How to Inspect a Used Graphics Card Before Buying: Checklist, Mining-Card Signs, Warranty & Return Questions

2026-10-18 ｜ GPU Repair ｜ 15 min ｜ HK GPU Fix Editorial

Key Takeaways

• A used graphics card should be judged by the exact card in front of you, not by its box, its listing photos, or a single benchmark score. Boxes can be reused, serial-number labels can be altered, and listing photos can be borrowed from another listing.
• The core verification method is cross-checking: compare what GPU-Z reports against the manufacturer's official specification page for that exact model, and back it up with a VRAM stress test, a visual inspection, and tests of the ports you actually plan to use.
• GPU-Z can flag some counterfeit cards with a [FAKE] label and a yellow warning, but that flag only covers specific older donor chips. It does not catch every fake, so a clean GPU-Z screen is not proof of authenticity on its own.
• There is no reliable usage-hours counter on consumer GPUs, no universal BIOS checksum repository for buyers, and no way to read prior workload (mining, gaming, rendering) from the card itself. Anyone who claims otherwise is guessing.
• Mining-card signs are clues, never proof. Dust, oil, fan wear, and even a modified vBIOS all have non-mining explanations. What matters is the card's present condition, measured by tests, not the label attached to its past.
• Agree return terms in writing before paying, and get the seller's answers to the warranty questions in writing too. In Hong Kong, a private second-hand sale has no automatic cooling-off or return right established by the sources used here.

Why a Box and a Benchmark Are Not Enough

Buying a used graphics card is different from buying most other used electronics. The card's value sits in silicon, memory, and a power delivery system that you cannot judge from the outside, and the most common traps are the ones that look convincing in a listing: a clean box, a neat photo, and a strong benchmark number.

A proper verification should show the exact card from multiple angles, the serial number, display output working, and the card running under load. A seller who refuses to show the serial number, says the card is already packed, or keeps finding reasons not to test it is a reason to walk away. A benchmark score alone is not proof that a card is healthy either. Temperatures, fan behaviour, clock stability, and visual output matter as much as the number at the end of the run. Second-hand cards may also carry hidden degradation, especially in VRAM, and many failures only appear under targeted memory stress. Visual stability alone is not enough.

This article walks through a complete inspection in the order a buyer should do it: before meeting the seller, physical inspection, authenticity checks, vBIOS checks, functional tests, mining-card clues, and finally the warranty and return questions that protect you if something turns out to be wrong.

Before You Meet the Seller

Most of the protection in a used-card purchase happens before any money changes hands. Ask the seller for:

• The exact model and board maker, plus photos of the exact card from multiple angles.
• A photo of the serial-number sticker.
• The original receipt or invoice.
• The card's stated history: first owner or not, any mining use, any repairs, any repaste, any BIOS flash.
• A short video of the exact card powered on and running under load.

Then do your own homework. Look up the manufacturer's official specification page for that exact model and note the GPU die, VRAM type and size, clocks, power connectors, and recommended power supply. That spec page is the reference everything else gets checked against.

Most importantly, agree the return terms in writing before payment: what counts as a defect, how many days you have to test, who pays return transport, and the refund method. Get the seller's identity and contact details in writing, and keep the payment record. If the seller will not put terms in writing, treat that as information in itself.

Physical Inspection

Inspect the card in person, or on a live video call, before paying where possible. Work through the table below and note anything that does not look right. A single small flaw may be cosmetic; a cluster of them is a pattern.

Two details deserve emphasis. First, the 16-pin connector family (12VHPWR and 12V-2x6) deserves a close look on high-wattage cards: incomplete insertion is one of the most common themes in melting reports and does not always look obvious at a glance. Avoid tight bends immediately after the plug, prefer a native 16-pin power supply cable when possible, and periodically inspect seating for discoloration or a hot-plastic smell. Second, tamper evidence is a clue, not proof. A sticker can itself be faked, so a perfect-looking sticker proves nothing on its own, and a disturbed one is a reason to ask questions, not an automatic verdict.

