NVIDIA Vera Rubin leads MLPerf training and agent benchmarks
NVIDIA topped every MLPerf Training v6.1 benchmark and was the only participant to submit results across all workloads, covering large language models, image generation, recommendation systems and reinforcement learning post-training. In its debut, Vera Rubin NVL72 delivered up to 2x more performance per GPU than NVIDIA GB300 NVL72. The training benchmarks measure how quickly systems bring models to a target accuracy.
On DeepSeek-V3, Vera Rubin NVL72 scaled from 256 to 1,024 GPUs with about 97% efficiency. NVIDIA also ran the benchmark on 10,240 GPUs using GB300 systems, while Microsoft Azure matched that scale with GB200 systems. NVIDIA attributed efficient scaling to its GPU interconnects, data-center networking and optimized software, which coordinate workloads across systems.
In NVIDIA’s first MLPerf Endpoints submission, Vera Rubin led the agentic benchmarks, which test coding and enterprise workflows while checking task completion. A provisional Kimi K3 result showed more than 10x the throughput per watt of GB300 NVL72 at the same response speed. The Vera Rubin inference results remain subject to MLCommons verification and may change following peer review. These efficiency gains matter for power-constrained data centers serving agents that repeatedly reason, call tools and check their work.