Supermicro NVIDIA Vera Rubin NVL72 Rack

Overview

The Supermicro NVIDIA Vera Rubin NVL72 represents an enterprise-grade rack-scale compute solution engineered for extreme exascale artificial intelligence and high-performance computing workloads. Operating as a unified computational engine, the rack integrates 72 NVIDIA Rubin GPUs alongside 36 NVIDIA Vera CPUs distributed across 18 individual 1U compute trays, with each tray containing four Rubin GPUs and two Vera CPUs. Inter-chip communication is orchestrated through nine sixth-generation NVIDIA NVLink switch trays, delivering an unprecedented aggregate scale-up bandwidth of 216 TB/s and providing access to 20.7 TB of high-bandwidth HBM4 memory paired with up to 54 TB of system-level LPDDR5X memory.

From a thermal and structural standpoint, the platform relies on Supermicro’s DLC-2 Direct Liquid Cooling infrastructure, incorporating liquid-cooled rack manifolds, custom cold plates, and optional rear-door heat exchangers backed by a 1.8MW in-row cooling distribution unit (CDU) featuring N+1 redundancy. Because this platform represents newly announced hyperscale hardware without public post-launch physical teardowns, acoustic SPL measurements and internal chassis-level component dampening analyses remain pending retail physical teardown and facility deployment benchmarks. Nevertheless, the physical form factor, confirmed chassis materials, manifold routing, and redundant plumbing establish a robust standard for next-generation AI infrastructure.

Technical Specifications