Hugging Face released 207 WebGPU kernels that run 2.57x faster than ORT WebGPU by geometric mean
Hugging Face Blog·medium signal
@huggingface/kernels, published September 1, is a JavaScript library that pulls versioned WebGPU kernels from the Hub and runs them in the browser. Across 809 comparable operations it measured 2.57x faster by geometric mean and 1.90x at the median against ORT WebGPU, with Add at 3.52x and LayerNormalization at 2.22x. Each kernel is its own repo with manifests, correctness tests and benchmarks, and a companion browser tool called Fleet crowdsources performance data across real consumer GPUs.