Technology

Add as a preferred source on Google

Nvidia Vera Rubin Completes DeepSeek-V3 Training Benchmark in 4.14 Minutes

The run used 1,024 GPUs; the MLPerf Training v6.1 round also ranked Vera Rubin first in six of seven tests and reported about 97% scaling efficiency.

Updated 1 min read
A gloved technician examines a GPU accelerator module in a Santa Clara, California, laboratory. / TokenPost.ai
A gloved technician examines a GPU accelerator module in a Santa Clara, California, laboratory. / TokenPost.ai

MLPerf’s latest round gives a like-for-like view of Vera Rubin and GB300 on a 1,024-GPU DeepSeek-V3 671B test, with Vera Rubin completing the training in 4.14 minutes.

Results released Monday, Nov. 16, from the MLPerf Training v6.1 round put Vera Rubin at the top in six of seven tests. The suite covered recommendation workloads alongside dense large language models and mixture-of-experts models.

For the DeepSeek-V3 comparison, Vera Rubin’s per-GPU performance was roughly twice GB300’s when both systems used the same number of GPUs. A scaling test recorded about 97% efficiency as Vera Rubin’s GPU count increased from 256 to 1,024.

The 72-GPU, 36-CPU NVL72 uses HBM4 and sixth-generation NVLink.

Related articles

View all