GPU server

A GPU server is a rack-mounted computer built around multiple accelerators rather than a single CPU doing the work. One or two host CPUs handle orchestration and I/O while four to eight GPUs do the math. High-wattage power supplies and directed airflow hold those GPUs under their thermal limits at sustained load. The board exposes wide PCIe or SXM connectivity so the cards can reach each other and the network at full bandwidth.

A 4U chassis with eight double-width PCIe cards at 350 W each draws roughly 4 kW once CPUs, drives, and fans are counted. It normally carries four power supplies in the 2,000 to 3,000 W class. An eight-GPU HGX system with 700 W SXM modules lands closer to 10 kW and usually needs 8U to fit the baseboard and its cooling. A plain 1U server, for comparison, draws a few hundred watts and holds no GPUs at all.

Start with power and cooling, not GPU count. A rack provisioned for 5 kW per node will not take a fully loaded HGX system without an electrical and cooling upgrade first.

Sources

Last verified August 29, 2026.