How to Test a Used GPU: The Checklist Before Your Return Window Shuts

By John published: August 17, 2026

A used graphics card has arrived. It looks fine. It probably is fine.

But you have a limited window in which “probably” can be converted into “definitely” at no cost to you, and once that window shuts you own whatever you were sent. eBay’s money-back guarantee runs 30 days from delivery. PayPal’s buyer protection runs 180. A private sale might give you nothing at all beyond the seller’s goodwill.

So the job is to find any fault fast, while it is still someone else’s problem. Here is the order to do it in, starting before the card goes anywhere near your machine.

The whole sequence takes about an hour, most of it unattended.

Before you open the box: film it

Start a video on your phone and film the unboxing in one unbroken take, including the packaging and the label.

This feels ridiculous until the day a card arrives with a snapped bracket or turns out to be a different model to the one advertised, and it becomes the single most useful sixty seconds of footage you own. Every disputed claims process comes down to whether damage happened in transit or in your hands, and unbroken footage settles it.

Step 1: the physical inspection, five minutes

Before power:

• PCIe fingers. Light scratching is normal. Deep gouges, corrosion, or a green tint mean it has lived somewhere damp.
• The power connectors. Look inside them. Any browning, melting or discolouration on a 12VHPWR or 8-pin connector is a hard stop - that card is a fire risk, not a bargain.
• The PCB. Look for scorch marks, bulging capacitors, residue from a leak, or the dull grey film that says it lived in a smoking household.
• Screws and stickers. Chewed screw heads mean it has been apart. That is not automatically bad, plenty of people repaste their own cards, but it is worth asking about. A broken warranty sticker tells you the warranty is gone.
• The fans. Spin each one by hand. They should turn freely and silently, with no grinding and no wobble on the axis. Fans are the most common failure on a used card and among the easiest things to replace.
• Sag and the bracket. Sight down the PCB. A permanent bend suggests it hung unsupported in a case for years, which stresses solder joints.

Step 2: is it actually the card you paid for?

Fake and rebadged cards exist. A flashed VBIOS can make a lesser GPU report itself as a better one, and the giveaway is in the details rather than the name.

Install GPU-Z on Windows, or use nvidia-smi -q on Linux, and check:

• The GPU die name, not just the marketing name. A card claiming to be a 3090 should report GA102.
• Memory size and bus width. 24 GB on a 384-bit bus for a 3090. A mismatch here is the classic fake.
• The VBIOS version, which you can cross-check against TechPowerUp’s VBIOS database for that model.
• PCIe link width and speed while under load. A card running at x8 when it should be x16, or dropping to Gen3 in a Gen4 slot, points at a dirty slot, a damaged connector, or a motherboard configuration issue.

On Linux, the equivalent one-liner:

nvidia-smi --query-gpu=name,memory.total,vbios_version,pcie.link.width.current,pcie.link.gen.current --format=csv

Step 3: test the memory properly

This is the step almost everyone skips, and it is the one that matters most if you bought the card for AI work.

Faulty VRAM does not announce itself. A card with a handful of bad memory cells will game happily for hours, because a corrupt texel for one frame is invisible. Load a language model into that same memory and you get garbage output, crashes on long generations, or an inference engine that dies twenty minutes in. Mining cards in particular ran their memory hot for years, and GDDR6X degrades.

Use memtest_vulkan. It is free, cross-platform, works on NVIDIA, AMD and Intel, and tests the whole of the card’s memory rather than the slice a game happens to touch.

Run it for at least an hour, ideally longer, and you want zero errors. Not “a couple”. Memory errors are not something you live with.

Step 4: sustained load and thermals

Short benchmarks prove nothing. Faults show up when the card is hot and has been hot for a while, which is exactly the state a benchmark run of two minutes never reaches.

Pick a sustained load - Unigine Superposition on loop, a 3DMark stress test, or OCCT’s GPU test - and run it for 30 minutes minimum while logging with HWiNFO64, or nvtop on Linux.

What you are watching for:

The 3090 deserves a specific warning here. Its memory modules sit on both sides of the board, the rear ones are poorly cooled from the factory, and years of load bakes the thermal pads hard. A memory junction temperature above 100°C is common on neglected examples. It is fixable with a pad replacement, and it is a legitimate reason to ask for money back.

Step 5: the AI-specific test

Gaming stability and inference stability are not the same test, and if you bought this card to run models, the gaming test alone will not catch what you care about.