Proving the Card Is Genuine

Counterfeit cards typically work by re-flashing firmware so an old or cheap GPU reports itself as a newer, more expensive model, including falsified memory capacity. The defence is cross-checking with GPU-Z, a free utility from TechPowerUp that reads live hardware data from the card, including name, GPU die, memory type, memory size, and BIOS version. Download it from TechPowerUp's official site, not a mirror, and use the latest version, because an old version may mislabel a card.

GPU-Z pulls information from the card's vBIOS and hardware. Version 2.12.0 added fake-NVIDIA detection: it compares the chip a card is supposed to have with the one it actually runs, and on a mismatch it adds a [FAKE] prefix in the Name field plus a large yellow warning triangle on the main info tab. The feature targets specific older NVIDIA chips used as donor silicon, including G84, G86, G92, G94, G96, GT215, GT216, GT218, GF108, GF106, GF114, GF116, GF119, and GK106. For example, GPU-Z should be able to detect a fake RTX 2060 that in reality uses a GK106 GPU.

The limitation matters as much as the feature: the fake flag covers the listed older donor chips, and newer tampering methods exist. Never claim, and never accept a seller's claim, that GPU-Z catches every fake or every modern counterfeit.

The cross-check method is:

1. In GPU-Z, confirm the Name field has no [FAKE] flag or yellow warning.
2. Compare the GPU die, memory type, and memory size against the manufacturer's official spec page or TechPowerUp's GPU database for that exact model.
3. Check shader and ROP counts and clocks where available.
4. Confirm the sub-vendor shown in GPU-Z matches the branding on the card.

As a worked example from a reputable outlet, a genuine RTX 4090 should show GPU die AD102, Memory Type GDDR6X, and Memory Size 24576 MB in GPU-Z. If any of those are off, stop testing and start the return. The same outlet notes that a real RTX 4090 should deliver substantially higher 4K frame rates than a 3090 Ti, so failure to hit expected performance is a warning sign. That comparison was that outlet's example for that pair of cards; do not turn it into a pass or fail threshold for other models.

Also check the serial number properly. Most graphics cards carry a unique serial number, normally on a physical sticker on the card. Windows Device Manager and hardware IDs generally show model and device ID, not the serial number. Some manufacturers' own software may display it, but GPU-Z usually does not. PowerColor's official FAQ states that the serial-number sticker on the back of the product must remain intact, and that removal or damage voids all warranties. If the sticker is missing, altered, or illegible, the product is not eligible for warranty service.

vBIOS: Read It, Never Flash It to Check

GPU-Z displays the BIOS version on its main screen and can save it. In newer versions, GPU-Z flags inconsistencies between the GPU's internal information and the data provided by the BIOS with a yellow warning icon, meaning it can detect that a BIOS does not actually correspond to the hardware even if someone flashed a vBIOS from another model.

The check is simple: note the vBIOS version in GPU-Z (on some cards, the Advanced tab shows NVIDIA BIOS details) and compare it with the version or versions the board manufacturer publishes for that exact model. A mismatch against the manufacturer's published BIOS for that exact board is a question to raise with the seller, not automatic proof of mining. Legitimate customisation and repair are also possible reasons for a modified BIOS.

Two hard rules apply. First, reading the GPU BIOS version is safe, but flashing a vBIOS is not: one wrong image or an interrupted flash can leave the card unbootable. If firmware is ever updated, only use the exact board-specific ROM from the GPU vendor or manufacturer support channel, and understand the recovery options first, such as dual BIOS, an external programmer, or vendor service. Second, do not re-flash a card to check it. Flashing is a risk to the card, not a diagnostic step.

Note also that vBIOS modification of many modern RTX cards is possible with advanced utilities. Do not accept the claim that modern cards cannot have a modified BIOS. And there is no known public checksum repository that lets a buyer definitively verify an arbitrary used card's dumped BIOS against the stock BIOS. Do not let a seller wave a checksum at you as proof.

Functional Tests

Before testing, revert all GPU and CPU settings to stock. If crashes persist, clean-install the GPU driver, and verify power supply capacity and cabling, especially on high-wattage cards. Use DDU to remove existing drivers before reinstalling, check separate PCIe power cables rather than split connectors, and disable CPU overclocks during GPU testing.

