NVIDIA Shatters Graph500 Record with H100 GPU-Powered Cloud Supercomputer

Details

• NVIDIA set a new Graph500 benchmark with 410 trillion traversed edges per second (TEPS), ranking first on the 31st Graph500 breadth-first search (BFS) list.
• The record was achieved using 8,192 NVIDIA H100 GPUs running in CoreWeave’s Dallas data center, processing a graph of 2.2 trillion vertices and 35 trillion edges.
• NVIDIA’s solution more than doubled the performance of comparable alternatives while using just over 1,000 nodes, far fewer than the 9,000+ nodes required by other top entries.
• The system delivered three times better performance per dollar, leveraging a full-stack approach with CUDA, Spectrum-X, H100 GPUs, and a new active messaging library.
• This advance aims to democratize access to supercomputing for sparse, irregular workloads like social network analysis, financial modeling, and weather forecasting.

Impact

NVIDIA’s breakthrough not only sets a new technical standard but signals a shift toward commercial cloud providers as key players in high-performance computing. The achievement underscores GPUs’ growing dominance in handling complex, irregular workloads and may accelerate the pace of innovation across science and industry. Expect cloud-based AI and graph processing to see broader adoption as cost and accessibility barriers fall.