---
title: "Run Massive-Scale UMAP in Minutes Using Multiple GPUs—Without Losing Accuracy"
description: "UMAP workflows on tens to hundreds of millions of vectors are constrained by the cost of constructing an all-neighbors k-nearest-neighbor graph, especially when repeated runs support exploration and parameter tuning. The NVIDIA cuML and cuVS 25.06 feature distributes this graph-construction stage across multiple GPUs while retaining the out-of-core design that partitions data into balanced clusters, overlaps vectors across nearby clusters, computes local graphs, and merges them into a global graph. knn_n_clusters and knn_overlap_factor expose the memory, runtime, and graph-quality trade-off: more clusters reduce per-GPU data, while more overlap can improve nearest-neighbor recall at higher computational and memory cost. On eight NVIDIA H100 GPUs, the reported evaluation found up to 74x end-to-end speedup over a projected CPU runtime for the 106M-vector MIRACL dataset, completing it in 8 minutes. The post states that performance continues to improve from one to eight GPUs while preserving comparable embedding quality across configurations."
---

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

[NVIDIA](https://yomu.fyi/company/nvidia-developer-blog) · Tanya Lenz · Aug 18, 2026

**Type:** Explainer

## Summary

UMAP workflows on tens to hundreds of millions of vectors are constrained by the cost of constructing an all-neighbors k-nearest-neighbor graph, especially when repeated runs support exploration and parameter tuning. The NVIDIA cuML and cuVS 25.06 feature distributes this graph-construction stage across multiple GPUs while retaining the out-of-core design that partitions data into balanced clusters, overlaps vectors across nearby clusters, computes local graphs, and merges them into a global graph. knn\_n\_clusters and knn\_overlap\_factor expose the memory, runtime, and graph-quality trade-off: more clusters reduce per-GPU data, while more overlap can improve nearest-neighbor recall at higher computational and memory cost. On eight NVIDIA H100 GPUs, the reported evaluation found up to 74x end-to-end speedup over a projected CPU runtime for the 106M-vector MIRACL dataset, completing it in 8 minutes. The post states that performance continues to improve from one to eight GPUs while preserving comparable embedding quality across configurations.

## Context

Repeated UMAP workflows become increasingly difficult to run interactively as datasets grow, because constructing the all-neighbors kNN graph becomes expensive for tens or hundreds of millions of vectors. The earlier out-of-core approach supported datasets too large for one GPU, but its training stage remained limited to a single GPU.

## Approach / What changed

The feature in NVIDIA cuML and cuVS 25.06 distributes out-of-core all-neighbors graph construction across GPUs. Data is partitioned into balanced clusters, vectors overlap across nearby clusters, local kNN graphs are computed independently, and the local graphs are merged into a global graph. The knn\_n\_clusters and knn\_overlap\_factor parameters control per-GPU memory, computation time, and neighborhood quality.

## Takeaways

- The 106M-vector MIRACL benchmark completed end-to-end processing in 8 minutes on eight NVIDIA H100 GPUs, with up to 74x speedup over a projected CPU runtime.
- Increasing knn\_n\_clusters reduces the amount of data assigned to each cluster and GPU, while increasing knn\_overlap\_factor generally improves kNN recall and embedding quality at higher memory and computation cost.
- Independent per-cluster graph construction avoids expensive all-to-all communication, and the reported scaling improves from one to eight GPUs while preserving comparable trustworthiness across configurations.

**Tags:** [Machine Learning](https://yomu.fyi/topic/machine-learning), [Performance](https://yomu.fyi/topic/performance), [Scalability](https://yomu.fyi/topic/scalability)

- Source: [NVIDIA](https://developer.nvidia.com/blog/run-massive-scale-umap-in-minutes-using-multiple-gpus-without-losing-accuracy)
- Source URL: https://developer.nvidia.com/blog/run-massive-scale-umap-in-minutes-using-multiple-gpus-without-losing-accuracy
- Ingested by Yomu: 2026-08-28T16:00:53.209Z

[Read original post](https://developer.nvidia.com/blog/run-massive-scale-umap-in-minutes-using-multiple-gpus-without-losing-accuracy)