DEV Community

IA Compa
IA Compa

Posted on Originally published at iacompa.com

NVIDIA Vera Rubin sweeps MLPerf Training v6.1 and leads MLPerf Endpoints in its debut

In MLPerf results released November 16, 2026, Vera Rubin NVL72 debuted sweeping every MLPerf Training v6.1 benchmark at up to 2x performance per GPU of GB300 NVL72, and led MLPerf Endpoints with more than 10x throughput per watt for agentic AI.

What changed in MLPerf Training v6.1 and Endpoints

Vera Rubin made its MLPerf debut in results released November 16, 2026, sweeping every MLPerf Training v6.1 benchmark. It was the only platform with results on every test, spanning LLMs, image generation, recommender systems and reinforcement learning post-training.

Vera Rubin NVL72 delivered up to 2x more performance per GPU than 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 4 minutes.

In the first MLPerf Endpoints submission — which tests agentic coding and enterprise workflows and checks that agents still complete tasks — Vera Rubin led across all of its agentic benchmarks. A provisional result on Kimi K3, published November 9, 2026, showed Vera Rubin NVL72 at more than 10x throughput per watt of GB300 NVL72 at the same response speed. NVIDIA notes the provisional Endpoints results are pending MLCommons verification and peer review.

Google Cloud published 10 MLPerf Training v6.1 results the same day, on 7th-generation Ironwood TPUs and Grace Blackwell GPUs, including the round's first agentic reinforcement learning benchmark.

Vera Rubin MLPerf by the numbers

  • Performance per GPU vs. GB300 NVL72 (MLPerf Training v6.1): Up to 2x per GPU (Nov 16, 2026)
  • Throughput per watt vs. GB300 NVL72 (MLPerf Endpoints, Kimi K3): More than 10x per watt (provisional, Nov 16, 2026)

Why Vera Rubin's MLPerf debut matters for business

Businesses serving AI agents are limited by data-center power, so Vera Rubin's throughput per watt on agentic workloads translates directly into how many agents a facility can serve per unit of electricity, and the dated MLPerf numbers give infrastructure buyers an independent yardstick for comparing NVIDIA, Google and other platforms before committing to a generation of hardware.

Sources

Top comments (0)