NVIDIA sweeps MLPerf benchmarks as Vera Rubin posts AI performance gains

NVIDIA sweeps MLPerf benchmarks as Vera Rubin posts AI performance gains
NVIDIA leads MLPerf race

AI infrastructure competition is increasingly centered on how quickly and efficiently systems can train models and run agentic workloads at scale. NVIDIA says its Vera Rubin platform leads its MLPerf debut across training, post-training and agentic inference, with results released today and early signs of further gains from software optimization.

Highlights

  • NVIDIA’s Vera Rubin platform achieved results across every MLPerf Training v6.1 benchmark, showcasing up to 2x GPU performance over GB300 NVL72 and over 10x throughput per watt on Kimi K3.
  • Vera Rubin NVL72 demonstrated 97% scaling efficiency on DeepSeek-V3 benchmarks, reducing training time from 16 to about 4 minutes when GPUs increase from 256 to 1,024.
  • NVIDIA reported MLPerf submissions from 20 organizations, including major cloud providers, and claimed leadership on SemiAnalysis AgentX and Artificial Analysis AgentPerf benchmarks, highlighting broad enterprise adoption.

Benchmark results and platform scaling

As reported by NVIDIA Blog, citing NVIDIA Newsroom, the company says Vera Rubin is the only platform with results across every MLPerf Training v6.1 benchmark in this round, covering large language models, image generation, recommender systems and reinforcement learning post-training.

In its first MLPerf Endpoints submission, NVIDIA says Vera Rubin also leads across its agentic benchmarks. The company says Vera Rubin NVL72 delivers up to 2x more performance per GPU than the NVIDIA GB300 NVL72 system in training tests, while a provisional Kimi K3 result shows more than 10x higher throughput per watt than GB300 NVL72 at the same response speed.

NVIDIA also highlights scaling efficiency on the DeepSeek-V3 benchmark. Vera Rubin NVL72 scales from 256 to 1,024 GPUs with about 97% efficiency, cutting training time from 16 minutes to about four when GPU count is quadrupled.

The company adds that it runs GB300 systems on 10,240 GPUs, the largest scale in this benchmark round, while Microsoft Azure matches that scale with 10,240 NVIDIA GPUs using GB200 systems. NVIDIA says NVLink handles scale-up connections within systems, while Quantum-X800 InfiniBand and Spectrum-X Ethernet link systems across data centers to keep large GPU clusters operating efficiently.

Implications for enterprise AI demand

These results matter for customers building AI agents for coding, enterprise workflows, customer service, scientific research and other power-intensive applications. Faster training can shorten model development cycles, while higher token output per watt can improve the economics of serving more users in power-constrained data centers.

NVIDIA says Vera Rubin shifts the trade-off between concurrency and response speed by allowing systems to serve more users while responding faster. The company also says performance should continue to improve through ongoing optimization of its CUDA-based software stack, potentially raising the long-term value of infrastructure already being deployed.

The broader partner ecosystem remains a key part of that positioning. NVIDIA says 20 organizations submitted MLPerf Training results using its platform, including Azure, Cisco, CoreWeave, Dell Technologies, Fujitsu, Google Cloud, HPE, Oracle, Supermicro and the University of Florida.

Beyond MLPerf, NVIDIA says Vera Rubin posts leading results on SemiAnalysis AgentX and Artificial Analysis AgentPerf. The company positions that breadth as evidence that the same infrastructure can support pretraining, post-training, agentic inference and high-performance computing workloads.

As we previously wrote, Saturn Funding Trust notes received preliminary A- ratings backed by lease payments on NVIDIA Vera Rubin NVL72 systems for SpaceX.

This material may contain third-party opinions, none of the data and information on this webpage constitutes investment advice according to our Disclaimer. While we adhere to strict Editorial Integrity, this post may contain references to products from our partners.