THE CRUNCH
NVIDIA's Vera Rubin platform has made its MLPerf debut with a clean sweep, topping every benchmark in MLPerf Training v6.1 and leading all of its agentic benchmarks in the first round of MLPerf Endpoints, a newer test that measures how well systems run AI agents on real coding and enterprise workflows. In training, the Vera Rubin NVL72 system delivered up to 2x more performance per GPU than the previous-generation GB300 NVL72, which NVIDIA describes as an early result. On the Endpoints side, the company reports more than 10x generational gains in throughput per watt for agentic AI, meaning how much agent work a system delivers for each watt of power.
The efficiency claim comes with conditions worth noting. The headline figure is a provisional result for Kimi K3, a large open model built for agentic workflows, comparing Vera Rubin NVL72 with GB300 NVL72 at a shared minimum interactivity of 50 tokens per second per user. NVIDIA notes these results are pending verification by MLCommons and subject to change after peer review.
Scale was also on show. On the DeepSeek-V3 training benchmark, Vera Rubin NVL72 scaled from 256 to 1,024 GPUs with about 97% efficiency, cutting training time from 16 minutes to roughly four. NVIDIA and Microsoft Azure both ran submissions at 10,240 GPUs, the largest scale this round, using GB300 and GB200 systems respectively. Twenty organisations submitted training results, including Google Cloud, Oracle, Dell and Supermicro.


