Nvidia Corp. swept every MLPerf Training v6.1 benchmark on November 16 with its Vera Rubin architecture's debut submission, delivering up to 2x the performance of Blackwell Ultra on the NVL72 platform1,2. The Vera Rubin NVL72 also led agentic inference results in the first MLPerf Endpoints v1.0 round, posting more than 10x generational gains in throughput per watt for agentic AI workloads.
Nvidia said its CUDA software stack for Vera Rubin is already highly mature at debut. MLPerf Endpoints v1.0 evaluates the balance between serving capacity and responsiveness using coding and enterprise workflows through API-level testing3.
Separately, Lambda published MLPerf Training v6.1 results on the same date showing 2.86x faster language-model training on newer Nvidia accelerated hardware[1].
ANALYSIS Vera Rubin's clean sweep across both training and inference benchmarks at first submission extends Nvidia's unbroken MLPerf dominance into its next hardware generation.