Empower WebAssembly ML Over Diverse AI Hardware Accelerators Outside the Browser
summary
A runtime backend interposer hooks into TensorFlow/PyTorch/ONNX APIs, compiles loaded models with a TVM graph compiler (or vendor toolchains like TensorRT/OpenVINO) into a shared library, injects it into the process at model load, and runs inference from a preloaded graph runtime so existing WebAssembly ML binaries get GPU/FPGA/TPU acceleration with no code changes. Demo shows a CPU-only WASM image recognition app converted to run on GPU; compilation happens once at load and inference proceeds at native-accelerated speed.
AI-written summary · about
chapters (24)
- 0:00Introduction
- 1:04What is WebAssembly
- 2:28Wasm vs containers and VMs
- 3:30Running Wasm outside browser
- 4:33WASI
- 6:16Wasm ML frameworks in browser
- 8:45Browser ML limits
- 9:27Wasm NN and hardware gaps
- 10:33AI hardware challenges
- 12:20Graph compilers in ML
- 13:44TVM compilation
- 15:06ML booster subsystem
- 16:31System agent config
- 17:55Runtime interposing
- 19:42C++ and Python ML support
- 20:44Hardware accelerator compat
- 21:49Enabling Wasm ML on accelerators
- 23:34Live demo setup
- 24:48Demo results and GPU
- 27:42Future directions
- 28:02GPU sharing and efficiency
- 29:10Open source status
- 29:51Wasm NN vs proposal
- 31:15Q&A
related talks
08mLightning Talk: Accelerate WebAssembly AI with Backend Graph Compliers on a Wide Variety of AI HW Accelerators
37mGetting Started with AI and WebAssembly
39mWebAssembly-Powered Machine Learning
37mAI/ML at the Extreme Edge with WebAssembly: A Path Forward
33mWebAssembly based AI as a Service with Kubernetes
25mLeveraging Wasm for Portable AI Inference Across GPUs, CPUs, OS & Cloud-Nativ...
27mAI inference on the Edge cloud using WebAssembly
01h 42mWebAssembly: Writing Once and Running Everywhere NG