Run the tests in this order:

1. Detection. The card appears in Device Manager with no yellow warning icons, and GPU-Z shows the correct model and VRAM size.
2. Benchmark or game loop. Run a demanding benchmark or game long enough to check for crashes, black screens, artefacts, flickering, or sudden clock-speed drops. Artefacts means dots, lines, or strange colours on screen, and also flashing textures, missing shadow maps, and sudden frame drops. During benchmarking, watch for crashes, lower frame rates than the card should produce, stuttering, and bizarre visual artefacts, which can indicate overheating or fan problems, an unsustainable or dangerous overclock, or VRAM errors.
3. OCCT VRAM test. This is the single most important test for a used card, because it targets the memory specifically. One guide's method: set memory allocation to 80 to 90 percent of total card capacity (for example, target 12 to 14 GB on a 16 GB card, leaving room for the Windows display driver), and run for 30 minutes. Zero errors are tolerated. Even a single memory error should be treated as a red flag, and stable temperatures do not guarantee healthy silicon or memory. If errors appear immediately, abort the test. Always baseline-test at stock settings before any overclocking, test one change at a time, and never ignore a single error during stress testing.
4. FurMark or Heaven. One diagnostic guide lists FurMark running for 5 to 10 minutes without crash or artefacts as a checklist item, alongside Heaven Benchmark completing without issues and an up-to-date display driver. Watch temperatures, fan behaviour, and clock stability during the run.
5. HWiNFO64. Pay attention to GPU Memory Junction Temperature (or VRAM Temperature) and the GPU Memory Error Counter, where the card exposes them.
6. Ports. Test every HDMI and DisplayPort output you plan to use, with a real display. A dead port is a defect even if the card benchmarks fine.
7. Second-PC cross-check. If available, test the GPU in a second PC, or test a second GPU in the same PC. This separates card faults from power supply, slot, and driver faults.
8. Power supply. Confirm the wattage and cabling are adequate for the card, and confirm the 16-pin plug is fully seated and not tightly bent at the plug.

On temperatures, be careful with single numbers. One buyer's guide suggests aiming for 65 to 85 degrees Celsius under stress in FurMark or UNIGINE Superposition, while another diagnostic guide lists idle below 50 degrees and load below 85 degrees as checklist items. A third source warns that extreme power tests can be misleading on laptops and may trigger protection mechanisms unrelated to real-world use. There is no single universal pass or fail temperature: limits vary by model, cooler design, ambient temperature, and case airflow. The safe framing is to compare against the same model's typical behaviour and watch for throttling, crashes, and artefacts, not one magic number.

Other methods exist for context. A 3DMark stress test method in one guide uses Speed Way or Time Spy Extreme Stress Test, all 20 loops (about 10 to 15 minutes), with a pass rating of at least 97.0 percent frame-rate stability. A high-resolution gaming validation method in the same guide uses demanding, VRAM-heavy games at native 4K or high render scales with ray tracing for at least 45 to 60 minutes, scanning for flashing textures, missing shadow maps, or sudden frame drops caused by error-correction throttling. Both are that guide's methods, not manufacturer standards. MemTestG80 and MemTestCL are named in one guide as GPU memory-test options; treat them as options, not as the standard. Factory-style diagnostics are not buyer tools: NVIDIA's DCGM GPU Memory Plugin validates memory integrity in data-centre settings, but DCGM diagnostics do not repair faults, do not replace NVIDIA Field Diagnostics, and do not determine RMA eligibility.

For more on reading GPU-Z and HWiNFO64 output, see the companion monitoring guide (article ID 108). For a card that is not detected at all, see the no-detection diagnostic guide (article ID 106). For no signal over HDMI or DisplayPort, see the display-output guide (article ID 105). For vBIOS problems, see the VBIOS guide (article ID 104).

Mining-Card Signs: Clues, Never Proof

The question buyers ask most is whether a card was used for mining. The honest answer is that prior workload cannot be read from the card. It can only be inferred from clues, and every clue has a non-mining explanation.

