NVIDIA has unveiled the Vera Rubin AI platform, a system designed to deliver supercomputer-level performance within a single server rack. This architecture allows data centers to achieve over 7 ExaFLOPS of AI processing power without requiring massive physical footprints. This architecture allows organizations to expand high-performance computing infrastructure more efficiently than with previous hardware generations.
New architecture packs supercomputer performance into dense liquid-cooled racks
The platform integrates the new Rubin GPU with the Vera CPU, linking them through NVIDIA's proprietary NVLink-C2C interconnect technology. This combination forms the core of the NVIDIA Vera Rubin NVL4 solution, which also incorporates the ConnectX-9 SuperNIC for networking and the BlueField-4 DPU for data processing tasks. These integrated components collaborate to handle the heavy computational demands characteristic of modern artificial intelligence tasks.
Specifications
- GPU Count per Rack: Up to 144 GPUs
- AI Performance: Over 7 ExaFLOPS
- FP64 Performance: 5 PetaFLOPS
- CPU: Vera CPU
- GPU: Rubin GPU
Each server rack can house up to 144 GPUs, providing the density needed for large-scale model training and inference. The system achieves 5 PetaFLOPS of native FP64 performance, supporting scientific computing requirements alongside AI tasks. To manage the heat generated by this density, the NVL4 solution utilizes a direct liquid cooling system.
The Vera Rubin platform is available globally starting in the fourth quarter of 2026. Major hardware partners including Bull, Dell, GIGABYTE, HPE, and Supermicro will offer the system to enterprise customers. Research institutions such as LANL, LRZ, and LBNL have already adopted the platform for their computing needs.
Discussion
0 comments
Log in to join the thread with a thoughtful take, question, or correction.