Vera Rubin NVL72 MLPerf Training v6.1 and MLPerf Endpoints results
NVIDIA's Vera Rubin platform debuts on MLPerf, with NVIDIA winning every Training v6.1 benchmark and claiming over 10x throughput per watt versus GB300 on MLPerf Endpoints.
Vera Rubin NVL72 delivered up to 2x more performance per GPU than GB300 NVL72 in MLPerf Training v6.1. On DeepSeek-V3, it scaled from 256 to 1,024 GPUs at about 97% efficiency, cutting training time from 16 minutes to about four. In the first MLPerf Endpoints submission, a provisional Kimi K3 result showed more than 10x the throughput per watt of GB300 NVL72 at the same response speed.
- Date
- Monday 16 November 2026
- Lab
- NVIDIA
- Kind
- infra
- Access
- app only
Figures
| Measure | Value | Measured by |
|---|---|---|
| Per-GPU training performance vs GB300 NVL72 | up to 2x MLPerf Training v6.1, early results | company |
| Scaling efficiency, 256 to 1,024 GPUs (DeepSeek-V3) | about 97% time to train fell from 16 minutes to about 4 | company |
| Throughput per watt vs GB300 NVL72 (Kimi K3, agentic) | more than 10x provisional, pending MLCommons verification; at 50 tokens/sec per user interactivity | company |
| Vera Rubin NVL72 system throughput (Kimi K3, 32 accelerators) | 41,414.2 tokens/sec at 100.89 kW provisional | company |
| Largest scale in round | 10,240 GPUs NVIDIA on GB300 systems; Azure matched with GB200 | company |
Figures come from NVIDIA's own blog. The Vera Rubin Endpoints results are provisional and not yet verified by MLCommons. The access field is a placeholder because this is a benchmark result for hardware, not a model.
Sources
This record was checked against its sources on 16 November 2026. How we check