Run Massive-Scale UMAP in Minutes Using Multiple GPUs—Without Losing Accuracy

NVIDIA reports processing 106 million vectors with UMAP in eight minutes on eight H100 GPUs. UMAP reduces complex numerical data to a smaller representation useful for visualization. The implementation divides data into overlapping groups, finds nearby points within each group on separate GPUs, then combines those relationships. Technical article

This makes the expensive neighborhood-search stage distributable without fitting the entire dataset into one GPU. The reported 74-fold speedup uses an extrapolated CPU runtime because the full CPU run failed to complete. Increasing overlap can preserve more neighboring relationships, but requires additional computation and memory.