Nvidia's Vera Rubin NVL72 debuts in MLPerf with 3.7x gain

Nvidia published its first MLPerf Inference preview submission for the Vera Rubin NVL72 on September 16, reporting up to 3.7 times the throughput of the shipping GB300 NVL72 on the Qwen3-VL benchmark and up to 2.5 times on DeepSeek-R1 in the v6.1 round. The company also submitted a 288-GPU, four-rack GB300 NVL72 configuration that held 99% scaling efficiency on DeepSeek-R1, and said software work alone lifted Qwen3-VL performance up to 1.6x over the v6.0 round. Nineteen partners filed results on Nvidia systems and cloud provider Nebius submitted its own Vera Rubin NVL72 preview, the first external numbers for the platform that reached customers in July.