The Development of an Application-Specific Instruction Set Processor Specialized on a Convolutional Neural Network Trained on MNIST

This paper will present the development of an ASIP design and it’s toolchain for integrating a convolutional neural network in the ONNX format that works on the MNIST dataset. Additionally, it will present the results of the design’s deployment onto a Xilinx Zynq UltraScale+ MPSoC ZCU102 FPGA. As a toolchain, the LLVM umbrella project is being used. LLVM started as a compiler project, aiming for modularity but later became an umbrella project due to the large number of contributions and subprojects. The three main components are the ONNX-MLIR project, the LLVM main project (mainly the clangd, the optimizer and a specific RISC-V backend developed by us) and an HLS Simulator project called Comet, used to design the architecture. After a series of modifications, Comet was renamed to AIDA.