NVIDIA says Vera Rubin tops all MLPerf Training benchmarks; Endpoints shows more than 10x throughput per watt
NVIDIA published Vera Rubin’s MLPerf benchmark results on November 16, 2026. The company said the platform ranked first in every MLPerf Training v6.1 benchmark. In its first MLPerf Endpoints submission, it reported more than 10x generation-over-generation throughput per watt for agentic AI.
Key facts
- NVIDIA published MLPerf benchmark results for its Vera Rubin platform on November 16, 2026.
- NVIDIA said Vera Rubin ranked first in every MLPerf Training v6.1 benchmark.
- According to NVIDIA, its first MLPerf Endpoints submission showed a generation-over-generation improvement of more than 10x in throughput per watt for agentic AI.
What happened
NVIDIA published MLPerf benchmark results for its Vera Rubin platform on November 16, 2026. The company said it ranked first in every MLPerf Training v6.1 benchmark and, in its first MLPerf Endpoints submission, showed a generation-over-generation improvement of more than 10x in throughput per watt for agentic AI.
Why it matters
NVIDIA’s announcement presents separate results for all MLPerf Training v6.1 benchmarks and for agentic AI throughput per watt, so those measures should be considered separately when comparing computing platforms. The more-than-10x figure is a claim based on the company’s benchmark results; it does not demonstrate power efficiency in real-world deployments or robot performance.
Analysis
What stands out in these results is that Vera Rubin was evaluated across distinct measures: all MLPerf Training v6.1 benchmarks and throughput per watt for agentic AI in Endpoints. This was NVIDIA’s first Endpoints submission, and the company said it showed more than 10x generation-over-generation performance on a metric for agentic AI. However, the figure is based on benchmark results presented by NVIDIA; the information provided does not show that power consumption, costs or processing quality in real-world use would improve by the same ratio.
In a related article, Germany’s Daisytuner raises €6 million to optimize AI models across chips and launches an inference engine, this newspaper covered an inference engine designed to improve execution efficiency across multiple chips. Daisytuner’s focus is execution optimization across chips, while NVIDIA’s announcement concerns benchmark performance using a computing platform; they address different levels of comparison. This highlights the need to evaluate efficiency in execution environments that include different chips separately from benchmark results for a specific platform when selecting AI computing infrastructure.
For physical AI, leaderobot.com reports limited sale of five-finger hand “Xingdong XHAND 1 Pro-U” for NVIDIA GEAR DexUMI covered a hand compatible with NVIDIA’s GEAR DexUMI. The MLPerf results here are measures of computing platform performance and do not directly demonstrate robot manipulation capabilities or learning outcomes in real-world environments. Computing performance and evaluations of the robot itself and its control systems should not be conflated; it is important to distinguish which stage is being measured.
The next details to check are individual scores and test conditions for each Training benchmark, the generation and configuration used as the comparison for Endpoints, and how throughput per watt was calculated. Without this information, it is difficult to compare the more-than-10x claim with other computing platforms or actual operating conditions. The announcement says Vera Rubin also showed high performance in post-training and inference, but provides no figures or task conditions for those results. Connecting benchmark results to deployment decisions will depend on how well the benchmarks’ scope corresponds to actual processing requirements.
Details
NVIDIA said Vera Rubin delivered high performance in training, post-training and agentic inference. It submitted results for every MLPerf Training v6.1 benchmark, and this was its first submission to MLPerf Endpoints.
Sources
Translated from the Japanese original: /news/nvidia-vera-rubin-de-mlperf-training-zenkoumoku-wo-seiha-end-891aa0. The Japanese version is authoritative if they differ.