WebNN backend for NPU acceleration

[sauravpanda]

Context

WebNN ships in Chrome 146. Provides access to NPUs, GPUs, and CPUs through a graph-based computation model. On devices with NPUs (Apple M-series, Intel Meteor Lake, Qualcomm), WebNN could be significantly faster than WebGPU.

Proposal

• Implement WebNN as a third compute backend alongside WebGPU and WASM SIMD
• Feature-detect at runtime: NPU > GPU > CPU
• Map core operations (matmul, softmax, rmsnorm) to WebNN graph ops
• Potentially much faster on NPU-equipped devices

References

• WebNN Spec
• ONNX Runtime Web already supports WebNN execution provider