Tesla not detected or no picture - step-by-step diagnosis (Above 4G, BAR/MMIO, power, cooling)

The short answer: if you have installed a Tesla and there is no picture, that is almost certainly normal, because NVIDIA data-centre cards (Tesla P4, P40, P100, V100, T4 and the A2) have no video output. If, however, the system does not see the card in nvidia-smi or the computer does not POST, there is no single magic option to switch on: you have to work through the diagnosis in order: whether the card is visible on the PCIe bus, whether the firmware correctly allocates PCI/MMIO space and the BAR (this is where Above 4G Decoding comes in), whether the driver works, whether the card has the correct power supply (watch the connector pinout) and whether it is overheating as a passive card. Below I break it down in exactly that order.

I write this from the perspective of someone who has unpacked, installed and started dozens of these cards, and who gets the same calls every time: "the card is dead, there is no picture". It is almost never dead. To be upfront: yes, we sell these cards (full disclaimer at the end). But this happens to be a guide that will save you returning a working card and an evening of frustration.

Why the card is "invisible", and what is normal

Let us start with the most common misunderstanding. A Tesla card is a compute accelerator for a server, not a graphics card for gaming. Two things follow from this that always surprise a new user:

• No video connector of any kind. There is no HDMI, no DisplayPort, no DVI. You connect the monitor to the CPU's integrated graphics or to a second, ordinary graphics card. A Tesla will never display a picture, and that is behaviour by design, not a fault. If you were hoping to plug a cable into the Tesla and see the desktop, you will not, on any of these cards.
• The card is a PCIe device and may be detected as early as POST: the firmware enumerates it on the bus. But because a Tesla has no display output, you will not see the BIOS screen on it. The easiest way to confirm the card is present is after the system has started: lspci (visibility on PCI) and nvidia-smi (the driver sees it).

So the first diagnosis is: do you have a picture from the integrated graphics or a second card? Then the Tesla is doing exactly what it is meant to do. The problem only begins when the system does not see it or the computer does not POST. Then you move on to the diagnosis below, in order, without shortcuts.

Above 4G Decoding: PCI/MMIO address space, not "squeezing in VRAM"

This is one of the most important settings for cards with a large BAR, such as the P40 or P100. Let us explain it precisely, because there is a myth about it.

Data-centre cards expose large address windows (BARs) that map their memory. It is not about literally "squeezing 24 GB of VRAM below 4 GB"; it is about PCIe/MMIO address space. On older platforms the firmware can only allocate devices a limited, 32-bit space below 4 GB, and that is too small even for one of these cards: in compute mode the P40 asks for a 32 GB BAR1 window and the P100 for 16 GB. Above 4G Decoding allows the firmware to use address space above 4 GB as well. Without a suitable allocation of address space, a large Tesla may not be initialised correctly: the card is physically in the slot and the system fans are spinning, but the card either does not work or shows an error (code 12, "not enough resources", in Windows; in Linux the card is listed by lspci but its BAR is unassigned and the NVIDIA driver refuses to load).

In practice:

1. Above 4G Decoding → Enabled. For cards with a large BAR this is usually a prerequisite.
2. The names differ between manufacturers: look in tabs such as PCI Subsystem Settings, Advanced, North Bridge, AMD CBS.
3. After the change, save, restart, and only then check lspci/nvidia-smi.

Resizable BAR: a separate feature that depends on the platform

There is an important distinction here that most guides blur: Resizable BAR (ReBAR / Smart Access Memory) is not the same as Above 4G Decoding. Above 4G is responsible for the ability to allocate a large address space above 4 GB. ReBAR is a separate mechanism (changing the size of the BAR window so that the CPU "sees" the card's entire memory through a single window).