Start with the physics. What fatigues a circuit board over time is not hours on the clock alone but thermal cycling: repeated heating and cooling makes materials expand and contract at slightly different rates, straining solder joints and connections. A gaming card cycles temperature frequently; a mining card typically sits at one stable temperature. Miners often undervolt because electricity is their main operating cost: maximum voltage gives diminishing hashrate returns while sharply increasing power bills and heat. So many ex-mining cards may have run at lower voltages and cooler core temperatures than gaming cards. Mining does not automatically damage a GPU simply because the workload is mining. It can contribute to permanent damage if a card is run under excessive heat, voltage, or poor cooling for long periods. The counterweight is that mining places sustained load on VRAM and raises memory temperatures, which can contribute to long-term wear. Good cooling and memory-temperature monitoring reduce the risk. There is no reliable fixed lifespan for a mining GPU; service life varies with hardware design, temperature history, voltage, fan quality, maintenance, workload, and environment.

Whether mining or gaming wears a card faster depends on the individual card's temperatures, tuning, operating hours, and cooling, not on the label mining or gaming alone. What actually ages first is usually not the GPU die itself: at the voltages these cards run, core degradation over a few years is described as negligible. What ages is thermal paste and pads drying out, and fans and bearings wearing. One forum thread argued the opposite on component degradation, saying capacitors and VRAM are the first to fail from heat and that running cards at high temperature around the clock for years is not safe from component degradation. That disagreement is worth knowing about, but it is forum discussion, not a settled number, and this article does not pick a winner.

Supported mining clues, each a clue and not proof on its own, are: a modified vBIOS that does not match the board maker's published BIOS for that exact model; heavy dust or oil staining; a non-stock fan, replaced cooler, missing screws, or mismatched tamper stickers; VRAM errors under the OCCT VRAM test; and seller-context clues, such as a seller listing many identical GPUs plus mining accessories like PCIe risers or power adapters. The last one is circumstantial, not a technical finding. Present it as a reason to ask the seller directly, never as proof.

The direct question is still worth asking: has this card been used for mining, and for how long, at what settings? Record the answer in writing. A refusal to answer is itself information. But whatever the answer, judge the card by its test results.

Warranty and Returns in Hong Kong

A seller advertising a used GPU with warranty remaining does not guarantee the next owner can use it. Always ask for the original invoice and check the manufacturer's transfer rules. A screenshot showing an expiration date is not enough by itself.

Brand policies differ, so the only safe rule is: varies by brand, check the exact model's policy, and never generalise one brand to all. The examples below illustrate the range; they are not a substitute for the current terms of the exact card you are buying.

• EVGA (transfer-friendly example). Transferable warranty for products in original factory condition retaining all factory labels and stickers. Transferred warranty will not exceed 3 years from the product's shipping date from EVGA and will not exceed the original warranty. Second-hand purchasers need valid proof of purchase of the transaction between current and previous owner. Some suffixes are excluded from transferable cover. Note that EVGA has exited the GPU market, so its policy is an example of transferability, not a buying recommendation.
• Sapphire (restrictive example). Warranty is described as non-transferable and covering only the original purchaser through an authorised dealer. Sapphire's RMA text states VGA products carry a 2-year warranty, with enquiries through the initial place of purchase, only by the original purchaser, requiring original invoice and documentation. Warranty is invalid if, among other cases, the serial number on the product and receipt do not match, or the product was not sold by official distributors or resellers. Dismantling the card or removing the heatsink or cooler fan renders the warranty invalid. Regional terms and time periods vary, so do not quote a single Sapphire warranty length as global.
• PowerColor. The warranty period begins from the purchase date on the invoice or receipt, proof of purchase is required, and the serial-number sticker must remain intact. PowerColor also states it does not provide a global warranty: the warranty period depends on the region of purchase, so check with the reseller. One secondary summary describes US and Canadian coverage as applying only to the original owner.
• ASUS (disputed). Two secondary sources disagree about ASUS transferability. One claims ASUS warranties are transferable if the new owner receives the original proof of purchase and registers the transfer (usually within 30 days), and can only be transferred once. Another argues ASUS warranties likely do not transfer to a second-hand buyer. Do not treat either as settled fact for Hong Kong. Check ASUS Hong Kong's current terms for the exact model before relying on any transferred warranty.
• Mining and commercial-use exclusions. Standard consumer warranties are normally written around ordinary desktop use. Cryptocurrency mining, blockchain computing, server deployments, and other commercial workloads may be excluded. Gigabyte is cited as stating that using its standard desktop graphics cards for blockchain computing or cryptocurrency mining renders the warranty void; Zotac and PNY are cited as including restrictions covering mining or enterprise-style use in some regional policies. Phrase this as some brands and regions exclude mining, and check the exact policy, not as a universal rule.
• Opening and repasting. Opening a card still under warranty can create problems. Sapphire states that dismantling the card or removing its heatsink invalidates the limited warranty, and damage caused during disassembly is unlikely to be covered by any maker. Check the official regional policy before removing the cooler, fitting a water block, replacing thermal pads, flashing a custom BIOS, or making any physical modification.

