Find an appropriate FPGA device for the linear algebra accelerator chip

[amacgillivray]

At present, the most attractive option is the Zynq-7000 chip.

From my current understanding, we could install and run Linux on the processing system (a core onboard the same chip as our programmable logic). This process is described in tutorials such as this one. The processing system could therefore act as our server and run our daemon program to handle requests and manage our (presumably bare-metal) chip.

• Zynq-7000 SoC ZC706 Evaluation Board
• Zynq-7000 SoC ZC702 Evaluation Board
• Links to board variants
• Zynq-7000 SoC Product Advantages
• Xilinx Cost-Optimized Portfolio

Another option is Pynq. From what I have read and heard so far, the Pynq Z2 board's python/linux support makes it very easy to access the programmable logic. See Pynq Overlays and Pynq Framework & Image for a breakdown of what Pynq is and how it can simplify some tasks. This seems to be most interesting for those that prefer python to less-abstracted languages like C or C++.

Miscellaneous Links re: FPGA

• https://www.linkedin.com/pulse/booting-embedded-linux-zynq-7000-soc-from-jtag-using-xsct-d%C3%B6kme
  • https://old.reddit.com/r/FPGA/comments/cv9w8v/pynq_vs_zynq_performance/
  • https://xilinx-wiki.atlassian.net/wiki/spaces/A/pages/18842203/Unsupervised+AMP?showComments=true&showCommentArea=true
  • https://xilinx-wiki.atlassian.net/wiki/spaces/A/pages/1837006921/OpenAMP+Base+Hardware+Configurations
  • https://docs.xilinx.com/r/2021.1-English/ug1137-zynq-ultrascale-mpsoc-swdev/Arm-GNU-Tools (Zynq Ultrascale)
  • https://docs.xilinx.com/v/u/2019.2-English/ug1165-zynq-embedded-design-tutorial
  • https://support.xilinx.com/s/question/0D52E00006iHqOeSAK/artix7-vs-zynq7000-pl-resource-count?language=en_US
  • https://old.reddit.com/r/FPGA/comments/4cel5v/zynq_vs_artix_mcu/

Links re: linear algebra for ML

• https://web.stanford.edu/~surag/posts/alphazero.html
• https://www.google.com/search?channel=fs&client=ubuntu&q=tree+search+with+nn+evaluation
• https://proceedings.neurips.cc/paper/2017/file/d8e1344e27a5b08cdfd5d027d9b8d6de-Paper.pdf

[amacgillivray]

Decision: We will use a Zynq board, preferably the Zynq-7000 ZC706 or ZC702. We have chosen this SoC due to the dedicated ARM cores / processing subsystem that we can use for debugging / interfacing with our IP, and because we think that running our daemon program in Linux on the PS would make for an interesting, fully-integrated solution (where all a user needs on their own system is the API headers / library).