Nvidia says Vera Rubin NVL72 led all seven MLPerf Training v6.1 benchmarks

Editorial illustration: Green computing modules lead gray counterparts across seven parallel metallic lanes, with glowing green trails against a dark background.

In brief

  • Nvidia's Vera Rubin submissions led all seven MLPerf Training v6.1 tests, fastest on six, Crypto Briefing reported.
  • Per-GPU gains reached up to 2x over the Blackwell Ultra GB300 NVL72, Nvidia reported.
  • DeepSeek-V3 671B: Nvidia's run finished training in 4.14 minutes on 1,024 Rubin GPUs.
  • CoreWeave said its 256-GPU Vera Rubin run beat its GB200 NVL72 setup by 2.7x.

What the benchmark numbers show

The comparison point matters. The GB300 NVL72 is the Blackwell Ultra system Vera Rubin replaces, so Nvidia is measuring the new rack against its own previous generation (and a heavily optimized one at that). The test workloads covered large mixture-of-experts models, dense large language models and advanced recommendation systems.

So what does "up to 2x per GPU" look like in practice? DeepSeek-V3 gives the clearest example. On the DeepSeek-V3 671B benchmark, Nvidia's submission finished training in 4.14 minutes using 1,024 Rubin GPUs. That's roughly double the GB300's per-GPU training speed at the same scale, according to the results. Nvidia also reported about 97% scaling efficiency when going from 256 to 1,024 GPUs on that benchmark.

These are Nvidia's own submissions, not independently verified figures.

Inference efficiency and CoreWeave's run

Vera Rubin also made its first appearance in MLPerf Endpoints v1.0. There, Nvidia said the system delivered more than 10 times the throughput per watt of the GB300 NVL72 on the Kimi K3 model under identical interactivity conditions.

CoreWeave submitted results as well. The cloud provider said its multi-rack Vera Rubin submission on DeepSeek-V3 at 256 GPUs ran 2.7 times faster than its previous GB200 NVL72 setup. That's CoreWeave's own figure, and it isn't a third-party check on Nvidia's numbers.

Same rack, new parts

The physical layout didn't change. Vera Rubin keeps the 72-GPU, 36-CPU rack configuration, and the platform uses HBM4 memory and sixth-generation NVLink interconnects.