Supermicro is engineering its NVIDIA Vera Rubin NVL72 with new DCBBS liquid-cooling components to fully support the power and thermal envelope at rack and cluster scale. This includes the manufacturing of optimized NVIDIA MGX racks, in-rack or in-row CDU, RDHx and L2A sidecar to streamline production and deployment of the rack-scale AI supercomputer at scale. The Vera Rubin NVL72 operates as a single rack-scale accelerator, unifying six co-designed chips — Rubin GPU, Vera CPU, NVLink 6, ConnectX-9, BlueField-4, and Spectrum-X — to deliver 3.6 Exaflops of inference, 75TB of fast memory, and 1.6 PB/s of HBM4 bandwidth, targeting up to 10x the throughput per watt and one-tenth the token cost compared to NVIDIA Blackwell.  

The 2U HGX Rubin NVL8 system provides the densest and most flexible HGX platform — and the first HGX platform to offer greater flexibility in CPU selections including NVIDIA Vera CPUs alongside next-generation AMD and Intel x86 processors. Built on the NVIDIA MGX rack architecture with Supermicro’s blind mate busbar and manifold for tool-free rack integration, it gives customers the freedom to pair eight Rubin GPUs with the CPU platform that best fits their workload and software stack.