WebAssembly-Powered Machine Learning
summary
Ports of C/C++ ML code and TensorFlow AOT into WebAssembly are evaluated for in-browser machine learning using Emscripten and wasm builds of C++ libraries. Results show concrete wins (e.g., a face+document pipeline improving from 45 FPS to 47–48 FPS with reduced frame jitter, and training microbenchmarks achieving 13× and 7× speedups for specific models) but also clear limits: SIMD, threads, and GPU access are not yet stable across browsers, so performance depends on model and workload.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 1:49Machine learning for frontend
- 10:26Browser ML challenges
- 18:13WebAssembly performance
- 23:05ML in browser with WASM
- 27:53Moving algorithm to WASM
- 33:44Outlook and future
- 35:10Q&A
related talks
42mWebAssembly? Web and Assembler? WTF!
32mEmpower WebAssembly ML Over Diverse AI Hardware Accelerators Outside the Browser
29mWebAssembly - the Final Frontier
19mML at Client Side
37mAI/ML at the Extreme Edge with WebAssembly: A Path Forward
16mWhy WebAssembly
33mWebAssembly based AI as a Service with Kubernetes
01h 42mWebAssembly: Writing Once and Running Everywhere NG