Question

A data center team is deploying several NVIDIA HGX B200 8-GPU nodes into an existing colocation facility. Each rack in the facility is rated for a maximum of 30 kW, and the team wants to place as many GPU nodes per rack as thermal limits allow while keeping deployment simple. A single HGX B200 node draws roughly 14.3 kW at full load. The facility currently supplies only chilled-air cooling with no rear-door heat exchangers or direct-to-chip liquid loops available. Which action should the team take to deploy these nodes reliably?

Accepted Answer

Limit each rack to two HGX B200 nodes and, if higher density is required, add liquid cooling (rear-door or direct-to-chip) to exceed the air-cooling thermal envelope

Answer

Install three HGX B200 nodes per rack and rely on higher fan speeds to dissipate the ~43 kW load

Answer

Place four HGX B200 nodes per rack since GPU power draw is only relevant during training, not at idle

Answer

Deploy the nodes but disable NVSwitch to lower total power below the rack's 30 kW rating