NVIDIA's Vera Rubin platform debuts in MLPerf with leading training and agentic inference results
In its first MLPerf submissions, NVIDIA's Vera Rubin platform topped every Training v6.1 benchmark and led agentic inference in Endpoints, with the company reporting more than 10x generational gains in throughput per watt.
In plain words
NVIDIA's newest computing platform, Vera Rubin, was tested for the first time in MLPerf, an independent benchmark suite that measures how fast AI systems train models and serve AI agents. NVIDIA reports it was the only platform with results on every training benchmark this round and that it leads the agentic inference tests. One caveat: the standout inference figure is a provisional result that has not yet been verified by MLCommons.
What was announced
According to NVIDIA's blog, the Vera Rubin NVL72 system delivered up to 2x more performance per GPU than the company's earlier GB300 NVL72 in MLPerf Training v6.1, which measures how quickly systems train AI models to a target accuracy. NVIDIA says its platform was the only one with results on every benchmark this round, covering large language models, image generation, recommender systems and reinforcement learning post-training. The company also cites scaling results: on the DeepSeek-V3 benchmark, Vera Rubin NVL72 quadrupled from 256 to 1,024 GPUs with about 97% efficiency, cutting training time from 16 minutes to roughly four.
In the first MLPerf Endpoints round, which tests agentic coding and enterprise workflows while checking that tasks complete correctly, NVIDIA says Vera Rubin leads all of its agentic benchmarks. The headline example is a provisional result for Kimi K3, a large open model for agentic workflows: Vera Rubin NVL72 reportedly delivered more than 10x the throughput per watt of GB300 NVL72 at a shared minimum interactivity of 50 tokens per second per user. NVIDIA notes the comparison is based on throughput per kilowatt, and that the Vera Rubin figures remain subject to MLCommons verification and peer review. Twenty partner organizations, including Azure, Google Cloud and Dell Technologies, also submitted training results on NVIDIA platforms.
What it does not show
- The results come from NVIDIA's own blog; the Vera Rubin Endpoints figures are provisional and pending MLCommons verification, so independent confirmation is not yet established.
- The 10x throughput-per-watt comparison relies on a shared minimum interactivity level and specific benchmark configurations, and the source does not provide full details of all tested workloads.
For security and IT teams
- Scope
- The reported results cover MLPerf Training v6.1 and the first MLPerf Endpoints v1.0 round, including agentic coding and enterprise workflow benchmarks, as described by NVIDIA.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.