The P40 does not require ReBAR. With multi-card problems, check PCI/MMIO allocation, the board firmware and BAR-related messages. ReBAR is not a prerequisite for running the P4, P40, P100, V100 or T4. In compute mode, the P40 has a 32 GB BAR1 aperture and the P100 PCIe has a 16 GB aperture, requiring allocation above 4 GB. The P4 and T4 have a 256 MB BAR1 aperture, so that aperture alone does not require Above 4G Decoding. On some boards the ReBAR switch only appears after Above 4G is enabled, but these are separate functions.

UEFI and CSM

On many modern boards, the configuration related to Above 4G and ReBAR is available, or works correctly, only in UEFI mode with CSM (Compatibility Support Module) disabled. If you have Above 4G enabled and it still does not work, check whether you are running in compatibility mode. If the relevant options are missing from the menu altogether, look at the boot configuration and the documentation for your specific board; some options only appear after switching to pure UEFI.

And the driver: use an NVIDIA driver branch that still supports your card. For the Pascal and Volta cards (P4, P40, P100, V100) the last branch is R580; the 590 branch and newer no longer support them, so after installing an unsupported driver branch, the card may remain visible in lspci while being unavailable through nvidia-smi. On Linux, pin the 580 branch. On Windows, use NVIDIA's data-centre driver for the card rather than a GeForce Game Ready package.

Power: check the pinout for the specific model

The second big trap and a real risk of physical damage, but not in the way it is usually told. It is not the case that "any PCIe cable = death of the card". The real danger is an incompatible pinout.

The power socket on cards such as the P40, P100 or V100 PCIe is described by NVIDIA as CPU 8-pin (the P100 brief says explicitly: "provides a CPU 8-pin power connector"). How to connect it, however, depends on the specific model and the power supply used: NVIDIA lists two supported options: one CPU (EPS) 8-pin cable straight from the power supply, or NVIDIA's adapter (the "CPU 8-pin to PCIe 8-pin dongle"), which plugs into the card's CPU 8-pin socket and is fed by two separate PCIe cables from the power supply. Supported combinations are two 8-pin, two 6-pin, or one of each, but each 6-pin cable must meet the card's requirements: up to 120 W for the P100, or half the average power and half the peak current for the P40. You must not assume that any 8-pin lead will be compatible with the card.

The rule worth remembering is not "PCIe is forbidden", but: do not use a cable with an unknown or incompatible pinout. Forcing in a cable with the wrong pin arrangement can put voltage on the wrong contacts, short the supply and damage both the card and the power supply.

What to do correctly:

• If your power supply has a spare, compatible CPU/EPS 8-pin cable, connect it directly to the card. Otherwise use an adapter for the model, fed by two separate PCIe cables; do not feed both inputs from one cable. Check the card's cable-capacity requirements before using 6-pin cables.
• With a modular power supply, use only the cables that came with that power supply: the pinout on the PSU end is not standardised between manufacturers.
• Make sure the adapter is described as being for this model: not every "8-pin to 8-pin" has the correct pinout.
• Any doubt at all about the cable? Stop and ask before you plug it in. An e-mail costs less than a burnt-out card.

Not every Tesla needs auxiliary power at all; it depends on the model:

Model	Auxiliary power	Notes
Tesla P4	none, from the PCIe slot (50 W or 75 W version)	passive, needs airflow
Tesla T4	none, from the PCIe slot (~70 W)	passive, low-profile, needs airflow
NVIDIA A2	no auxiliary power connector	40–60 W TDP, passive, low-profile, needs airflow
Tesla P40	yes, CPU/EPS 8-pin (~250 W)	check the specific pinout and adapter
Tesla P100 PCIe	yes, CPU 8-pin (~250 W)	check the specific pinout and adapter
Tesla V100 PCIe	yes, CPU 8-pin (~250 W)	check the specific pinout and adapter

Do not apply one power scheme to every Tesla. The P4/T4/A2 run from the slot alone; the P40/P100/V100 need an external plug with the correct pinout. I have gathered more about putting the kit together, and about what to look at when collecting a used card, in the buying checklist for a used data-centre card.

