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

Accelerating Massive-Scale Data Visualization

NVIDIA has introduced multi-GPU support for the Uniform Manifold Approximation and Projection (UMAP) algorithm in its cuML and cuVS 25.06 libraries, enabling a significant reduction in processing time for massive-scale datasets. The update allows data scientists to run UMAP on vector datasets as large as 870 GB in just eight minutes using eight NVIDIA H100 GPUs. This development is important for workflows in exploratory data analysis and topic modeling, where iterative analysis on datasets with tens or hundreds of millions of vectors was previously hindered by prohibitive runtimes that could stretch into hours or days.

The new capability works by distributing the most computationally intensive part of the UMAP algorithm—the all-neighbors k-nearest neighbors (kNN) graph construction—across multiple GPUs. The implementation partitions the dataset into balanced clusters, allowing each GPU to independently compute a local kNN graph. These local graphs are then merged to form the global graph, an approach that cleverly avoids the all-to-all communication that often limits scalability in distributed systems. Users can manage the trade-off between performance, memory usage, and embedding quality through two key hyperparameters: `knn_n_clusters`, which determines the number of data partitions, and `knn_overlap_factor`, which controls the overlap between clusters to preserve neighbor relationships.

• Software: NVIDIA cuML and NVIDIA cuVS version 25.06
• Core Feature: Multi-GPU support for all-neighbors kNN graph construction in UMAP.
• Performance: Up to 74x speedup over projected CPU runtimes; processes 870 GB of vectors in 8 minutes on eight NVIDIA H100 GPUs.
• Method: Partitions data into balanced clusters for independent, parallel processing on each GPU, avoiding all-to-all communication.
• Key Controls: `knn_n_clusters` and `knn_overlap_factor` hyperparameters for tuning speed, memory, and quality.

This enhancement effectively transforms UMAP from a time-consuming batch process into a more interactive tool for very large datasets. By making multi-hundred-gigabyte workloads feasible in minutes, NVIDIA lowers the operational barrier for researchers and data scientists who need to visualize and extract features from high-dimensional vector spaces. This directly impacts the practicality of iterative analysis in fields like single-cell biology and natural language processing, where dataset sizes are consistently growing and rapid insight is a competitive advantage. The ability to use precomputed kNN graphs also offers flexibility, integrating directly with other tools in the data science ecosystem.

By parallelizing the primary bottleneck in UMAP, NVIDIA is not just offering an incremental speedup; it is changing the cost-benefit analysis for data scientists working with massive datasets. This makes complex, iterative vector analysis more accessible on standard multi-GPU systems, further cementing the CUDA ecosystem's role in large-scale data science beyond just model training.