Graphics Cards

Server GPU selection guide

Data center GPUs differ most in board power and slot width, and NVIDIA's data sheets group them like this:

What a server needs for a data-center GPU

Four things decide whether a server can run a data-center GPU: airflow, auxiliary power, slot space and the maker's support list. NVIDIA's passive data-center cards rely on the server's airflow: its PCIe data sheet for a 250 W dual-slot accelerator describes a passive heat sink that requires system airflow to stay within thermal limits, and its 70 W low-profile card lists a passive thermal solution. Power comes through the slot and, on larger cards, an auxiliary connector, a PEX 8-pin on NVIDIA's 150 W single-slot card and a CPU 8-pin on the 250 W accelerator. The server has to support the card as well: Dell's R740 guide offers three GPUs of 300 W or six of 150 W, only in the 24 x 2.5-inch drive chassis and with at most two on NVMe configurations. Dell's PowerEdge GPU matrix lists the NVIDIA cards each model supports, the R760xa takes up to four 400 W double-width accelerators or twelve 75 W single-width cards, and riser options are on riser cards. A card's maximum power adds to the server's load, so check the supplies too: Dell PowerEdge PSUs and HPE ProLiant PSUs are listed separately.

GPU servers instead of single cards

To buy GPU capacity as a tested system, configure a GPU-ready Dell PowerEdge such as the R760xa or R750xa from Dell PowerEdge servers, or browse our Supermicro GPU servers.

Not sure which GPU fits your server? Match the right accelerator to your chassis, VRAM, and power budget with our buying guides: picking a GPU server for AI training and inference and Dell PowerEdge GPU Server compatibility; for running an LLM on your own servers, see our on-prem LLM hardware guide.

Server GPU questions

Can I put a data-center GPU in any server?

No. The server has to provide the slot width and length, an auxiliary power cable for cards that need one, enough supply capacity and enough airflow across the card, and it has to list the card. Dell's R740 guide, for example, supports GPUs only in its 24 x 2.5-inch drive chassis, with room for three 300 W cards or six 150 W ones, and Dell publishes a GPU matrix of supported NVIDIA cards per PowerEdge model.

Do passive GPUs need special cooling?

They need the server's own airflow. NVIDIA's PCIe data sheet for its 250 W dual-slot accelerator states that the passive heat sink requires system airflow to operate within its thermal limits, and it accepts airflow in either direction across the card. A passive card in a desktop case or an unsupported slot can overheat, so fit it in a server built and listed for it.