NVIDIA, CoreWeave, IBM: The largest GB200 NVL72 cluster in MLPerf benchmark testing.

NVIDIA, CoreWeave, and IBM have submitted the largest MLPerf Training v5.0 report in history around the GB200 Grace Blackwell super chip. They used 2496 Blackwell GPUs to run CoreWeave's AI-optimized cloud platform. This submission features the largest NVIDIA GB200 NVL72 cluster in MLPerf compliance testing to date, 34 times larger than the only previous submission from a cloud provider, showcasing the immense scale of CoreWeave's cloud platform and its thorough preparation for today's demanding AI workloads. This submission made significant progress on the benchmark suite's largest and most complex base model Llama 3.1 405B, completing the run in just 27.3 minutes. Compared to other participants' submissions at similar cluster scales, CoreWeave's GB200 cluster showed a performance improvement of over 2 times. This result highlights the significant performance leap brought by the GB200 NVL72 architecture and the strong capabilities of CoreWeave's infrastructure in providing consistent, top-tier AI workload performance.