NVIDIA Vera Rubin NVL72 Debuts with Fastest MLPerf Training Times
NVIDIA introduced the Vera Rubin NVL72 architecture in MLPerf Training v6.1, delivering leading time-to-train and demonstrating a mature CUDA software stack.
What changed?
The NVIDIA Vera Rubin NVL72 architecture made its debut in the MLPerf Training v6.1 benchmark. In its first appearance, the Vera Rubin NVL72 achieved the fastest time-to-train results, surpassing previous platforms such as Blackwell Ultra. NVIDIA also reports that the CUDA software stack for Vera Rubin is already highly mature.

Why does it matter to an everyday developer?
Faster time-to-train directly affects the iteration speed for developers working on complex AI agents and models. A mature CUDA stack ensures that developers can leverage the platform's improved performance without significant migration or compatibility hurdles. This can lead to reduced development cycles and more rapid experimentation or deployment of large-scale AI workloads.
What can the developer do now?
Developers using GPU acceleration for training large AI models should monitor ecosystem support and availability for NVIDIA Vera Rubin NVL72. As hardware and cloud providers begin offering this architecture, teams can plan to benchmark workloads or migrate training pipelines to take advantage of the reported speed gains. The mature CUDA stack should simplify this transition for existing CUDA-based projects.
