Why Your GPU Won't Work in Your Server: 8 Installation Problems and the Fix for Each (Dell PowerEdge & HPE ProLiant)

A data-center GPU that won't work in a server is almost never a bad card. It's a server that hasn't been told, cabled, cooled or powered for that card. Dell PowerEdge and HPE ProLiant machines can run an NVIDIA A40, A100, V100 or T4 well, but only once eight separate conditions are met, and the symptom of a missed one rarely points at the cause. This guide works backwards from what you're seeing to what's wrong.

Start With the Symptom

What you see	Most likely cause	Jump to
Server won't POST, or hangs at boot with the card in	BIOS isn't mapping the card's large BAR above 4 GB	Problem 1
Card missing from lspci, iDRAC or iLO	Slot wired to an empty CPU socket, or no auxiliary power	1, 5
Cable "almost" fits, or the card never lights	PCIe 8-pin where the card needs EPS-12V (or the reverse)	Problem 2
Fine at idle, slow after 20 minutes of load	Passive card starved of airflow	Problem 3
Lid won't close, card won't seat	Card length, width, heat sinks or riser	Problem 4
Server powers off or loses redundancy under load	Power supplies sized for the old role	Problem 6
Second card misbehaves	Mixed GPU models, or slot/lane limits	Problem 7
nvidia-smi says "No devices were found"	Driver, Secure Boot or nouveau	Problem 8

Problem 1: The Server Won't Boot, or the Card Isn't There

Data-center GPUs ask the system for a very large window of memory-mapped address space. NVIDIA's A40 product brief lists a 64 GiB BAR1 in its default display-off mode. A BIOS that only maps devices below 4 GB can't place that window, and the result ranges from a card that silently doesn't enumerate to a server that won't finish POST.

The fix is to enable 64-bit memory-mapped I/O above 4 GB. Supermicro and consumer boards call it "Above 4G Decoding"; on Dell PowerEdge it lives under Integrated Devices as Memory Mapped I/O above 4GB; on HPE ProLiant Gen10 it's PCI Express 64-Bit BAR Support in the system utilities. Boot in UEFI mode, not legacy BIOS, and update the BIOS before you start — GPU support lists are added in firmware releases.

If the card still doesn't appear, look at which slot it's in. On two-socket servers, the riser slots on one half of the board are wired to the second processor. With one CPU installed, those slots are electrically dead. Dell goes further on the R740xd and requires both sockets populated for any GPU configuration. A single-socket server isn't a GPU candidate until you add a CPU.

Problem 2: The Power Cable Fits — Almost

This is the expensive one. NVIDIA data-center cards take a CPU-style (EPS-12V) 8-pin auxiliary connector. Workstation cards take PCIe 8-pin and 6-pin graphics power, or the newer 16-pin connector. The EPS and PCIe 8-pin plugs are the same size and shape, but the 12 V and ground pins sit in different positions and the keying differs. The keying exists to stop you. If a plug only goes in with force, it's the wrong plug, and powering on can damage the card, the riser or both.

Card	Board power	Auxiliary connector
NVIDIA A40	300 W	1× CPU (EPS-12V) 8-pin
Tesla V100, P100, P40 (PCIe)	250 W	CPU (EPS-12V) 8-pin
A100 80GB PCIe	300 W	CPU (EPS-12V) 8-pin
RTX 6000 Ada	300 W	1× PCIe CEM5 16-pin
Tesla T4 / L4	70 W / 72 W	None — powered from the slot

The card decides the cable, not the server. Read the auxiliary power line on the card's spec sheet before you order anything.

HPE ProLiant DL380 / DL385 Gen10: data-center cards use the 8-pin keyed cable from kit 871829-B21, which is HPE part 869820-001. The neighbouring kit, 871830-B21, is for PCIe-style workstation power and is not a substitute. The HPE 869820-001 cable is $49 at Resilient Tec, in stock, sold on its own so you don't buy the whole kit to power one card.

