IREE: An AI Subsystem for Zephyr?
summary
IREE runs on Zephyr POSIX with a new RISC‑V 64-bit BSP, serial driver, filesystem image and a run_module flow, packaged reproducibly with Guix. Benchmarks compile TF‑Lite/ONNX to VMFB, run unchanged on gem5 and an FPGA (AMD VC128), and calibration hits ~14% mean absolute error between simulator and FPGA. Models include MNIST, MobileNet and GPT‑2, with GPT‑2: ~40 minutes per character on FPGA vs ~4 hours on gem5 — inference works, FPGA gives usable speed.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 3:18AI-enabled mega cities
- 6:30Zephyr for AI workloads
- 9:59Porting IREE to Zephyr
- 14:23IREE compiler overview
- 17:58Hello world AI on Zephyr
- 20:08RISC-V for integration
- 23:28FPGA and Gem5 performance
- 25:39SPEC and ML workloads
- 27:52Benefits of integrated stack
- 30:37Guix for reproducible builds
- 34:15Building and running with Guix
- 35:41Summary
related talks
01h 19mMini-conference: RISC-V Support in Zephyr
10mAvoiding Amdahl's Law RISC-V Architecture Exploration for AI & ML Compute
07mLightning Talk: Zephyr Portability with an AI Application on Very Different MCUs
10mLightning Talk: Debugging Edge AI on Zephyr and Lessons Learned
13mKeynote: Taking RISC-V Intelligence Everywhere
19mAccelerating Edge AI with the Help of RISC-V
36mTinyML at the Edge: Deploying and Optimizing AI Workloads on Zephyr RTOS
19mDemocratizing Inference of Open-weight Models on RISC-V Manycore Acc...