Building FPGA-based Machine Learning Accelerators in Python
summary
Demo shows a real ML accelerator built for an entry‑level Xilinx FPGA with materials under $200 using a Python toolchain: PyTorch → ONNX → Brevitas/FIN → Vivado, plus Migen and LiteX for SoC work. It compiles a quantized model to a bitfile, loads onto a Z1/Z2‑class board, and runs a sine‑wave predictor and an MNIST classifier that reaches about 93% accuracy and ~133,000 images/sec on a ~$130 board, so the pipeline actually works end‑to‑end.
AI-written summary · about
chapters (10)
- 0:00FPGA-based ML accelerators intro
- 3:34FPGAs vs ASICs
- 7:23Hardware vs software dev
- 13:21PYNQ board intro
- 16:51Sine wave predictor in Python
- 22:12MNIST classification demo
- 25:01LiteX and NightFury next steps
- 27:08Why FPGAs in 2024
- 28:54Open-source tools & communities
- 33:11Q&A on HLS & performance
related talks
31mCFU Playground
27mOpen source FPGA NVMe accelerator platform for BPF driven ML processing with Linux/...
01h 55mNeural Networks and Accelerator Co-Design
12mDemo: NVIDIA's use of Catapult HLS for Building ML Inference Accelerators
01h 11mMachine Learning Acceleration in an Embedded World
37mExtending the Python Ecosystem to Xilinx
11mLightning Talk: Accelerating Experimentation...
25mMachine Learning on Microcontrollers; Tiny, Useful and Fun