NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance

Nvidia Blog
Read full story· 5 min
NVIDIA's Vera Rubin platform debuted in MLPerf benchmarks, sweeping all Training v6.1 categories and delivering over 10x efficiency gains in agentic AI inference. It showed up to 2x GPU performance improvement and near-linear scaling up to 10,240 GPUs.

What happened

NVIDIA introduced its Vera Rubin platform in the latest MLPerf benchmarks, with results released on November 16, 2026. The platform achieved top results in all MLPerf Training v6.1 categories and made its first appearance in MLPerf Endpoints, where it demonstrated significant efficiency gains. Vera Rubin NVL72 systems showed up to double the performance per GPU compared to the previous generation and maintained high efficiency when scaled to thousands of GPUs. The benchmarks covered a range of AI tasks, including large language models and agentic inference workloads.

Key details

  • The system achieved over 10x throughput per watt in agentic AI inference compared to the previous generation.
  • NVIDIA and Microsoft Azure both demonstrated training at a scale of 10,240 GPUs in this MLPerf round.
  • NVIDIA was the only platform with results on every MLPerf Training v6.1 benchmark.

Why it matters

Organizations running large-scale AI workloads, such as cloud providers and research institutions, can now train and deploy models more quickly and efficiently using the Vera Rubin platform. This reduces both the time and energy required for AI development, enabling more users to access advanced AI capabilities at lower operational costs. The improvements in efficiency and scalability directly affect the capacity of data centers to serve more users and handle more complex AI tasks.


More in Machine Learning