How to choose a data-center GPU
Start from the workload. Model size and batch concurrency set the VRAM requirement — 48 GB on the L40S, 141 GB on the H200, 192 GB on the MI300X — and training versus inference decides how much HBM bandwidth you actually need. Inference fleets often get better value from more, smaller cards than from fewer flagships.
Choose the form factor next. SXM and OAM modules are baseboard-bound for HGX and UBB platforms and are not electrically or mechanically standalone cards; standard servers take PCIe variants — H100 PCIe, H200 NVL, L40S, L4 or MI210. Confirming tray and baseboard compatibility up front prevents the most expensive mismatch in GPU procurement.
Check the power and cooling envelope before ordering: 700 W SXM parts need the right manifold and PSU headroom, and dense PCIe builds need confirmed airflow. For 8-GPU nodes we kit the complete baseboard, host CPUs, memory, storage and NICs as one BOM with one warranty window.
Budget the fabric with the silicon — NVLink and NVSwitch topologies and InfiniBand east-west networks change the node design — and note that export-control rules apply to some SKUs and destinations. We flag restrictions at quotation stage and offer compliant alternatives where a part cannot ship.