Nvidia’s next AI platform just made its benchmark debut — and it did not leave anything on the table. In MLPerf Training v6.1 results released November 16, the Vera Rubin NVL72 system posted the fastest time to train on benchmark after benchmark, while a first MLPerf Endpoints submission pointed to a huge efficiency jump for the AI agents now flooding data centers.

Rack-scale AI supercomputer with glowing GPU modules linked by light conduits, representing Nvidia Vera Rubin MLPerf performance

A clean sweep in training

According to NVIDIA’s official blog, its platform “swept every MLPerf Training v6.1 benchmark” and was the only one with results on every test in the round, spanning large language models, image generation, recommender systems and reinforcement-learning post-training. In its debut, Vera Rubin NVL72 delivered up to 2x more performance per GPU than the highly optimized GB300 NVL72 it succeeds.

The headline number came on DeepSeek-V3, a 671-billion-parameter model. Nvidia’s technical breakdown puts Vera Rubin at 4.14 minutes on 1,024 GPUs, against 8.03 minutes for GB300 on the same GPU count. Nvidia also says Vera Rubin scaled from 256 to 1,024 GPUs with about 97% efficiency, cutting training time from roughly 16 minutes to about four. The largest run of the round still belonged to the older hardware: Nvidia ran GB300 systems on 10,240 GPUs, a scale Microsoft Azure matched with GB200 systems.

10x per watt for agents — with an asterisk

Training is only half the story. In its first MLPerf Endpoints submission, Nvidia says Vera Rubin delivered “more than 10x generational gains in throughput per watt for agentic AI” on Kimi K3, a large open model built for agentic workflows, at the same response speed as GB300. That matters because agent workloads generate token after token as they reason, call tools and check their own work — and data centers are limited by power, not just by chip count.

The asterisk: Nvidia itself labels the Kimi K3 result provisional, pending MLCommons verification and peer review, so the 10x figure could still change. The training results, retrieved from MLCommons on November 16, stand on firmer ground.

Why it matters

This is Nvidia’s second big SC26 statement in a day, after the Rubin HPC GPU built for AI agents doing science and the Vera CPU rack packing 192 custom CPUs into one cabinet. Together they show a platform being positioned for every stage of the agent era — training, post-training, inference and the CPU work around it — just as rivals push custom chips into frontier labs. Twenty partner organizations submitted MLPerf results on Nvidia platforms this round, a reminder of how wide that ecosystem already is. If the provisional efficiency numbers survive verification, the cost of running agents at scale is about to drop hard.

Source: NVIDIA Blog — https://blogs.nvidia.com/blog/mlperf-training-endpoints-vera-rubin/

Mohsen Daemi