Nvidia’s 16 November technical report says Vera Rubin NVL72 completed all seven MLPerf Training v6.1 benchmarks in its debut, recording the fastest training time on six. The suite covers language models, image generation, recommendations and reinforcement learning.
On DeepSeek-V3 671B, Nvidia measured 4.14 minutes with 1,024 Vera Rubin GPUs, against 8.03 minutes with the same number of GB300 NVL72 GPUs. Training that model requires its specialist components to exchange work across GPUs, making the connections between them important to speed. Nvidia says Vera Rubin maintained above 97% scaling efficiency as it grew from 256 to 1,024 GPUs.
Answering customer questions could draw on a newly trained model sooner if the reduction in training time held for that work.
GB300 still recorded the fastest DeepSeek-V3 time overall: 1.38 minutes using 10,240 GPUs. On the suite’s new reinforcement-learning post-training test, Qwen 3.5 397B GRPO, Nvidia reported 10.44 minutes for Vera Rubin and 13.78 minutes for GB300, each using 1,024 GPUs. Vera Rubin also completed that test in 10.94 minutes using vLLM as its inference backend.