NVIDIA says Vera Rubin systems deliver up to twice GB300 training speed per GPU
The Vera Rubin NVL72 system made its first appearance in MLPerf Training v6.1. MLCommons, which runs the benchmark, confirmed that this round was the first to include Rubin-generation GPUs.

NVIDIA's Vera Rubin NVL72 system posted its first MLPerf Training results on 16 November 2026. The company says that, measured per GPU, the system reached up to 2x the performance of its GB300 NVL72 system1. MLCommons, the group behind the benchmark, said the v6.1 round was the first to include NVIDIA Rubin-generation GPUs2.
MLPerf Training measures how quickly hardware can bring various AI models to a chosen accuracy level1. NVIDIA says it swept every benchmark in v6.1.
NVIDIA also reported results on scaling. In the DeepSeek-V3 benchmark, it says Vera Rubin NVL72 went from 256 to 1,024 GPUs at about 97% efficiency, and four times as many GPUs cut training time from 16 minutes to about four1. NVIDIA ran the same benchmark on 10,240 GPUs in GB300 systems, the largest scale in the round, and Microsoft Azure matched it with 10,240 NVIDIA GPUs in GB200 systems1.
According to MLCommons, 24 organizations submitted results, including AMD, Google, Dell, HPE, Oracle, CoreWeave and Crusoe, a first-time entrant2. Together they entered 153 systems and produced 250 test results. Cloud-based systems made up 37% of the total, compared with 29% in v6.0, and a dozen entries ran on at least 1,000 accelerators2. Hardware came from NVIDIA, AMD and Google, and Google's TPUv7 (Ironwood) also appeared for the first time2.
Version 6.1 adds two new benchmarks2. One is DLRMv4-HSTU, a recommendation test that replaces the earlier DLRM-dcnv2 and trains on the Yambda dataset. The other is Qwen3.5-397B, a large-scale workload for post-training a large language model2.
NVIDIA also entered MLPerf Endpoints for the first time. The figures for Kimi K3, a large open model designed for agent-style workflows, are provisional. NVIDIA says they are pending MLCommons Association verification and may change after peer review1.
Sources
- 1NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance
PrimaryNVIDIA BlogAccessed
- 2MLPerf Training v6.1 Results: Record Participation
IndependentMLCommonsAccessed
ITANU’s newsroom system produced this article with AI and checked it against the sources listed above. How we work