Ask all of these questions in writing before paying:

1. Are you the original purchaser? Can you show the original receipt or invoice with store name, date, model, and price?
2. Does the serial number on the card match the serial number on the receipt and box?
3. Has the card ever been opened, repasted, had pads or fans replaced, had its cooler removed, or had its BIOS flashed?
4. Has the card been used for mining, blockchain, or commercial workloads?
5. Which brand and exact model is it, and what does that brand's current Hong Kong warranty policy say about second-owner transfer?
6. If the brand requires transfer registration, who will provide the documents, and by when?
7. What is the agreed return window and defect definition? For example: if GPU-Z shows a [FAKE] flag, VRAM errors above zero, artefacts, crashes under the agreed test, or any port dead within an agreed number of days, full refund.
8. Who pays return shipping or transport if the card fails the agreed test? What is the refund method and timing?
9. How will payment be made, and will the buyer keep a receipt or transaction record showing the seller, date, model, serial number, and agreed terms?
10. If buying from a shop rather than a private seller: is the return policy in the shop's written terms? Does the shop offer its own in-shop warranty or exchange period, separate from the manufacturer's warranty?

Two Hong Kong realities frame all of this. First, the Consumer Council found that most surveyed online second-hand platforms said they only served as a trading-venue provider and would neither intervene in the transaction process nor bear liability for the quality and authenticity of items traded. None of the surveyed platforms implemented counterfeit-check mechanisms or return safeguards, and only two platforms had a dispute-resolution mechanism. Buyers must be vigilant, especially for high-value transactions: ask the seller for relevant proof or seek expert advice.

Second, under Hong Kong's Sale of Goods Ordinance, as explained by CLIC, a buyer is entitled to reject goods and get money back if the goods are faulty, not of satisfactory quality, do not match their description, or are different from the ones ordered. A valid rejection needs a clear indication that the goods were not accepted and remain at the seller's risk. Under section 38, where goods are delivered and the buyer refuses them with a lawful reason, the buyer is not bound to return them to the seller, though they may do so; it is sufficient to intimate refusal to the seller. But if problems are found after accepting the goods, the buyer may not be entitled to return them and demand a full refund, though they can sue the seller for compensation if the seller refuses to replace them. In practice, the return policy for a private second-hand sale is whatever the two parties agree in writing before payment. There is no automatic cooling-off or return right on a private used-GPU sale in Hong Kong that the sources used here establish. A Consumer Council cooling-off proposal cited in search results is a proposal for specific trades, not an enacted general right. Keep receipts, manuals, and maintenance and warranty documents, and note that some shops may willingly give refunds or exchanges for change of mind even though there is no legal obligation to do so; traders are free to decide.

If a card you already own needs professional inspection, the published reference range is HK$400-2,800 total; actual fee set by the post-inspection quote.

Red Flags: Walk Away

Any one of these is enough to walk away, or to require a full refund under the pre-agreed terms:

1. GPU-Z shows [FAKE] or a yellow warning, or the die, memory type, or memory size do not match official specs.
2. The OCCT VRAM test shows any error above zero.
3. Artefacts, crashes, black screens, or flickering under load at stock settings with a clean driver and adequate power supply.
4. Burn marks, a melted or discoloured power connector, a hot-plastic smell, or a damaged PCIe connector.
5. A serial-number sticker that is missing, altered, or illegible, with no matching receipt.
6. A seller who refuses to show the serial number, refuses a powered-on or under-load test, or says the card is already packed.
7. A seller who cannot or will not answer the warranty questions above, or whose story changes between messages.
8. A BIOS version that does not match the board maker's published BIOS for that exact model, with no explanation from the seller.
9. Any port you need is dead.
10. A deal that requires immediate payment with no written terms, no receipt, and no way to contact the seller afterwards.