Dell PowerEdge R740 / R740xd: data-center cards use Dell 4VPD3 (04VPD3), an 8-pin riser-to-EPS-12V cable. The PCIe-style cable, TR5TP (8-pin to 6-pin + 6+2-pin), is for workstation cards. Our R740xd GPU Ready builds ship with the PCIe-style cable, so an A40 added to one needs a 4VPD3 as well. The Dell 4VPD3 cable is $47 at Resilient Tec, one per card, shipping in about 7 days.

The full HPE breakdown, including the unkeyed 871828-B21 kit, is in 871828, 871829, 871830: which HPE GPU cable kit you need and 871829-B21 vs 871830-B21.

Problem 3: Fast at First, Then It Crawls

The A40, A100, V100, T4 and L4 are passive. There's no fan on the card. It's a heat sink in a shroud, designed to sit in a server's front-to-back air path and let the chassis fans do the work. Put one in a server still running its standard fans and it idles normally, climbs past 80 °C under sustained load, and clocks itself down. Nothing errors. The job just takes three times as long, and the usual next suspect is the driver or the model.

Three things fix it:

• The high-performance fan kit. Dell lists it as a requirement for GPU configurations in the R740xd, not an option. HPE's DL38x Gen10 equivalent is 867810-B21, paired with the 826706-B21 high-performance heat sink kit.
• The GPU air shroud and baffles. They force air through the card's fins instead of around them. The mylar strips in Dell's enablement kit seal bypass paths; they aren't packing material.
• Leave the fan logic alone. When iDRAC sees a PCIe card it doesn't recognise as Dell-qualified, it applies a default "third-party card" cooling response and raises fan speed. Dell's own cooling white paper describes it. Plenty of forum guides show how to switch it off because it's loud. On a fanned NIC that's harmless. On a passive GPU it removes the airflow the card depends on. On HPE, if a passive card runs hot, raise the BIOS Thermal Configuration from Optimal Cooling to Increased Cooling rather than lowering anything.

Inlet temperature matters too. Dell caps an R740xd with a GPU and higher-TDP CPUs at 30 °C inlet. A rack that was fine for a storage node can be too warm for the same chassis with a 300 W card in it.

Problem 4: It Doesn't Fit

Most data-center GPUs are full-height, full-length and double-width; the A40 is 10.5 inches long and two slots wide. Four physical traps:

• Riser. Standard risers in many servers are single-width or low-profile. A double-width card needs a GPU riser with an x16 slot and room above it.
• Heat sinks. In the R740xd, standard CPU heat sinks foul the GPU shroud. The GPU kit's low-profile heat sinks go on both sockets.
• x16 mechanical isn't x16 electrical. Many slots that accept an x16 card are wired x8. The card will run, at half the host bandwidth. Check the riser map, not the slot.
• Support. A full-length card needs its rear end supported by the shroud or a retention bracket, especially in a server that ships on a truck.

Problem 5: No Auxiliary Power at All

A PCIe slot delivers up to 75 W. That's why a 70 W T4 or 72 W L4 needs no cable, and why a 250–300 W card that has its slot but not its auxiliary cable will either not enumerate or come up in a low-power state. If the riser has no GPU power header, it isn't a GPU riser, whatever it looks like.

If you'd rather not deal with auxiliary power at all, slot-powered cards remove the problem. Our R740xd – Dual T4 Build ($4,693 at Resilient Tec) runs two 70 W Tesla T4s with no GPU power cables.

Problem 6: The Power Supplies Were Bought for a Different Job

Most used 2U servers on the secondary market left their first owner as storage or virtualization nodes with 750 W supplies. Dell sets an 1100 W floor for GPU configurations in the R740xd, and notes they may run in non-redundant mode. Add it up: two CPUs, memory, a full drive bay, a RAID controller, a NIC and a 300 W card. A pair of 750 W supplies in that box isn't redundancy. It's one supply and a spare that can't carry the load alone. 1600 W supplies put genuine redundancy back, which is why every GPU build we ship leaves with them. HPE offers 1600 W supplies for the same reason.

