Will This GPU Fit My Server? Slots, Power Connectors and Airflow

"Will this GPU fit?" is really four questions: will it physically go in, can the server power it, can the server cool it, and will the platform recognize it. A card that passes three of the four still fails. The good news is that each check can be answered from documents you already have: the GPU's product brief and your server's technical guide or GPU configuration matrix. Here is how to work through them.

Will this GPU physically fit in my server?

Check three dimensions, and check them against the riser slot the card will actually use, not the server as a whole.

• Height. Low-profile cards fit short slots found in 1U and many 2U servers. Full-height cards need a full-height riser. The L4 is a single-slot, low-profile card.
• Length. Full-length cards run most of the depth of the riser area and usually need a support or retention bracket at the far end. The H100 PCIe and A100 40GB PCIe are full-height, full-length, dual-slot cards, listed with others on the FHFL dual-slot page.
• Width. A dual-slot card covers the neighboring slot. In a riser with slots at single spacing, one dual-slot card can consume two positions, and the second position may be the one your network card was planned for.
• Obstructions. Cable routing, memory heat spreaders, internal drive cages and the chassis lid can all collide with a full-length card. Server guides list which slots accept full-length cards for this reason.
• Electrical width and generation. A physical x16 slot wired at x8 will run most cards, but at half the link bandwidth. PCIe is backward compatible, so a PCIe 5.0 card such as the H100 PCIe works in a PCIe 4.0 slot, at Gen4 link speed.

What power connector does a data center GPU need?

It depends on the card's power draw and generation, so confirm the exact connector in the product brief before you order a cable. The general pattern:

• Up to 75W: slot power only. A PCIe x16 slot can supply up to 75W, which covers cards like the 72W L4. No auxiliary cable is needed.
• Above 75W, older data center generation: many passive data center cards use an 8-pin EPS-12V style connector, the same keying as a CPU power cable, rather than the 8-pin PCIe connector used on desktop graphics cards. The two are not interchangeable.
• PCIe 5.0 generation and high-power cards: the 16-pin 12V-2x6 connector (the revised 12VHPWR). For example, the RTX 5090 in our RTX 5090 GPU Server is listed at 575W TDP on a 600W PCIe Gen5 12V-2x6 connector.
• Server-side cables are specific to the server vendor and riser. The riser end of a GPU power cable often uses a proprietary pinout, so a cable from a different platform, or a desktop adapter, can fit mechanically and still be wrong.

This is the most common reason an otherwise compatible card cannot be installed on delivery day: nobody ordered the right power cable. OEM GPU kits remove that risk. The Dell NVIDIA L4 GPU kit ships with GPU, riser bracket, power cable and shroud for PowerEdge R750, R760, R7525 and XE series, and the HPE NVIDIA L4 kit ships with riser cage, power cable and air baffle for ProLiant DL380, DL385 and Apollo series.

Can my power supplies handle the GPU?

Add the GPU TDP to the server's existing peak load and compare it to the capacity of one power supply, not the total of all of them. With redundant PSUs, the server must be able to run on the surviving supply or supplies if one fails. For example, four H100 PCIe cards at 350W each add 1,400W of GPU load before CPUs, memory, drives and fans. That is why our 4 GPU AI Server uses a representative configuration of four 2700W Titanium supplies in a 2+2 redundant arrangement, rather than the PSUs of a general-purpose 2U server. For very high-power cards such as the 600W RTX PRO 6000 Blackwell 96GB, check the platform's PSU requirement per card before anything else.

Can the server cool it?

Most data center GPUs, including the H100 PCIe, A100 PCIe and L4, are passively cooled: the server's fans must push air through the card's heatsink. A server that is qualified for a 72W low-profile card in its standard fan configuration may need a high-performance fan option or GPU enablement kit for a 250W or 350W card. The passive, directed chassis airflow page lists which cards depend on server airflow. Actively cooled workstation cards are the opposite problem: a blower card in a server built for passive cards can fight the chassis airflow, so use the passive server variant where one exists.

Will the platform recognize the card?

• BIOS: large-memory GPUs usually need Above 4G Decoding (sometimes called MMIO above 4GB) enabled so the card's memory can be mapped. Without it, the system may not boot or the card may not appear.
• BMC and fan control: OEM-qualified cards report temperature to the management controller so fans respond correctly. Unrecognized cards may get a fixed high fan speed on some platforms and too little cooling on others.
• Firmware levels: GPU support is often added in a specific BIOS or BMC release, so update before installing.
• Driver and OS: confirm the GPU driver supports your OS and hypervisor, especially for virtualization or partitioned use.

How many double-wide GPUs can my server take?

The server's GPU configuration guide gives the only reliable number, because four limits apply at once: how many x16 slots sit at double-width spacing, which risers are available, how much PSU capacity is left, and how much airflow the fan wall provides. As a broad pattern, 1U servers usually take only low-profile single-slot cards, general-purpose 2U servers take a small number of double-wide cards with a GPU riser option, and GPU-optimized 4U systems are designed for four or more.

If you need more double-wide GPUs than your current platform allows, a purpose-built chassis is usually the cheaper path than forcing it. Our 4 GPU AI Server is a 4U system taking four GPUs (RTX 5090, RTX PRO or H100 PCIe), the RTX 5090 GPU Server takes four to eight RTX 5090 cards, and the 8 GPU AI Server moves to NVLink and NVSwitch for H100, H200 and B200 class GPUs. Our comparison of 4-GPU versus 8-GPU H100 servers helps decide which way to go.

A GPU fits when size, power, cooling and firmware all check out against the specific slot it will occupy. Three out of four is a return.

Nexus Compute supplies new GPUs and OEM GPU kits through authorized distribution, with the full manufacturer warranty, and checks compatibility against your platform before dispatch, including the power cable and riser. Send us the server make, model and generation, the current riser layout and occupied slots, the installed PSU wattage and redundancy mode, and the GPU and quantity you want, and we will confirm fit and quote within 48 business hours.