NVIDIA Vera Rubin NVL72 Achieves Top Performance in MLPerf

In the world of AI infrastructure, system performance, efficient scaling, and continuous software optimization are crucial elements that determine the economic impact of AI inference. Higher system performance results in more generated tokens and thus higher revenue. Concurrent scaling means that throughput increases proportionally as hardware is added, minimizing resource consumption. Furthermore, continuous optimization ensures that organizations derive more value from their infrastructure investments. The underlying factor for all these elements is platform fungibility, where the same infrastructure can run any model and workload, from training to inference.

The NVIDIA platform is specifically designed to perform optimally in these areas, as confirmed by the recently published results of MLPerf Inference v6.1. The latest Vera Rubin NVL72 delivers up to 3.7 times better throughput in its first MLPerf Inference preview compared to the GB300 NVL72. Additionally, four racks of the GB300 NVL72 with 288 GPUs achieve a scaling efficiency of 99%, meaning that throughput increased almost linearly, even from the baseline of a single rack.

The performance is further driven by continuous software optimizations, yielding up to 1.6 times higher performance compared to earlier versions. This process of optimal software development has also led to further improvements, even after the submission of v6.1. The use of innovative techniques such as disaggregated service and large-scale expert parallelism maximizes the efficiency of the models powered by Vera Rubin NVL72.

Notably, concurrent scaling efficiency, as demonstrated by the DeepSeek-R1 benchmark, shows that additional GPUs actually lead to a proportional increase in throughput. This proves that the architecture, interconnectivity, and software must work well together to achieve sustainable infrastructure performance. NVIDIA's results not only underscore the company's progress in AI infrastructure but also highlight the crucial role of partners in cloud and edge technology ecosystems.

Read the full article from Nvidia.