Supermicro announced expansions in manufacturing and liquid-cooling capabilities to support the deployment of data center-scale systems optimized for the Nvidia Vera Rubin and Rubin platforms. The company said the enhancements are intended to enable faster delivery of Nvidia's Vera Rubin NVL72 and HGX Rubin NVL8 systems to hyperscale and enterprise customers.
According to a press release, Supermicro highlighted its Data Center Building Block Solutions (DCBBS) approach as a framework for streamlining production, enabling customization, and reducing time-to-deployment for high-density AI infrastructure. The company noted that the expanded manufacturing capacity, combined with liquid-cooling technology, is designed to support first-to-market availability of the flagship Nvidia platforms.
The Nvidia Vera Rubin NVL72 SuperCluster integrates 72 Nvidia Rubin GPUs with 36 Nvidia Vera CPUs, Nvidia ConnectX-9 SuperNICs, and BlueField-4 DPUs, using NVLink 6 and Quantum-X800 InfiniBand or Spectrum-X Ethernet for high-speed connectivity. The system offers 3.6 exaflops NVFP4 performance, 1.4 PB/s HBM4 bandwidth, and 75 TB of memory, with a liquid-cooling stack including in-row Coolant Distribution Units for warm-water operation.
The 2U liquid-cooled HGX Rubin NVL8 system is designed for AI and HPC workloads and supports up to eight GPUs. It delivers 400 petaflops of NVFP4 performance, 176 TB/s of HBM4 bandwidth, 28.8 TB/s of NVLink bandwidth, and 1600 Gb/s of ConnectX-9 networking. The system accommodates x86 CPUs such as Intel Xeon or AMD EPYC processors and includes a high-density 2U busbar option for rack integration with advanced Direct Liquid Cooling technology.
Key features of the Nvidia Vera Rubin platform include NVLink 6 high-speed interconnects for GPU-to-GPU and CPU-to-GPU communication, Nvidia Vera CPUs with 88 cores and 176 threads, third-generation Transformer Engine acceleration, third-generation confidential computing for model and data isolation, and second-generation reliability, availability, and serviceability (RAS) capabilities. Networking support is provided through Nvidia Spectrum-X Ethernet Photonics, offering up to 102.4 Tb/s switching with co-packaged optics. Storage integration is supported through Supermicro Petascale all-flash servers and JBOF systems leveraging BlueField-4 DPUs.
Supermicro stated that its investments in manufacturing and liquid-cooling, combined with the modular DCBBS architecture, are intended to facilitate rapid configuration, validation, and scaling of high-density systems. The company said these measures aim to reduce deployment timelines and support accelerated adoption of the Nvidia Vera Rubin and Rubin platforms across enterprise and hyperscale environments.
Article edited by Jack Wu
