What changed
NVIDIA announced that its Vera Rubin platform achieved leading performance across MLPerf Training v6.1 benchmarks and its first MLPerf Endpoints submission. The platform delivered more than 10x generational gains in throughput per watt for agentic AI, measuring how much agentic work a system can perform with each watt of power. NVIDIA also demonstrated performance at scale, with Vera Rubin NVL72 scaling from 256 to 1,024 GPUs with about 97% efficiency, quadrupling the GPUs cut time to train from 16 minutes to about four.
What this means in practice
The results highlight NVIDIA's leadership in AI training and agentic inference, with Vera Rubin NVL72 systems outperforming the NVIDIA GB300 NVL72 in throughput per watt for agentic workflows. For example, in a provisional result for Kimi K3, a large open model built for agentic workflows, Vera Rubin NVL72 systems delivered more than 10x the throughput per watt of NVIDIA GB300 NVL72 at the same response speed. These results are part of the MLPerf Training v6.1 benchmarks, which measure how fast a system can train a range of AI models to a target level of accuracy. NVIDIA was the only platform with results on every benchmark this round, spanning large language models, image generation, recommender systems, and reinforcement learning post-training.
Limitations and availability
The MLPerf results are subject to verification by the MLCommons Association, and the provisional Vera Rubin results are pending peer review. NVIDIA's Vera Rubin platform is available for customers to invest in, with continuous optimization of NVIDIA’s CUDA-powered software stack expected to enhance performance over time. The MLPerf name and logo are trademarks of MLCommons Association, and all rights are reserved.
