Buying a used data-centre GPU (Tesla P40, V100, A2): what to check

Buying a second-hand compute card is different from buying a used gaming GPU. The biggest risk is not that the card is "old": a Tesla P40 from 2016 can still do solid work for the next few years. The real risk is ending up with a unit that is faulty or worn out: a card that has spent years in a data centre under constant 24/7 load, sometimes at the edge of its thermal limits. How to check this before you spend your money is covered below.

Why "it worked when I removed it" is not enough

Server accelerators (Tesla, now referred to as NVIDIA Data Center GPUs) differ from consumer cards in their workload profile. In typical home use a GPU spends most of its time idle or under moderate load. In a data centre a card can run close to 100% utilisation for years: model training, inference, rendering. That does not mean such a card is worse. It only means that its condition cannot be judged "by eye" or simply from the system booting.

A card can initialise correctly and appear in nvidia-smi, yet still have damaged memory cells that only show up under full load, or a deteriorating cooling system that makes the card throttle, or even produce errors, after a few minutes of work. The mere fact that the card "comes up" in the system says nothing about whether it will survive a real workload.

A load test is your test drive

The only sensible way to verify that a used compute card is in working order is to put it through a test under full load; the most popular tool for this is gpu-burn. Load the card for several minutes and check that gpu-burn finishes with OK and no computational errors. This is a screening test, not a guarantee of complete health. Before and after the test, read the available ECC and retired-page counters with nvidia-smi -q -d ECC,PAGE_RETIREMENT; for the A2 (Ampere), also inspect nvidia-smi -q -d ROW_REMAPPER. Check that ECC is enabled, and do not interpret N/A as zero errors. Investigate new uncorrectable errors, retired pages or remapped rows with the seller. Increasing corrected-error counts also need assessment, but nonzero historical counters alone do not establish that a card is faulty. It is exactly the same mechanism as a test drive when buying a car: you are not interested in whether the engine starts when the car is parked, but in whether the car behaves properly in real conditions.

When buying a card, ask about this wherever and from whomever you buy, whether it is a seller on Allegro (Poland's largest marketplace), a listing on eBay or someone with a local advertisement. Simply asking "has the card passed a gpu-burn test and what was the result?" filters out sellers who do not check what they ship at all. If the seller has no such test, you can run it yourself once you receive the card: under Linux it takes a few commands, and the result gives you solid grounds for a claim if something is wrong.

It is worth distinguishing between two levels of confidence. "The card worked when I removed it from the server" is an unsupported claim: there is no knowing under what conditions or for how long that "test" ran. A test certificate tied to the serial number of a specific unit is in a different league entirely: it means that exactly the unit you receive has been verified, not some other card from the same batch.

Inspection checklist: what to check physically

Apart from the load test itself, it is worth examining the card for a few things that are easy to overlook:

• Condition of the heatsink and thermal paste: traces of dried-out, cracked paste or visible discolouration of the heatsink may suggest that the card overheated in the past.
• Completeness of the package: for the 250 W cards (P40, P100, V100 PCIe) the most important item is the power adapter. These cards take a CPU-style 8-pin (EPS) connector, not the PCIe 8-pin familiar from gaming graphics cards, and the 12 V and ground pins sit on opposite sides in the two connectors, so a PCIe plug must never be forced in. NVIDIA's adapter takes two PCIe 8-pin cables and outputs one CPU 8-pin. Low-power cards such as the P4 and A2 have no power connector at all: they draw everything from the PCIe slot.
• Signs of tampering: scratched screws, signs of unauthorised removal of the heatsink, accidentally bent pins in the connector. All of this is worth spotting in the photographs before buying and asking the seller about.

Compatibility with the hardware you already have

A card can be in working order and still not work properly in your system unless you take care of a few things in advance:

• Above 4G Decoding in the BIOS: cards exposing large BAR apertures need PCIe resources to be mapped above 4 GB, usually enabled through a BIOS option called Above 4G Decoding. The Tesla P40, for example, exposes a 32 GB BAR1 aperture in compute mode. Without mapping above 4 GB, the system cannot allocate it and the card may fail to initialise; insufficient resources can produce Code 12 in Windows. The requirement follows from the BAR size, rather than VRAM capacity alone. Resizable BAR is a separate option and is not required to run these cards. It is worth checking the motherboard documentation before you buy the card, not after the fact.
• No video output: Data-centre cards of this kind have no display output connectors. You need either the CPU's integrated graphics or a separate graphics card just for the display, if you plan to work locally on the same machine.
• Operating system: NVIDIA provides drivers for these cards on supported Linux and Windows versions. Before buying, check compatibility between the specific GPU, driver version, operating system and tools you intend to use.
• Cooling passive cards: many used server accelerators are passive designs, intended to work in a server chassis with a very strong airflow forced by the chassis fans. Put into an ordinary desktop case without a plan, such a card will overheat. The cooling plan (additional fans, a 3D-printed shroud, mounting in the right place in the case) must be in place BEFORE the purchase, not after the card starts to throttle. I have written about this separately in the article on cooling passive cards.

