Hugging Face's WebGPU kernels beat ORT by 2.57x

Hugging Face released @huggingface/kernels on September 1, according to the WebGPU kernels announcement. The library pulls versioned WebGPU kernels from the Hub and runs them straight in the browser.

For anyone running inference client-side, the speedups are concrete. Its 207 kernels beat ORT WebGPU by a 2.57x geometric mean and a 1.90x median across 809 comparable operations. Element-wise Add reached 3.52x, and LayerNormalization reached 2.22x.

Each kernel lives in its own repo on the Hub, with a manifest, correctness tests, and benchmarks attached. Developers can check what they're pulling in before it runs in production.

A companion tool called Fleet crowdsources those benchmark numbers across real consumer GPUs, not a single lab's test bench. WebGPU performance swings hard between an integrated laptop chip and a discrete desktop card.

Fleet's crowdsourced numbers matter more than the 2.57x figure. If they hold up, browser inference gets a real performance baseline instead of vendor demos. Hugging Face published the comparison on September 1, as more builders start treating in-browser inference as production infrastructure rather than a demo trick.