Problem 7: The Second Card

Dell requires every GPU in an R740xd to be the same type and model. You can't pair an A40 with a V100 to save money. The R740xd tops out at three double-width or six single-width cards, and every additional 300 W card turns a comfortable thermal design into a tight one, especially in the 24-bay chassis where the drives draw from the same front panel. Two NVLink-bridged A40s need adjacent slots that match the bridge spacing. Plan the slot layout before buying the second card.

Problem 8: The Hardware Is Fine, the OS Isn't

• nouveau loaded first. Many Linux installs bind the open-source nouveau driver to the card. Blacklist it and install NVIDIA's data-center driver.
• Secure Boot. With Secure Boot on, an unsigned NVIDIA kernel module won't load. Enroll the signing key, or use your distribution's signed driver packages.
• "My A40 has no video." Correct. NVIDIA ships the A40 configured for virtualization and compute with its three DisplayPort outputs disabled. They can be enabled with NVIDIA's display-mode tool, which also shrinks the BAR1 window. Use the iDRAC or iLO console for the server's own display.
• vGPU is licensed separately. Bare-metal CUDA, inference and rendering need only the driver. Splitting one card across virtual machines needs NVIDIA vGPU software and a license.

Build It Yourself, or Buy It Built?

Done piecemeal from the secondary market, a GPU server is a chassis, a card, an enablement kit, two low-profile heat sinks, a fan kit, two 1600 W supplies, possibly a second CPU, the right power cable, rails, and the labor to find out which of those a seller left out. Each is a separate listing with its own condition claim and return window.

Our refurbished R740xd – A40 Build ($10,959 at Resilient Tec) arrives with all eight problems above already solved, burned in, with a lifetime warranty. If you'd rather choose the card, the R740xd – GPU Ready ($2,995) has the enablement kit, fan kit, low-profile heat sinks and 1600 W supplies in place; add a 4VPD3 cable ($47) for each A40 or Tesla-class card. For the full Dell part list, see every part Dell doesn't tell you you need for an A40 in an R740xd. Already own the server? Send us the configuration and we'll quote the missing parts, not a whole new machine.

FAQ

Why is my GPU not detected in my Dell or HPE server?

The four usual causes, in order: the card is in a riser slot wired to an empty second CPU socket; the BIOS is not mapping 64-bit memory above 4 GB, which data-center cards with large BARs (the A40 asks for 64 GiB) need; the auxiliary power cable is missing or wrong; or the OS is loading the open-source nouveau driver instead of NVIDIA's data-center driver. Check them in that order before suspecting the card.

Can I use a PCIe 8-pin cable on an NVIDIA A40, A100 or V100?

No. NVIDIA data-center cards such as the A40 take a CPU-style (EPS-12V) 8-pin auxiliary connector. It looks almost identical to a PCIe 8-pin, but the 12 V and ground pins are in different positions. Use the server vendor's GPU cable for that card, or the adapter the card's vendor specifies. If a plug only goes in with force, it is the wrong plug.

Which power cable does an NVIDIA A40 need in a Dell R740xd?

The Dell 4VPD3 (04VPD3), an 8-pin EPS-12V GPU power cable, one per card. Resilient Tec sells the Dell 4VPD3 cable for $47, available to order with about a 7-day lead time. It is not the TR5TP PCIe-style cable that workstation cards use. (Prices checked October 12, 2026.)

Why does my passive GPU throttle in my server?

Passive cards such as the A40 have no fan and rely completely on chassis airflow. Throttling almost always means the server is missing its high-performance fan kit or GPU air shroud, the inlet air is too warm, or someone turned down the fans. On Dell PowerEdge servers a common culprit is disabling the third-party PCIe card cooling response in iDRAC to quiet the fans.

Do I need two CPUs to run a GPU in a server?

