Nvidia’s Vera Rubin Posts Fastest Times in Six MLPerf Training Tests
Its 1,024-GPU DeepSeek-V3 run finished in 4.14 minutes, with about twice GB300’s per-GPU performance at the same scale.

Nvidia’s Vera Rubin NVL72 recorded the quickest training time in six of seven tests in the MLPerf Training v6.1 results released Monday, Nov. 16. The suite covered mixture-of-experts models, dense large language models and recommendation systems.
On DeepSeek-V3 671B, a system with 1,024 Vera Rubin GPUs completed training in 4.14 minutes. At the same GPU count, Vera Rubin delivered about twice GB300’s per-GPU performance. Expanding the Vera Rubin system from 256 to 1,024 GPUs produced about 97% scaling efficiency on that test.
Vera Rubin NVL72 pairs 72 GPUs with 36 CPUs and uses HBM4 memory and sixth-generation NVLink connections. The benchmark adds a measured training run to TokenPost’s earlier coverage of a fixed-workload Vera Rubin training estimate, which calculated hardware needs for a specified workload rather than reporting benchmark performance.