Load a model that nearly fills the card. On a 24 GB card, that means something in the 32B class at Q4_K_M rather than a comfortable 8B.

ollama run qwen3:32b "Write a 2000 word essay on the history of the steam engine."

Then watch three things:

ollama ps

Processor must read 100% GPU. Anything else means it did not fit and the test is invalid.

nvtop

Memory should sit near capacity and stay there. A card that crashes only when its last gigabyte is exercised is a card with a bad memory module at the top of the address range, and a smaller model would never have found it.

And read the output. Coherent prose means healthy memory. Repeated tokens, sudden gibberish, or a crash partway through a long generation are the symptoms of faults that gaming benchmarks routinely miss.

Do that on a long generation, twice. It takes ten minutes and it is the most relevant test in this article.

Step 6: compare against expected performance

Run a standard benchmark and compare your score with published results for the same card. 3DMark keeps a database, and Superposition scores are widely posted.

Within about 5% of typical is fine. 10-15% below means something is wrong: throttling, a PCIe link running at half width, a power limit set low in a flashed VBIOS, or paste that gave up years ago. It is worth chasing down rather than shrugging at, because it is usually fixable and it is always leverage in a conversation with the seller.

What is worth fixing, and what is a return

Fix it yourself:

• Dried thermal paste and hard pads. An afternoon, cheap materials, and typically 10-15°C off both core and memory temperatures.
• Worn fans. Replacements are readily available for common models and swapping them is straightforward.
• Dust. Obviously.

Send it back:

• Any memory errors at all.
• Visible artifacts under load.
• Scorched or melted power connectors.
• A card that is not the model advertised.
• Crashes that survive a clean driver reinstall.

The middle ground is negotiation. A card that runs 8°C hot with tired pads is still a decent buy at a discount that covers the repaste kit and your Saturday afternoon. Message the seller with your logged numbers before opening a case - most private sellers had no idea, and a partial refund suits everyone better than a return.

If it fails

Open the dispute inside the window. Do not let a seller talk you into “trying a different driver” for three weeks while your protection expires - that is not always a stalling tactic, but the clock does not care either way.

Have your evidence ready: the unboxing video, the memtest output, the temperature logs, the benchmark comparison. A claim with logs attached is settled quickly. A claim that says “it seems slow” is not.

Frequently asked questions

How long should I test a used GPU for?

About an hour of active testing, and longer if you can leave it unattended. An hour of memtest_vulkan with zero errors, 30 minutes of sustained graphical load with sensible temperatures, and one long AI generation covers everything that commonly goes wrong.

Are ex-mining GPUs bad to buy?

Not automatically. A card run at a steady undervolted load in a ventilated room has often had an easier life than a gaming card thermally cycled hard every evening. The real risks are worn fans and degraded memory thermal pads, both of which are testable in advance and repairable afterwards.

How do I know if a used graphics card has bad VRAM?

Run memtest_vulkan for at least an hour and expect zero errors. Faulty memory rarely shows up in games, because a corrupted pixel for a single frame is invisible - but it produces crashes and garbled output when a language model occupies the whole of the card’s memory.

What temperature is too hot for a used GPU?

Core above roughly 80°C under sustained load is worth investigating, and a hotspot more than about 15°C above core suggests uneven contact. On GDDR6X cards, memory junction temperature above 95°C means the thermal pads need replacing, which is common on used 3090s.

Can I test a used GPU on Linux?

Yes, and for AI buyers it is the more relevant environment. nvidia-smi reports the model, memory, VBIOS and PCIe link state, memtest_vulkan runs natively, nvtop gives live temperatures and memory use, and loading a large model in Ollama is the most realistic stress test available.

What if the seller says it worked fine for them?

They may be telling the truth. Gaming loads exercise a card very differently from inference, and a fault that only appears when the last gigabyte of memory is in use will genuinely never have shown itself. Send your logs, stay civil, and open the dispute inside the window regardless.

Related guides

• Best Graphics Card for Gaming: A Resolution-First Buyer’s Guide - what to buy if games are the point
• Best GPU for Running Local LLMs: A VRAM-First Buyer’s Guide - what to buy if AI is the point
• Local LLM Jargon Explained: Tokens, Weights, VRAM and Quantisation - the vocabulary behind the tests above
• Build a Headless Ubuntu LLM Server with Ollama on 8 GB VRAM - what to do with the card once it passes

This article may contain affiliate links to Amazon, eBay and other retailers. I may earn a small commission from qualifying purchases made through these links at no additional cost to you.