Run Massive-Scale UMAP in Minutes Using Multiple GPUs—Without Losing Accuracy | NVIDIA Technical Blog

This post details the introduction of multi-GPU support for the all-neighbors kNN graph construction step in UMAP within NVIDIA cuML and cuVS 25.06. It explains how the dataset is partitioned into balanced clusters, local kNN graphs are computed independently per cluster, and then merged to form a global all-neighbors graph. This approach avoids all-to-all communication, enabling end-to-end scaling for datasets with tens to hundreds of millions of vectors. It also introduces and explains the `knn_n_clusters` and `knn_overlap_factor` hyperparameters for managing the space, time, and quality trade-offs, and provides a Python example using the cuVS all-neighbors API.