← The feed Many Minded · The brief
KNOWLEDGE · forward · impact 3/5 · 2026-11-16 · NVIDIA

Vera Rubin's first benchmark round: faster training, better tokens per watt

NVIDIA's Vera Rubin debuts in MLPerf with up to 2x per-GPU training speed over GB300 and a claimed 10x efficiency gain, though key results are provisional.

NVIDIA announced on November 16, 2026 that its Vera Rubin platform made its MLPerf debut, posting leading results in training, post-training and agentic inference. The company says it swept every benchmark in MLPerf Training v6.1 and was the only platform to submit results on all of them. Those tests span large language models, image generation, recommender systems and reinforcement learning post-training.

The numbers: Vera Rubin NVL72 delivered up to 2x the per-GPU training performance of the GB300 NVL72. On the DeepSeek-V3 benchmark it scaled from 256 to 1,024 GPUs at about 97% efficiency, cutting training time from 16 minutes to about four. For scale, NVIDIA ran DeepSeek-V3 on 10,240 GB300 GPUs, the largest in the round, and Microsoft Azure matched that count with GB200 GPUs. Twenty organizations submitted, including Dell, HPE, Google Cloud, CoreWeave and Oracle.

In its first MLPerf Endpoints submission, NVIDIA claims more than 10x generational gain in throughput per watt for agentic AI. In a provisional Kimi K3 run, a 32-accelerator Vera Rubin NVL72 reached 41,414.2 tokens/sec at 100.89 kW, versus 2,210.4 tokens/sec at 60 kW for a 48-accelerator GB300 NVL72.

Why it matters: the cost of intelligence is largely compute and electricity. Finishing training runs in fewer GPU-hours and serving more tokens per kilowatt lowers the price of building and running frontier models, which is what eventually widens who can use them.

Caveats are significant. The Vera Rubin Endpoints results are provisional, pending MLCommons verification, and only the GB300 Endpoints results are verified. The 10x figure compares runs with different accelerator counts and concurrency, at a shared 50 tokens/sec per user floor. Training results are early, and all of this comes from NVIDIA's own announcement. Watch for verified results and whether efficiency gains reach customer prices.

Source: NVIDIA