BlockBeats News, August 19th - NVIDIA released a technical blog post stating that its cuML and cuViz libraries have added multi-GPU UMAP capability. This enables the distributed dimensionality reduction of large-scale vector data to be performed across multiple GPUs, significantly reducing the runtime while preserving embedding quality.

NVIDIA indicated that on a DGX system equipped with 8 H100 GPUs, testing was conducted on the MIRACL dataset containing 106 million vectors, with a data size of approximately 870GB. The cuML multi-GPU UMAP was able to complete end-to-end processing in 8 minutes, achieving a speedup of up to 74 times compared to the expected CPU implementation speed. The previous CPU solution could not complete the full processing even with 2TB of memory.

This approach divides the data into multiple clusters, parallelly constructs local k-nearest neighbor (kNN) graphs on different GPUs, and then merges them into a global graph, thus overcoming the single-GPU memory limit. NVIDIA stated that this technology can compress UMAP tasks of several hundred gigabytes that previously took hours or even days to complete to just a few minutes.