NVIDIA reports Vera Rubin training gains while its AI agent result awaits verification

Executive Briefing
“NVIDIA has reported Vera Rubin’s first results in MLPerf Training and MLPerf Endpoints. The...”
Key Takeaways
- 1.NVIDIA says its platform was the only one with results in every MLPerf Training v6.1 test this round. The tests covered large language models, image generation, recommendation systems and reinforcement-learning post-training.
- 2.In its initial results, NVIDIA says Vera Rubin NVL72 delivered up to twice the performance per GPU of an optimized GB300 NVL72 system.
- 3.NVIDIA reports that, on the DeepSeek-V3 test, increasing Vera Rubin NVL72 from 256 to 1,024 GPUs reduced training time from 16 minutes to about four minutes, with roughly 97% scaling efficiency.
- 4.For an AI agent test using Kimi K3, NVIDIA reports that Vera Rubin NVL72 achieved more than ten times the throughput per watt of GB300 NVL72 at a shared minimum interactivity of 50 tokens per second per user. The Vera Rubin result is provisional, awaits MLCommons verification and may change after peer review.
NVIDIA has reported Vera Rubin’s first results in MLPerf Training and MLPerf Endpoints. The company reports a training performance gain over its GB300 system, but the power-efficiency result for AI agent workloads is still awaiting MLCommons verification.
Key Developments and Technical Context
NVIDIA says its platform was the only one with results in every MLPerf Training v6.1 test this round. The tests covered large language models, image generation, recommendation systems and reinforcement-learning post-training.
In its initial results, NVIDIA says Vera Rubin NVL72 delivered up to twice the performance per GPU of an optimized GB300 NVL72 system.
NVIDIA reports that, on the DeepSeek-V3 test, increasing Vera Rubin NVL72 from 256 to 1,024 GPUs reduced training time from 16 minutes to about four minutes, with roughly 97% scaling efficiency.
For an AI agent test using Kimi K3, NVIDIA reports that Vera Rubin NVL72 achieved more than ten times the throughput per watt of GB300 NVL72 at a shared minimum interactivity of 50 tokens per second per user. The Vera Rubin result is provisional, awaits MLCommons verification and may change after peer review.
Strategic Analysis and Industry Implications
The Kimi K3 throughput-per-watt comparison uses a specific response-speed threshold. It does not establish the same advantage for every model or AI agent deployment.
The DeepSeek-V3 scaling result suggests that adding GPUs can substantially shorten this test. It does not establish the operating cost of a real-world training run.
Outlook and Key Factors to Watch
Will Vera Rubin’s MLPerf Endpoints result change after MLCommons verification and peer review?
How will these benchmark results translate into performance, power use and cost in customer systems?







