAM 2D • GEMM - • ALU
  16. 16. VTA © 2018 Halide 勉強会 https://arxiv.org/abs/1807.04188 VTA Compute Load Store WAR
  17. 17. VTA © 2018 Halide 勉強会 https://docs.tvm.ai/vta/tutorials/matrix_multiply.html#sphx-glr-vta-tutorials-matrix-multiply-py (1, 256) x (256, 256)
  18. 18. VTA © 2018 Halide 勉強会 https://docs.tvm.ai/vta/tutorials/matrix_multiply.html#sphx-glr-vta-tutorials-matrix-multiply-py FPGA 32bits -> 8bits
  19. 19. VTA © 2018 Halide 勉強会 https://docs.tvm.ai/vta/tutorials/matrix_multiply.html#sphx-glr-vta-tutorials-matrix-multiply-py compute_at
  20. 20. VTA © 2018 Halide 勉強会 https://docs.tvm.ai/vta/tutorials/matrix_multiply.html#sphx-glr-vta-tutorials-matrix-multiply-py Tensorize VTA
  21. 21. © 2018 Halide 勉強会 (‘A`)
  22. 22. © 2018 Halide 勉強会 : (256 x 256) * 2 = 131072 Ops : 0.0441 / 10 = 0.00441 : 29.7 MOPS Python Timeit
  23. 23. © 2018 Halide 勉強会 1. 1. TVM FPGA 2. 2. 1. FPGA ( ) 2. 32bit( ) ? 3. ( )
  24. 24. © 2018 Halide 勉強会 https://arxiv.org/abs/1807.04188
  25. 25. © 2018 Halide 勉強会 https://arxiv.org/abs/1807.04188
  26. 26. VTA 3 • Fetch • Load • Store • Compute 2 • Tensor ALU • GEMM Core © 2018 Halide 勉強会 https://arxiv.org/abs/1807.04188

×