ONNX Runtime vs NeuronWriter

ONNX Runtime now moves on two tracks: the core runtime on its 1.2x cadence, and the WebGPU execution provider shipping independently as a plug-in. Core 1.28 upgraded to ONNX 1.22, made cuDNN and cuFFT optional at runtime to shrink the CUDA redistributable, and introduced an experimental C API surface. The WebGPU plug-in's second release is almost entirely attention performance: fused FlashAttention decode kernels for any sequence length, a generalized prefill path, and model-specific fusions for Qwen3 and Gemma 4.

Decoupling WebGPU from the core binary let it move on its own schedule, and it is spending that freedom on LLM inference specifically rather than broad operator coverage. The work keeps narrowing on what makes transformer decode fast in a browser — attention fusions, KV-shared decoder layers, per-device tuning down to specific Apple silicon. The core runtime is heading the other way, trimming dependencies and formalizing API surface instead of adding to it.

Expect the WebGPU plug-in to keep landing model-family fusions shortly after each new open-weight release, since the Qwen3 and Gemma 4 paths both arrived that way; whether the experimental C API stabilizes is not something these entries indicate.