Architecture and service life: are you buying for now or for years?

It is worth understanding where in its life cycle the architecture of the card you are considering stands, because this translates into driver support and the length of sensible use:

• Pascal (e.g. P4, P40, P100) and Volta (V100): CUDA 13.0 removed support for both, so CUDA 12.9 is the last toolkit that compiles for them, the R580 driver branch is their last (supported until mid-2028) and PyTorch 2.14 is the last release with a ready-made build for them. The cards keep working on that software stack. That does not make them a bad buy: for an inexpensive way into the field or a project of a few years, Pascal in particular can be a sensible choice on price.
• Ampere (A2): has by far the longest expected period of support in this group, because the CUDA 13 cut-off affects only architectures below compute capability 7.5, and the A2 is 8.6. The V100 shares Pascal's timetable, so for the next few years the A2 is a safer choice than the V100, even though both look like "newer generation" architectures than Pascal on paper. If you are planning to build something that has to work for the next few years, it is worth taking this difference into account when choosing the model, not only once driver support starts to run out.

If you are considering a specific model, it is worth reading separately about the P40, V100, P100, A2 and P4: the differences in VRAM, power consumption and performance are considerable, and not every model suits every application. Keep in mind, however, that the models currently listed for sale are not an exhaustive list: if you need a different model, a newer architecture or a card with more VRAM, accelerators of this type can be sourced to order, with a quotation on request.

Warranty on used hardware: what to expect

The warranty on used server hardware is usually short, counted in months rather than years. That is normal for this category of equipment, and a short warranty in itself should not automatically put you off. If you buy as a consumer from a business in the EU, the seller's statutory liability for non-conformity applies on top of any warranty; for second-hand goods it can be shortened, but to no less than one year. It does not apply to business purchases or to buying from a private individual. What matters more is that, before buying, you know exactly what the warranty covers (only hardware faults, or also, for example, non-conformity with the specification) and how to make a warranty claim. A certificate of a test performed before dispatch can serve as a point of reference here: if the card had a documented gpu-burn PASS on the day of dispatch and starts producing ECC errors two weeks later, you have clear grounds for a conversation with the seller, instead of it being your word against theirs.

A myth that needlessly puts people off: vGPU licences

Some people give up on buying data-centre cards of this kind because they have heard that an NVIDIA vGPU licence must be purchased. That applies only to NVIDIA vGPU software, which shares a GPU between virtual machines through vGPU profiles, for graphics or for compute. Passing the whole card through to a single virtual machine for compute with the ordinary data-centre driver (no vGPU software) does not need it; passing the card through for professional graphics with the vGPU driver requires a vWS licence. For ordinary use with CUDA, that is, training or inference of AI models directly on the host or in a container, no additional licence is needed. It is worth making this distinction so as not to reject sensible hardware because of a misunderstanding.

Checklist: before you click "buy"

• The seller has shown a load-test result (gpu-burn: OK with no computational errors, or an equivalent result from another tool), plus before/after ECC and retired-page readings and, for the A2, row-remapping data
• The test is tied to the serial number of the specific unit, not a general "tested" claim
• Photographs of the heatsink/paste show no signs of overheating or unauthorised tampering
• For a 250 W card (P40, P100, V100): the package includes a 2× PCIe 8-pin to CPU 8-pin (EPS) adapter, or you know where you will get one (the P4 and A2 need none)
• You have checked in the motherboard documentation that it supports Above 4G Decoding (Resizable BAR is not required)
• You have a planned source of display output (integrated graphics or a second card), because data-centre cards have no video output
• If the card is passive, you have a cooling plan ready BEFORE the purchase, not after
• You know the card's architecture and its realistic period of support (Pascal and Volta: last driver branch R580, supported until mid-2028; Ampere: current)
• You know what the warranty covers and how to make a warranty claim
• You understand that you do not need a vGPU licence for CUDA