Exhibitor login
Nvidia 16 November 2026

NVIDIA Vera Rubin excels in MLPerf performance

NVIDIA Vera Rubin excels in MLPerf performance

The latest MLPerf benchmarks show that the NVIDIA Vera Rubin platform delivers outstanding performance, achieving strong scores in training, post-training, and agentic inference with its debut results. NVIDIA won every benchmark in the MLPerf Training v6.1 and achieved more than 10x generative gains in throughput per watt for agentic AI in its first participation in MLPerf Endpoints. This highlights not only the power of the Vera Rubin but also the potential of continuous software optimization to enhance infrastructure effectiveness over time.

AI agents have become indispensable across various sectors, contributing to code writing, executing business processes, and answering customer inquiries. Efficient training means smarter AI models are available to users faster and can respond more quickly to new data and threats. During the latest benchmarks, the NVIDIA platform was the only one to successfully complete all tests, with impressive results for various AI applications, ranging from large language models to image generation and recommendation systems. Vera Rubin NVL72 even demonstrates up to 2x better performance per GPU compared to previous generations, underscoring NVIDIA's leadership in AI training.

Moreover, NVIDIA demonstrated the scalability of the Vera Rubin by achieving a training efficiency of around 97% on the DeepSeek-V3 benchmark by increasing the number of GPUs from 256 to 1,024, reducing training time from 16 minutes to four minutes. The impressive results are not only crucial for AI development but also for the efficiency of data centers, where the use of NVIDIA NVLink and Quantum-X800 InfiniBand facilitates smooth cooperation among thousands of GPUs. This enables developers to create more advanced models and significantly shortens turnaround times.

Vera Rubin also leads in agentic inference within the MLPerf Endpoints, which is critical for applications requiring multiple steps to successfully complete tasks. With more than 10x throughput per watt compared to previous systems at equal response times, Vera Rubin proves its value in a time when speed and efficiency are essential. The strong collaboration with 20 different organizations underscores the versatility of the NVIDIA platform and the possibilities for innovation within the AI training landscape. This not only demonstrates NVIDIA's leading position in MLPerf but also in other independent benchmarks that now substantiate the effectiveness of Vera Rubin.

Read the full article from Nvidia.