Copied


NVIDIA Vera Rubin Platform Leads MLPerf Training and Endpoints Benchmarks

Terrill Dicki   Nov 17, 2026 02:24 0 Min Read


NVIDIA's Vera Rubin platform has made a commanding debut in the latest MLPerf benchmarks, achieving top performance across training, post-training, and agentic inference workloads. According to results released on November 16, 2026, NVIDIA swept all MLPerf Training v6.1 benchmarks and delivered over 10x generational gains in throughput per watt in its first MLPerf Endpoints submission.

Breakthrough Performance in MLPerf Training v6.1

MLPerf Training benchmarks measure the speed at which systems train AI models to a target level of accuracy. NVIDIA's Vera Rubin NVL72 platform demonstrated up to 2x more performance per GPU compared to the previous-generation NVIDIA GB300 NVL72 system. The platform achieved results across all tested workloads, including large language models, image generation, recommender systems, and reinforcement learning post-training.

In large-scale testing, the Vera Rubin NVL72 system scaled from 256 to 1,024 GPUs with 97% efficiency on the DeepSeek-V3 benchmark, reducing training time from 16 minutes to just four. NVIDIA also showcased the largest-scale submission of this round, running the benchmark on 10,240 GPUs using GB300 systems. Efficient scaling was enabled by NVIDIA's NVLink networking within systems and Quantum-X800 InfiniBand and Spectrum-X Ethernet for data center-wide connectivity.

Agentic AI Leadership in MLPerf Endpoints

In its MLPerf Endpoints debut, the Vera Rubin NVL72 platform delivered more than 10x the throughput per watt of the GB300 NVL72 system for agentic AI workloads. These workloads involve AI agents performing complex tasks, such as planning trips or resolving IT issues, by reasoning, calling tools, and verifying results.

For example, in the Kimi K3 benchmark for agentic workflows, the Vera Rubin NVL72 system achieved a system throughput of 41,414.2 tokens per second while maintaining an average interactivity of 50.4 tokens per second per user, with an operating power of 100.89 kW. In comparison, the GB300 NVL72 system achieved 2,210.4 tokens per second at 60 kW. These results highlight the Vera Rubin platform's ability to serve more users while maintaining faster response times.

Extensive Ecosystem and Broader Implications

NVIDIA's MLPerf Training submissions included contributions from 20 partners, such as Microsoft Azure, Google Cloud, Dell Technologies, and Oracle, showcasing the adaptability of NVIDIA's programmable platform. Beyond MLPerf, the Vera Rubin platform has also demonstrated leading performance in independent benchmarks like SemiAnalysis AgentX and Artificial Analysis AgentPerf.

With continuous optimization of its CUDA-powered software stack, NVIDIA aims to further enhance the performance of the Vera Rubin platform, ensuring that customer investments deliver increasing value over time.


Read More