The Tesla works but overheats: cooling passive cards

This is an important stage that is easy to forget once the card has finally "come up". Most of these cards (P4, T4, A2, P40, P100, V100) have passive cooling: a heatsink with no fan, designed to work in a server chassis with forced airflow. The mere fact that the card appears in nvidia-smi does not mean that a desktop provides the right thermal conditions for it.

In an ordinary desktop case without adequate airflow, the card can quickly reach a high temperature and start reducing its clocks (thermal throttling). In extreme conditions the GPU's operation may be cut back as an emergency measure. That is why:

• Provide direct airflow through the heatsink: a duct/tunnel with an axial fan, or a server blower fan blowing along the fins.
• Under sustained load, monitor the GPU temperature (nvidia-smi -l 1), the power draw and any throttling.

Several cards: when a 2–4× P40 rig will not POST

With a single card the issue is usually "the system does not see it"; with several the problem changes character: the computer does not start at all (black screen, no POST, sometimes a restart loop). The core is the same: each card wants to map a large BAR window, and the more cards there are, the larger the total address space to be shared out, and the easier it is for the firmware to stumble over it. But it does not come down to a single switch. With a multi-card setup, check in order:

• Above 4G Decoding → Enabled: with several cards this is a basic condition.
• Allocation of PCI/MMIO/BAR space: whether the firmware has enough address space available to allocate the windows (see the dmesg diagnosis below).
• The board's capabilities and firmware: the number and configuration of PCIe lanes, the slot layout, the version of the board's BIOS.
• Resizable BAR: not a prerequisite for running the P4, P40, P100, V100 or T4; leave the default setting and focus on correct MMIO allocation and Above 4G for cards with large BARs.
• Other PCIe devices taking up address resources (additional controllers, network cards): sometimes they are the ones that "eat up" the space.
• Power and risers. Each P40/V100 draws ~250 W plus momentary spikes: in practice a powerful, stable power supply and solid EPS cables. Cheap risers and splitters can be a source of instability with several cards.
• Cooling: several passive cards side by side are a serious thermal challenge.

And the most important tactic: isolate the problem. Will it not start with four cards? Go back to one, confirm POST and nvidia-smi, then add them one at a time. That is the quickest way to pinpoint whether it is the BIOS, the MMIO space, the power or a particular slot/riser.

Step-by-step diagnosis: from PCIe to hardware failure

Instead of "switch on three options and hope for the best", go through this order. It usually settles the matter within a quarter of an hour or so and, more importantly, tells you where the problem really lies.

Step 1: Is the lack of a picture normal? Yes, a Tesla has no video output. Connect the monitor to the integrated graphics or a separate card. If that was all it was, you are done.

Step 2: Is the device visible on PCIe?

lspci -nn | grep -i nvidia
lspci -vv -s 01:00.0    # address from the output above; 01:00.0 is only an example

Step 3: Is there a BAR/MMIO problem or a driver problem?

sudo dmesg | grep -i -E 'nvrm|nvidia|no space|failed to assign'
nvidia-smi

If dmesg shows BAR 1 [mem ...]: can't assign; no space (older kernels: BAR 1: no space for ...), failed to assign, or the driver's NVRM: This PCI I/O region assigned to your NVIDIA device is invalid with BAR1 is 0M, the problem concerns the allocation of PCI/MMIO address space; a common cause is the firmware/BIOS configuration or the limitations of a particular platform. Reinstalling the driver will not fix it; go back to the BIOS (step 4). If the BIOS has no Above 4G option, the Linux kernel parameter pci=realloc sometimes lets the kernel reassign the windows itself; it is a workaround, not a guarantee.

Step 4: Does the BIOS allocate the space correctly? Check: Above 4G Decoding, UEFI mode, CSM, the PCIe/slot configuration and the board's firmware version.

