NVIDIA Unveils 'Rubin' Architecture, Surpassing Blackwell

The Next Leap in AI Hardware

Just as the industry finishes adopting the Blackwell architecture, NVIDIA has announced its successor: Rubin. Named after astronomer Vera Rubin, the new architecture is explicitly designed to handle the training and inference of multi-trillion-parameter AI models with unprecedented efficiency.

Breakthrough Memory Bandwidth

The Rubin GPUs feature a completely redesigned memory subsystem utilizing HBM4 (High Bandwidth Memory 4), effectively doubling the memory bandwidth compared to previous generations. This drastically reduces the bottleneck associated with loading massive AI models into memory, allowing for much faster inference times.

CEO Jensen Huang emphasized that Rubin is not just a GPU, but a complete datacenter solution. "The scale of AI requires a full-stack approach. Rubin integrates networking, computing, and software into a single unified engine," Huang stated.