What No Check Can Tell You

Be honest about the limits. No usage-hours counter exists: a consumer GPU does not itself track power-on or usage hours in any way, and no known reliable counter or app can show GPU hours. Hours would not answer the real question anyway: power-on time does not tell whether the GPU ran at low load showing a browser tab or at full load. Prior workload cannot be read from the card. Remaining lifespan cannot be measured. A card that passes every test today can still fail later; a test is a snapshot, not a guarantee. And whether a warranty will be honoured for a second owner cannot be known from a screenshot or a seller's promise. Only the brand's current regional policy plus proof of purchase plus the card's physical state (stickers, no disassembly, no mining exclusion) together determine it, and the final decision rests with the brand or distributor.

That is also why the checklist matters more than any single test. No one check proves a used card is good. The combination of a genuine-card cross-check, a zero-error VRAM test, a clean physical inspection, working ports, and written return terms is what turns a gamble into a purchase you can defend.

FAQ

Can a benchmark pass prove a used card is healthy?

No. A benchmark score alone is not proof that a card is healthy. Temperatures, fan behaviour, clock stability, and visual output matter as much, and memory faults often only appear under targeted VRAM stress. A benchmark pass plus a zero-error OCCT VRAM test, a clean physical inspection, and working ports is a far stronger picture than a benchmark pass on its own.

Does GPU-Z catch every fake card?

No. GPU-Z version 2.12.0 and later can flag some counterfeit cards with a [FAKE] label and a yellow warning, covering specific older donor chips such as G84, G86, G92, G94, G96, GT215, GT216, GT218, GF108, GF106, GF114, GF116, GF119, and GK106. Newer tampering methods exist, so a clean GPU-Z screen is not proof of authenticity. Always cross-check the die, memory type, and memory size against the manufacturer's official spec page.

Can I tell whether a card was used for mining?

Not definitively. Prior workload cannot be read from the card. Clues such as a modified vBIOS, heavy dust or oil, a replaced cooler, or VRAM errors all have non-mining explanations, and seller-context clues such as many identical listings are circumstantial. Ask the seller directly and record the answer in writing, but judge the card by its test results: a zero-error VRAM test, stable behaviour under load, and a clean inspection count for far more than any story about its past.

If the seller says warranty remains, am I covered?

Not necessarily. Transferability varies by brand and region, some brands or regions exclude mining or commercial use, and opening or repasting a card can void cover. Ask for the original invoice, check the brand's current Hong Kong policy for the exact model, and get the seller's answers in writing. A screenshot of an expiry date is not enough by itself.

Is there a cooling-off period for a private used-GPU sale in Hong Kong?

No automatic cooling-off or return right for private used-GPU sales is established by the sources used here. The Consumer Council's cooling-off item found in search results is a proposal for specific trades, not an enacted general right. Your return protection is whatever you and the seller agree in writing before payment, plus the Sale of Goods Ordinance position on rejecting faulty or misdescribed goods promptly.

What temperature should a used card run at under load?

There is no single universal pass or fail temperature. Sources disagree: one guide suggests aiming for 65 to 85 degrees Celsius under stress, another lists idle below 50 degrees and load below 85 degrees, and limits vary by model, cooler design, ambient temperature, and case airflow. Compare against the same model's typical behaviour and watch for throttling, crashes, and artefacts rather than one magic number.

Sources: TechPowerUp GPU-Z documentation and fake-card reporting via Notebookcheck and Hexus; SlashGear counterfeit-card guide; PowerColor FAQ; Sapphire RMA and warranty documentation; EVGA warranty pages; Fix Gaming Channel warranty explainer; Consumer Council Hong Kong second-hand platform findings and guidance; CLIC Hong Kong Sale of Goods Ordinance guidance; OCCT guides; PCMasterInsider VRAM stability guide; Position Is Everything GPU stress-test guide; XDA Developers and Changelly mining-wear analysis; DOTA Tech GPU trade-safety guide; MakeUseOf used-GPU tests; NVIDIA DCGM documentation.