Supermicro 2-OU HGX B300 OCP System — 8-GPU Liquid-Cooled AI Node

Specifications

Overview

Supermicro's 2-OU OCP system packs eight NVIDIA HGX B300 GPUs into a half-height node built to the OCP ORv3 21-inch rack specification. Because each node is only 2 OU high, a rack carries up to 18 nodes and 144 total GPUs — GPU density that air cooling cannot sustain at the B300's 1,100 W-per-GPU TDP.

Cooling is therefore integral, not optional: Supermicro uses DLC-2 direct liquid cooling with blind-mate manifold connections so nodes can be slid in and out without draining the loop, and a modular GPU/CPU tray architecture for serviceability. Front-facing OSFP ports with NVIDIA ConnectX-8 SuperNICs provide 800 Gb/s scale-out to Quantum-X800 InfiniBand or Spectrum-X Ethernet fabrics.

For deployment planning, the practical constraint is facility-side: a 144-GPU rack at this TDP requires liquid cooling capacity in the hundreds of kilowatts per rack, which is why ORv3/ORW racks and CDUs are specified alongside the compute.

Key Benefits

144 GPUs per 21-inch rack, 1,100 W per-GPU support with DLC-2 direct liquid cooling, blind-mate manifolds for hot service, and 800 Gb/s scale-out via ConnectX-8 SuperNICs.

Applications

Hyperscale AI training clusters, AI-factory deployments, large-scale LLM training and inference, and OCP ORv3 liquid-cooled data center builds.

Related Products