Nvidia says multi-GPU UMAP can process 870GB of vector data in 8 minutes

Nvidia said on Aug. 19 that its cuML and cuVS libraries now support multi-GPU UMAP, a feature designed to spread large-scale vector-data dimensionality reduction across several GPUs while keeping embedding quality intact and cutting runtime. The company said the approach is aimed at UMAP jobs that can run into the hundreds of gigabytes. Those workloads, Nvidia said, could previously take hours or even days. In testing on a DGX system with eight H100 GPUs, Nvidia said the MIRACL dataset — which contains 106 million vectors and totals about 870GB — was processed end to end in 8 minutes. Nvidia said the result was up to 74 times faster than an estimated CPU implementation. The company also said a CPU-based setup was unable to complete the full workload even with 2TB of memory. Nvidia said the method works by splitting the data into multiple clusters, building local k-nearest-neighbor graphs in parallel on different GPUs, and then merging them into a global graph. That design, the company said, helps bypass the memory limits of a single GPU.