Nvidia's Vera Rubin NVL72 made its MLPerf debut in the MLPerf Training v6.1 round, delivering up to 2x more performance per GPU than the GB300 NVL72 system, according to Nvidia. MLCommons published the v6.1 results on November 16, and Nvidia said its platform was the only one to post results on every benchmark in the round, spanning large language models, image generation, recommender systems and reinforcement-learning post-training.
The scaling numbers carry the most weight. On the DeepSeek-V3 training benchmark, Nvidia said Vera Rubin NVL72 scaled from 256 to 1,024 GPUs at about 97% efficiency, and quadrupling the GPU count cut time to train from 16 minutes to about four.
The round's largest systems ran at 10,240 accelerators. Nvidia submitted at that scale with GB300 systems, and Microsoft Azure matched it with 10,240 Nvidia GPUs on GB200 systems. Nvidia credited NVLink for scale-up and Quantum-X800 InfiniBand and Spectrum-X Ethernet for scale-out networking.
Key facts
- 01Nvidia says Vera Rubin NVL72 delivered up to 2x the per-GPU performance of GB300 NVL72 in MLPerf Training v6.1.
- 02On DeepSeek-V3, Vera Rubin scaled from 256 to 1,024 GPUs at about 97% efficiency, cutting training time from 16 to about 4 minutes.
- 03MLCommons' v6.1 round drew 24 submitters, 153 systems and 250 results; the largest system used 10,240 accelerators.
- 0437% of v6.1 systems were cloud-based, up from 29% in the v6.0 round.
- 05Nvidia Rubin-generation GPUs and Google TPUv7 Ironwood posted first-time MLPerf Training results.
MLCommons said 24 organizations submitted to v6.1, producing 153 systems and 250 test results. Submitters included AMD, Google, Fujitsu, MiTAC, tiny corp and the University of Florida alongside Nvidia and its partners. Twelve systems used 1,000 or more accelerators.
The round also carried first-time MLPerf Training results for Nvidia's Rubin-generation GPUs and Google's TPUv7, known as Ironwood. Accelerators from Nvidia, AMD and Google all appear in the results.
“The MLPerf Training benchmark suite provides essential guidance for those deploying AI training systems”— Pavan Yalamanchili, MLPerf Training working group co-chair
Cloud infrastructure took a larger share of the field. MLCommons said 57 systems, or 37%, were cloud-based, up from 29% in the v6.0 round. Crusoe made its first MLPerf Training submission.
MLCommons added two benchmarks this round. DLRMv4-HSTU replaces DLRM-dcnv2 as the recommender test and trains on the Yambda dataset, while a new Qwen3.5-397B benchmark measures large-scale post-training with GRPO.
Nvidia also made its first MLPerf Endpoints submission. On an agentic workload using the Kimi K3 model, Nvidia said a 32-accelerator Vera Rubin NVL72 configuration delivered more than 10x the throughput per watt of a 48-accelerator GB300 NVL72 at the same response speed. Nvidia noted that the Vera Rubin Endpoints result is provisional, pending MLCommons verification and peer review.
Nvidia named 20 organizations that submitted training results on its platform, including CoreWeave, Dell Technologies, Google Cloud, HPE, Lambda, Nebius, Oracle and Supermicro. MLCommons' announcement did not list per-benchmark training times; the full v6.1 results are posted on the MLCommons results site.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.



