Supermicro B300 SYS-822GS-NB3RT

Eight HGX B300 GPUs joined by NVLink and NVSwitch, 288GB per GPU and 2.3TB of HBM3e across the system. Dual Intel Xeon 6700 processors with P-cores sit alongside, and 32 DIMM slots take up to 8TB of DDR5.

The networking is what separates this from a generic GPU server. Eight OSFP 800GbE ports come from NVIDIA ConnectX-8 SuperNICs, so the fabric is provisioned per GPU rather than shared across the chassis. That is the difference between a node that scales into a cluster and one that becomes the bottleneck in it.

The facility question comes before the order, not after it. Six 6600W Titanium supplies in a 3+3 arrangement, 119 kg net, 950 mm deep, air cooled at up to 350W per CPU. The rack needs the airflow and the containment to take that. We confirm the power envelope and the aisle temperature with you before quoting, because on a machine like this the facility usually governs the delivery date rather than the hardware.