Step 5: Does the driver work? If lspci sees the card and nvidia-smi reports "No devices were found" or fails, the system detects the device on the bus, so check the software first: the driver branch (R580 for Pascal and Volta), Secure Boot (if the kernel requires module signatures, sign the NVIDIA module with a key trusted by the kernel and enroll its certificate if necessary, for example through MOK, or disable Secure Boot), whether nouveau is still loaded (lsmod | grep nouveau), and the kernel logs (dmesg). Do not automatically assume, however, that the problem is purely a software one: confirm it with the logs.

Step 6: Does the card have the correct power supply? Check the specific model and the pinout of the cable/adapter (table above). A wrong pinout means a risk of a short circuit.

Step 7: Is the card stable under load? Monitor the temperature, power draw, ECC errors and NVIDIA logs; check the riser and the PSU. A passive card without airflow may only start disappearing under load.

Step 8: Only now consider a hardware failure. If you have gone through steps 1–7 and nothing fits, get in touch with us before you write the card off as dead.

Diagnostic table: symptom → what to check → what to do

Symptom	Check first	What to do
No picture after installing the Tesla	A Tesla has no video output (normal)	Connect the monitor to the integrated graphics or a card with a display output
Card not visible / code 12 (Windows)	PCI/MMIO allocation	Enable Above 4G; check UEFI/CSM and the BIOS's capabilities
BAR ... no space in dmesg	PCIe address space	Check Above 4G, the firmware and the PCI/MMIO configuration
lspci sees the GPU, nvidia-smi does not	Driver / kernel	Check the driver for the model and dmesg
The GPU disappears under load	Power / temperature / riser	Check the PSU, cables, temperature and riser
A rig with 2–4 cards will not POST	PCI/MMIO / firmware / power	Start with one card and add them one at a time
The GPU works but runs hot	Airflow	Provide forced airflow through the heatsink
The GPU died after a cable was connected	Power pinout	Use a cable/adapter compatible with the specific model

---

A disclaimer to finish, as always: yes, we sell these cards, so read the above with that in mind. But I am writing it exactly as I myself would have wanted to hear it before I first put a passive Tesla into a desktop. We test every card under load and include a burn-in certificate (temperature, power draw, ECC errors), because with used cards that is the difference between "it worked when I removed it" and real proof that the card is healthy. Before you decide a card is faulty, go through the diagnosis above, and if something does not add up or you are planning a multi-card rig, write to us: we will help you choose adapters, power, cooling and BIOS settings.

(See also: What to watch out for when buying a used data-centre card)

Sources

Technical basis: the manufacturer's documentation. For a specific power configuration and BAR/MMIO settings, always also check the documentation for your motherboard/server.

• NVIDIA: Tesla P40 Product Brief (PB-08338-001): specification, power connector (250 W, 8-pin) and power adapter (Table 5, Figure 6): https://images.nvidia.com/content/pdf/tesla/Tesla-P40-Product-Brief.pdf
• NVIDIA: Tesla P100 PCIe Product Brief (PB-08248-001): "CPU 8-pin power connector" and the supported power options (one CPU 8-pin cable, or a dongle fed by two PCIe cables): https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/tesla-product-literature/NV-tesla-p100-pcie-PB-08248-001-v01.pdf
• NVIDIA: Tesla V100 PCIe Product Brief (PB-08744-001): form factor, power, power draw: https://images.nvidia.com/content/tesla/pdf/Tesla-V100-PCIe-Product-Brief.pdf
• NVIDIA: T4 Product Brief (70 W Low Profile): passive, powered from the PCIe slot, requires airflow: https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/tesla-product-literature/T4%20Product%20Brief.pdf
• NVIDIA: A2 Tensor Core GPU Datasheet: low-profile, passive, configurable 40–60 W TDP, no auxiliary power connector: https://www.nvidia.com/content/dam/en-zz/solutions/data-center/a2/pdf/a2-datasheet.pdf
• Microsoft: Device Manager error messages (code 12: the device cannot find enough free resources; for these cards, one possible cause is a failure to allocate address space above 4 GB): https://learn.microsoft.com/en-us/windows-hardware/drivers/install/device-manager-error-messages