CFU Playground
summary
A practical pipeline for building FPGA ML accelerators with an open source toolchain that actually runs on hardware and speeds up real models: it uses nMigen and LiteX, extends RISC-V with a custom function unit, and patches TensorFlow inference to call the CFU. The worked example drops full-model evaluation from 147 million → 8 million cycles and 2D convolutions from 95 million → 33 million cycles, implemented in ~20 lines Python plus ~15–20 lines C++.
AI-written summary · about
chapters (4)
related talks
30mCFU Playground: Model-specific Acceleration on FPGAs
34mBuilding FPGA-based Machine Learning Accelerators in Python
27mOpen source FPGA NVMe accelerator platform for BPF driven ML processing with Linux/...
12mDemo: NVIDIA's use of Catapult HLS for Building ML Inference Accelerators
28mAn Efficient Implementation of TensorFlow Lite for RISC-V Vectors
40mHands-on TinyML Deployment on RISC-V: Building an Optimized Edge AI Pipeline - Corrected Audio
19mBuilding Tool Chains for RISC-V AI Accelerators
11mLightning Talk: Accelerating Experimentation...