blogs.nvidia.com web signal

NVIDIA Vera Rubin NVL72 Sweeps MLPerf Training v6.1, Hits 2x Per-GPU Throughput Over GB300

Summary

Nvidia's new Vera Rubin NVL72 platform swept every benchmark in MLPerf Training v6.1 at debut, delivering up to 2x the per-GPU performance of the GB300 NVL72 and scaling from 256 to 1,024 GPUs at ~97% efficiency on DeepSeek-V3 (cutting training time from 16 to ~4 minutes). On a provisional Kimi K3 agentic inference test, Vera Rubin NVL72 posted more than 10x the throughput per watt of GB300 at the same response speed. CoreWeave reported first multi-rack Vera Rubin cloud results in the same round.