WebGPU AccelerationThe SGD optimization loop runs across thousands of GPU shader cores in parallel. Expect dramatic speedups on large datasets compared to CPU-only UMAP.
HNSW k-NNUses hnswlib-wasm for O(n log n) approximate nearest neighbor search — fast and accurate regardless of dataset size.
Transparent CPU FallbackWhen WebGPU is unavailable, umap-gpu silently falls back to an identical CPU implementation. Same API, same output, everywhere.