NVIDIA's Vera Rubin debuts on MLPerf: up to 2x performance per GPU in training and over 10x throughput per watt for agentic inference
Synopsis
NVIDIA's Vera Rubin platform entered MLPerf for the first time, with submitted results across training, post-training and agentic inference. In training, Vera Rubin NVL72 delivered up to 2x more performance per GPU than the highly optimized GB300 NVL72 in its debut; 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. In its first MLPerf Endpoints submission, Vera Rubin NVL72 showed more than 10x the throughput per watt of GB300 NVL72 at the same response speed in a provisional Kimi K3 result.
Interpretation
NVIDIA's Vera Rubin platform covered training, post-training and agentic inference in its first MLPerf appearance, and was the only platform with results on every MLPerf Training v6.1 benchmark this round. The comparison is the highly optimized NVIDIA GB300 NVL72 system; Vera Rubin NVL72 delivered up to 2x more performance per GPU than it in its debut, described as early results. MLPerf Training v6.1 spans large language models, image generation, recommender systems and reinforcement learning post-training, and Vera Rubin NVL72 has a result on each, while no other system covered them all.
On scale, Vera Rubin NVL72 ran the DeepSeek-V3 benchmark from 256 to 1,024 GPUs, quadrupling the GPU count and cutting training time from 16 minutes to about four. The scaling held about 97% efficiency, meaning added GPUs still shortened training close to proportionally. In the same round, NVIDIA also ran the benchmark on 10,240 GPUs with GB300 systems, the largest scale this round, and Microsoft Azure matched that scale with 10,240 GB200 GPUs.
In its first MLPerf Endpoints submission, Vera Rubin NVL72 led across all of its agentic benchmarks, which measure how many tokens per second a system generates for each watt of power, or how much agentic work it serves at a given power. The comparison is again GB300 NVL72; in a provisional result for Kimi K3, a large open model built for agentic workflows, Vera Rubin NVL72 delivered more than 10x the throughput per watt of GB300 NVL72 at the same response speed. MLPerf Endpoints tests systems on real agentic coding and enterprise workflows and checks that the agents still complete their tasks correctly, not throughput alone.
Perspective
The demonstrated scope is the MLPerf Training v6.1 and MLPerf Endpoints submissions, covering large language models, image generation, recommender systems, reinforcement learning post-training, and agentic coding and enterprise workflows; the generational comparison is explicitly called early results and the Kimi K3 figure is provisional.
NVIDIA says Vera Rubin performance will keep climbing through continuous optimization of its CUDA software stack, so the current numbers are not a final point; the 2x per-GPU generational comparison is described as early results, and the Kimi K3 throughput-per-watt comparison is provisional.
