NVIDIA cuML 25.06 Speeds Up UMAP Training on Multiple GPUs

NVIDIA has released version 25.06 of its cuML and cuVS libraries, introducing multi-GPU support for the all-neighbors k-nearest neighbors (kNN) graph construction in the Uniform Manifold Approximation and Projection (UMAP) algorithm. While previous versions restricted UMAP training to a single GPU, this update allows practitioners to distribute the most computationally expensive phase of the algorithm across multiple graphics cards. The development enables end-to-end scaling for datasets containing tens to hundreds of millions of vectors, drastically reducing training runtimes.

The multi-GPU implementation works by partitioning datasets into balanced clusters, computing local kNN graphs independently on each GPU, and merging them into a global all-neighbors graph. This method eliminates the need for expensive all-to-all communication between GPUs. Users can control the trade-offs between processing speed, memory usage, and embedding quality using two main hyperparameters: knn_n_clusters and knn_overlap_factor. For instance, when processing 100 million float32 vectors of 1024 dimensions on 80 GB GPUs, setting the overlap factor to two and the cluster count to 24 keeps the local data size per cluster at roughly 34 GB, leaving plenty of memory headroom.

In empirical tests, NVIDIA evaluated the new system on the Wiki and MIRACL datasets using an NVIDIA DGX system equipped with eight NVIDIA H100 GPUs and an Intel Xeon 8480CL 224-core CPU with 2 TiB of RAM. On the MIRACL dataset, which contains 106 million vectors, the multi-GPU setup executed UMAP over 870 GB of vectors in just eight minutes. This represents a 74x speedup compared to projected runtimes for traditional CPU-based implementations. At this scale, the standard CPU reference implementation failed to complete the task entirely due to excessive memory consumption, despite the system's 2 TiB of RAM.

This performance leap makes UMAP practical for massive-scale visualization, topic modeling, and single-cell analysis workflows that were previously computationally intractable. By accelerating the training phase without sacrificing the trustworthiness scores of the resulting embeddings, the update allows machine learning practitioners to run iterative, exploratory data analysis on massive datasets in minutes rather than hours or days.