WebGPU vs WebAssembly for running models in a browser

WebAssembly runs ONNX models on the CPU and is supported everywhere; WebGPU hands the same graph to your graphics card and is typically 5–20× faster on transformer workloads, with the gap widening as the model grows. WASM still wins on tiny models where kernel launch overhead dominates, on browsers with no WebGPU, and when you need multi-threading via SharedArrayBuffer. This page lays out the trade-offs — precision support, buffer size limits, load time, and the fp16 trap — and explains how this playground chooses a backend automatically.

webgpu vs wasmwebgpu inference speedonnx runtime web backendbrowser ml performancewasm simd threads aiwebgpu browser supportclient side inference benchmark