Nvidia said in a technical blog post on Aug. 19 that cuML and cuVS now include multi-GPU UMAP support, allowing large-scale vector-data dimensionality reduction to run across multiple GPUs while preserving embedding quality and cutting runtime. The company said that on a DGX system with eight H100 GPUs, testing on the MIRACL dataset, which contains 106 million vectors and about 870GB of data, the end-to-end process finished in 8 minutes. Nvidia said the result represents up to a 74x speedup versus an estimated CPU implementation. It also said a prior CPU approach could not complete the full workload even with 2TB of memory. The method splits data into multiple clusters, builds local k-nearest-neighbor graphs in parallel across GPUs, and then merges them into a global graph to bypass single-GPU memory limits.