GPU Servers: What Actually Fits in 2U and 4U

Most GPU deployments that go wrong go wrong on geometry and airflow, not on compute. A card that fits the slot electrically may not fit the chassis physically, may not get the power connector it needs, and may not survive a warm aisle. Before choosing an accelerator, it is worth working through the four constraints that actually decide the answer: height, power, cooling and PCIe topology.

The numbers that decide the fit

Rack units are 44.45 mm each. A full-height PCIe card is roughly 111 mm tall at the bracket, and the standard data-center GPU length is 267 mm (10.5 inches). A double-width card occupies two slot positions, about 40 mm.

Put those together and the consequence is immediate: 2U of external height is 88.9 mm, which is already less than the 111 mm a full-height card needs before you subtract the chassis itself. A full-height GPU cannot stand upright in a 2U server. It has to lie flat, parallel to the motherboard, on a riser card. In 1U (44.45 mm) even a low-profile card at 68.9 mm needs a riser. Only at 4U — 177.8 mm — can a full-height card sit vertically in a slot the way it would in a workstation, which is why dense GPU chassis are almost always 4U or taller.

The second consequence is that riser availability, not slot count, sets the GPU count in a general-purpose server. Fitting a GPU riser assembly in a 2U typically consumes the slots your NICs and HBAs were going to use, and the GPU-capable riser is usually a specific part number, not the one the server shipped with.

Chassis	External height	Card orientation	Realistic accelerator count	Typical role
1U general purpose	44.45 mm	Low-profile only, on a riser	1–3 single-width low-profile cards	Inference, video transcode, light vGPU
1U purpose-built GPU chassis	44.45 mm	Riser, dedicated airflow tunnels	Up to 4 double-width	Density where noise and heat are not your problem
2U general purpose	88.9 mm	Riser, card lies flat	2 double-width, 3 on some models	Virtualization host with accelerators
4U	177.8 mm	Vertical, direct or short riser	8–10 double-width, or an SXM baseboard	Training, HPC, dense VDI

Power: the card, the supply, and the circuit

A PCIe slot delivers up to 75 W. Anything beyond that needs auxiliary power: a 6-pin adds 75 W, an 8-pin PCIe adds 150 W, and the newer 12V-2x6 connector carries up to 600 W. This is where the card catalogue splits neatly into two groups.

Cards that draw 75 W or less need no auxiliary cable at all, which is what makes them the practical choice for 1U and for 2U servers that were never specified with a GPU power harness. The NVIDIA T4 at 70 W and the L4 at 72 W are both single-slot, low-profile and bus-powered; the A2 sits in the same bracket at a configurable 40–60 W. The A10 is a step up at 150 W, single-slot but full height and full length, so it needs both a cable and a riser.

Above that, everything is double-width and cabled: A30 at 165 W, A16 and A100 40 GB at 250 W, A40 and A100 80 GB at 300 W, H100 PCIe and L40S at 350 W. SXM modules go further still — 400 W to 500 W for A100 SXM4 depending on the baseboard, and up to 700 W for H100 SXM5 — which is precisely why they live on baseboards in 4U and larger chassis with dedicated cooling.

Three power details cause real damage. First, the auxiliary connector on the server side is often a vendor-specific cable, and the 8-pin EPS connector used for CPU power is not interchangeable with the 8-pin PCIe power connector; the shells look almost the same, the keying and the pinout do not match. Never adapt one to the other. Second, many high-output server power supplies deliver their full rated wattage only on 200–240 V input and derate significantly on 100–127 V — a detail that has stranded more than one lab build. Third, do the rack arithmetic: eight 350 W cards is 2.8 kW of GPU alone, and once CPUs, fans and drives are added a single 4U node lands around 4 kW. A rack provisioned for 6–8 kW takes one of those nodes comfortably and a second only at the very top of the budget — not four.

Cooling: passive cards depend entirely on the chassis

Data-center GPUs are passively cooled. They have a heatsink and no fan, and they assume a chassis pushing high static pressure air through them front to back. Drop one into a tower or an unducted chassis and it will throttle within minutes, then shut down. Conversely, a workstation card with its own axial fan can perform badly in a rack server, because its fan fights the chassis airflow instead of adding to it. Match the cooling model of the card to the cooling model of the box.

Server vendors sell GPU enablement kits for exactly this reason, and they are not optional decoration. A kit typically includes higher-performance CPU heatsinks, high-static-pressure fans, air baffles and a GPU shroud that seals the airflow path. Populating GPUs also usually lowers the server's maximum supported inlet temperature and requires full fan redundancy, so a room that was comfortable running general compute may not be comfortable running the same chassis with accelerators in it.

Depth is the quiet constraint. A 267 mm card plus riser plus power cabling makes GPU-configured chassis among the deepest in any catalogue. Measure the rack, including the rear door and any cable management arms, before ordering.

PCIe topology, NVLink, and NUMA

Every double-width accelerator wants a full x16 link. In a two-socket 2U, the risers are wired to specific CPUs, so a GPU on riser 2 is remote from a NIC on riser 1 — and traffic between them crosses the inter-socket link. Pin processes and network queues to the socket that owns the card, or accept the penalty.

Dense 4U systems solve the lane shortage with onboard PCIe switches that fan a smaller number of CPU lanes out to eight or more cards. That enables fast peer-to-peer transfers between GPUs under the same switch, which is what multi-GPU training wants, in exchange for oversubscribing the uplink to the CPU. If your workload streams data from storage to every GPU continuously, that uplink is the number to check.

NVLink bridges on PCIe cards need two GPUs in adjacent slots at exactly the right slot spacing, which is frequently impossible on 2U riser layouts and one of the more common late surprises in a build. For workload partitioning rather than aggregation, MIG on A100 and H100 splits a single GPU into as many as seven isolated instances; for VDI, remember that vGPU is a licensed software layer on top of supported cards, not a property of the hardware alone. The accelerators listed in the GPU collection note the form factor and TDP for each, which is the fastest way to filter out cards your chassis cannot take.

Where to go from here

Start from the chassis you have or the density you need, then work down through height, TDP, connector and airflow. If you need two accelerators next to a virtualization workload, a 2U with the right riser and enablement kit is the efficient answer; if you need eight, specify the 4U and the power to feed it. Browse the GPU options to see what is available by form factor, or open Build Your Server to put a chassis, riser configuration, power supplies and accelerators together and send the configuration over for a quote. Tell us the rack depth and the circuit you are feeding it from, and we will check the build against both before quoting.