Often, yes. On two-socket servers some riser slots are wired to the second CPU and are dead with one CPU installed, and Dell requires both sockets populated for GPU configurations in the R740xd. Check the riser map for your chassis before you buy the card.

What power supply does a GPU server need?

Size for the whole system, not the card. Dell sets an 1100 W floor for GPU configurations in the R740xd; for a dual-CPU chassis with a 300 W GPU, 1600 W supplies are what give you genuine redundancy rather than a nominal pair.

Why are the display ports on my NVIDIA A40 not working?

They are disabled by default. NVIDIA ships the A40 configured for virtualization and compute, with its three DisplayPort outputs turned off. They can be switched on with NVIDIA's management tooling, which also changes the card's BAR1 size.

How much does a GPU-ready server cost?

At Resilient Tec, the refurbished R740xd – GPU Ready (no GPU) is $2,995, the R740xd – A40 Build is $10,959 and the R740xd – Dual V100 Build is $5,849. Each ships with the GPU enablement kit and high-performance cooling installed, tested, with a lifetime warranty. (Prices checked October 12, 2026.)

Where can I buy an NVIDIA A40 and the cable to power it?

Resilient Tec sells the NVIDIA A40 48GB for $5,490 refurbished or $6,495 new, and the HPE 869820-001 GPU power cable for $49 for HPE ProLiant DL380 and DL385 Gen10 servers. Both are in stock at Resilient Tec with a lifetime warranty. (Prices checked October 12, 2026.)

Ready to Buy?

Every server below is a refurbished Dell PowerEdge R740xd with the GPU enablement kit, high-performance fans, low-profile heat sinks and 1600 W supplies installed and tested. Prices below checked October 12, 2026.

• R740xd – A40 Build — $10,959 — 1× NVIDIA A40 48 GB, ready to run.
• R740xd – Dual V100 Build — $5,849 — 2× Tesla V100 for workloads that split across two cards.
• R740xd – Dual T4 Build — $4,693 — 2× Tesla T4 16 GB, slot-powered, no GPU cables needed.
• R740xd – GPU Ready (no GPU) — $2,995 — no card; bring your own GPU.
• R740xd – GPU Ready, Gold 6248 / H740p (no GPU) — $3,493.33 — no card; Xeon Gold 6248 and PERC H740p.
• NVIDIA A40 48GB card — $5,490 refurbished / $6,495 new — for a chassis that's already enabled.
• Dell 4VPD3 8-pin EPS-12V GPU power cable — $47 — ships in about 7 days; powers one A40 or Tesla-class card in a Dell R740 / R740xd.
• HPE 869820-001 8-pin keyed GPU power cable — $49 — powers one data-center GPU in an HPE DL380/DL385 Gen10.
• Dell H4X6X ReadyRails II 2U sliding rail kit — $139 — sliding rails for the R740xd and related 2U PowerEdge chassis.

Different GPU, more drives, or several nodes? Request a quote — builds to order typically ship in 1–2 business days.

You Might Also Need

• Putting an NVIDIA A40 in a PowerEdge R740xd — the complete Dell bill of materials
• What can you actually run on a 48 GB A40? — model sizes and tokens per second
• Custom Dell R740xd GPU servers — build options and real prices
• Which HPE GPU cable kit do you need? — 871828, 871829 and 871830 compared
• GPU servers & on-prem AI guide and enterprise servers guide
• What cable do I need? — networking the node once it's racked

Sources: NVIDIA A40 Product Brief and A40 tech specs (power connector, BAR1 sizes, display-off default); NVIDIA Quadro/RTX power guidelines DA-07261 (CEM5 16-pin on RTX 6000 Ada); Dell, "PCIe Card Cooling with Dell EMC PowerEdge Servers" and KB 000135682 (third-party PCIe card cooling response); Dell PowerEdge R740xd GPU installation guidance (enablement kit, 1100 W floor, both sockets, 30 °C inlet, identical GPUs); NVIDIA Developer Forums, "A100 PCIe isn't recognized by BIOS" (Above 4G decoding and MMIO).