Empower ML Framework on RISC-V by Backend Graph Compilers
summary
A transparent backend interposer intercepts ML framework calls (TensorFlow/TensorFlow Lite/PyTorch) and automatically forwards models and predict calls to graph compilers such as TVM backend, removing model conversion and code changes. The system includes a node agent, controller for selecting compilers, and a Jupyter kernel extension demo; an unchanged image‑recognition script runs on RISC‑V while inference executes on the graph compiler, showing measurable inference speedup. Designed to accept vendor or open‑source graph compilers as drop‑in backends.
AI-written summary · about
chapters (8)
- 0:00Introduction and motivation
- 1:03Challenges for ML on RISC-V
- 4:19Proposed solution overview
- 7:37Graph compiler integration
- 11:55Demo with TensorFlow Lite
- 13:42Demo with Jupyter extension
- 15:09Summary and future work
- 17:18Q&A
related talks
23mEmpower Upstream ML Frameworks on RISC-V
17mExploring Static Code Generation and SIMD-Acceleration for Machine Learning on RISC-V
34mBridging the Gap: Compiling and Optimizing Triton Kernels Onto RISC-V Targets Based on MLIR
32mEmpower WebAssembly ML Over Diverse AI Hardware Accelerators Outside the Browser
40mHands-on TinyML Deployment on RISC-V: Building an Optimized Edge AI Pipeline
10mLightning Talk: Software Development for ML and RISC-V Vector Acceler...
10mSupport TVM QNN Flow on RISC-V with SIMD Computation
13mZML: ML framework for Zig