Sign in
← AI PulseAug 18, 2026

Wire · news · Single-source brief

NVIDIA cuML and cuVS 25.06 Introduce Multi-GPU UMAP for Large Datasets

NVIDIA cuML and NVIDIA cuVS 25.06 now support multi-GPU processing for Uniform Manifold Approximation and Projection (UMAP), enabling faster dimensionality reduction on datasets up to hundreds of gigabytes.

By Illumora Editorial

Source · Aug 18, 2026, 4:48 PM · On Illumora · Aug 18, 2026, 4:52 PM

Media from the primary source — shown here so you can stay on Illumora.

Rewritten from one allowlisted primary — not independent enterprise reporting. Lanes →

Brief drafted by Illumora’s editorial model from the linked primary source. Ops desk reviews flagged pieces. How we write →

Read the source →NVIDIA Developer Blog — Run Massive-Scale UMAP in Minutes Using Multiple GPUs—Without Losing Accuracy | NVIDIA Technical Blog
Save

NVIDIA has introduced multi-GPU support for Uniform Manifold Approximation and Projection (UMAP) within its cuML and cuVS 25.06 libraries. This enhancement targets the computationally intensive all-neighbors kNN graph construction, a critical step in the UMAP algorithm. The new capability allows for end-to-end scaling, significantly reducing training runtimes for datasets containing tens to hundreds of millions of vectors.

Key Points

  • NVIDIA cuML and NVIDIA cuVS 25.06 introduced multi-GPU support for UMAP's all-neighbors kNN graph construction.
  • This feature enables end-to-end scaling and reduces training runtimes for datasets with tens to hundreds of millions of vectors.
  • The multi-GPU implementation partitions datasets into balanced clusters, computes local kNN graphs independently, and merges them into a global graph.
  • This approach avoids all-to-all communication, preserving embedding quality at massive scales.
  • Empirical results showed up to a 74x speedup over projected CPU runtimes using eight NVIDIA H100 GPUs.
  • The method makes UMAP feasible for multi-hundred-gigabyte workloads, executing them in minutes.
  • Two hyperparameters, knn_n_clusters and knn_overlap_factor, control the trade-off between space, time, and quality.

Context

UMAP is a dimensionality reduction technique used for visualization and feature extraction in applications like exploratory data analysis, topic modeling, and single-cell analysis. According to the NVIDIA Developer Blog, iterative and exploratory workflows require UMAP to run repeatedly, and the cost increases substantially with larger datasets. The all-neighbors kNN graph construction is a critical and expensive step, especially as datasets scale.

Why It Matters

This update addresses a key bottleneck in UMAP's scalability, allowing researchers and data scientists to process significantly larger datasets more quickly. The ability to run UMAP on multi-hundred-gigabyte workloads in minutes, rather than hours or days, can accelerate iterative analysis and exploration, potentially leading to faster insights and model development.

What To Do

  • Review the NVIDIA cuML and NVIDIA cuVS 25.06 documentation for details on multi-GPU UMAP implementation.
  • Experiment with the knn_n_clusters and knn_overlap_factor hyperparameters to optimize for specific dataset sizes and quality requirements.
  • Compare the performance of multi-GPU UMAP against single-GPU or CPU-based implementations for large-scale dimensionality reduction tasks.
  • Explore the cuVS all-neighbors API for applications requiring standalone all-neighbors graph construction.