Pathfinder to GPU Offload in WASM
summary
Claims a feasible path to GPU offload for WASM by combining WNN model import and a WASI webGPU runtime; Rust code compiles to .wasm and runs with wasmEdge using an OpenVINO backend to execute a road segmentation model and emit a tensor image. Prototype compute backend uses wgpu and runs a vector-add on 2 million elements with 256-thread workgroups (1 then 1000 iterations) and a 16x16 matrix multiply with 4x4 groups on a remote GPU; demos complete and results validate the approach.
AI-written summary · about
chapters (17)
- 0:00Intro
- 0:15Agenda
- 0:45Motivation
- 1:20Native offload APIs
- 2:43Current offload APIs
- 3:33WASM offload APIs
- 5:24Demo
- 8:13Web GPU
- 9:41Web GPU vs WebGL
- 10:43WASM Web GPU
- 11:40Compute Pass
- 12:18WGPU vs WASM
- 14:20Demonstration
- 18:55Apache TVM
- 19:17Other APIs
- 20:42Summary
- 21:17Questions
related talks
30mWasi-Webgpu: Build Beautiful Graphics and Safely Run AI with Any GPU
33mBringing AI to the Heterogeneous Edge with WebAssembly & ONNX
31mRunning AI workloads with WASM - is it production ready? [Panel]
31mGPUs Unleashed! Make Your Games More Powerful With wasi-gfx
10mDeploying machine learning in wasm to CPU, GPU and NPU with the new wa...
24mContent Authenticity Initiative Trustmark With WASI WebGPU
37mGetting Started with AI and WebAssembly
14mWasm, WebGPU, & WebNN: How compute abstraction are enabling client side AI