Nvidia Vera Rubin NVL72 sweeps MLPerf Training v6.1 at debut
TL;DR
- Vera Rubin NVL72 posted results on every MLPerf Training v6.1 benchmark and delivered up to 2x the per-GPU performance of Nvidia's own GB300 NVL72.
- On DeepSeek-V3, the system scaled from 256 to 1,024 GPUs at about 97% efficiency, cutting training time from 16 minutes to 4.
- On a provisional Kimi K3 agentic test, 32 Vera Rubin accelerators beat a 48-accelerator GB300 by more than 10x on throughput per watt.
Nvidia's Vera Rubin NVL72 showed up at MLPerf Training v6.1 and posted results on every benchmark, up to 2x per-GPU over Nvidia's own GB300 NVL72, according to Nvidia's blog.
"In its debut, Vera Rubin NVL72 delivered up to 2x more performance per GPU than the highly optimized NVIDIA GB300 NVL72 system," the post states. On the DeepSeek-V3 benchmark, the platform scaled from 256 to 1,024 GPUs at about 97% efficiency, cutting training time from 16 minutes to 4.
The bigger number sits inside a new category. In its first MLPerf Endpoints submission, Nvidia claimed "more than 10x generational gains in throughput per watt for agentic AI." On a provisional Kimi K3 run, 32 Vera Rubin accelerators beat a 48-accelerator GB300 by more than 10x on throughput per watt at the same response speed. The Endpoints benchmark, per the post, "tests systems on real agentic coding and enterprise workflows, and checks that the agents still complete their tasks correctly."
Twenty partners submitted on Nvidia platforms this round, including CoreWeave, Microsoft Azure, Oracle, Dell, HPE, Google Cloud, Lambda, Nebius, Crusoe and Supermicro. Azure matched the top GB200 scale at 10,240 GPUs; the Vera Rubin figures are at 1,024.
The post names no Nvidia executive and offers no third-party corroboration; the Kimi K3 figure is flagged provisional and no competing accelerator vendor appears in the sweep. It lands inside a dense run of coverage, 284 tracked Nvidia stories in the past 90 days on our radar, and is the sharpest per-watt argument Nvidia has made yet for rolling off Blackwell.
Originally reported by blogs.nvidia.com
Read the original article →Original headline: NVIDIA Vera Rubin NVL72 Sweeps MLPerf Training v6.1, Hits 2x Per-GPU Throughput Over GB300