Nvidia’s Vera Rubin platform made its MLPerf benchmarking debut with results released November 16, delivering what the company calls leading performance across training and inference tasks.
In MLPerf Training v6.1, Nvidia submitted results for every benchmark category—the only participant to do so this round. The Vera Rubin NVL72 system achieved up to 2x more performance per GPU compared to Nvidia’s GB300 NVL72 system, according to the company. Benchmarks spanned large language models, image generation, recommender systems and reinforcement learning post-training.
Nvidia also demonstrated scaling efficiency on the DeepSeek-V3 benchmark, where Vera Rubin NVL72 maintained approximately 97% efficiency when scaling from 256 to 1,024 GPUs. Quadrupling the GPU count reduced training time from 16 minutes to about four minutes. The company ran the same benchmark on 10,240 GB300 GPUs, representing the largest scale submitted this round. Microsoft Azure matched that scale using 10,240 Nvidia GB200 GPUs.
In its first MLPerf Endpoints submission, Nvidia reported more than 10x generational gains in throughput per watt for agentic AI workloads—a measure of how much agentic work a system delivers per watt of power consumed. MLPerf Endpoints measures inference performance for AI agent systems.
Nvidia stated that Vera Rubin performance will continue improving through ongoing optimization of its CUDA software stack.