How can I tell if a used GPU has a defective VRAM module?

The first-48-hours test protocol: run memtest_vulkan for at least 2 passes, sustained load at full throttle for 1+ hour, and monitor core/memory temps. Silent VRAM failures show up under sustained pressure or during long test passes.

What's the difference between ex-mining, gamer, and refurb cards?

Mining cards ran hot for months but at stable clocks; gamer cards have unknown history and may be thermally abused; refurb cards are officially reconditioned but carry no historical use data. Pricing reflects risk — miners are cheapest, refurbs highest. All are valid if tested properly.

Is a 48-hour return window enough to catch defects?

It is tight but often enough for catastrophic failures (won't boot, immediate thermal throttle, outright VRAM errors). Silent VRAM degradation and failure-under-load patterns may not surface in 48 hours. If you plan to keep the card long-term, longer return windows (30+ days) or seller guarantees matter more.

Should I buy from eBay, Facebook Marketplace, or a specialist GPU seller?

eBay: seller feedback is verifiable, return policy is clear, and disputes are arbitrated. Marketplace: cheaper, no platform protection, and refund risk is yours. Specialists (Newegg Refurbished, GPU retailers): higher prices but official guarantees. Start with eBay for used cards, verify seller reputation first.

What thermal pads and paste should I use if I repaste?

Use thermal paste with 5+ W/m·K conductivity (Thermal Grizzly Kryonaut, GD900) and 5mm 1–2 W/m·K thermal pads on memory modules. Avoid paste alone — VRAM modules need pad contact to the cooler. Apply thin and even; excess paste traps air and raises temps.

How to Buy a Used GPU Without Getting Burned: Vetting, Testing, and Red Flags

Buying a used GPU can cut your cost in half — but half the savings evaporate if the card has a hidden defect and fails outside the return window. The used market is where the real value lives for local LLM inference (a used RTX 3090 at ~$500–$800 is the community standard for good reason), but the risk is real: defective VRAM modules that do not fail on boot, thermal damage from mining, and cards with creeping clock instability.

This guide is the infrastructure under every used 3090 buying guide and sub-$500 card recommendation on the site. It walks you through the signals you can read before you bid, the first-48-hours test protocol that catches most defects, and the return-window discipline that lets you walk away guilt-free if something is wrong.

Before you bid: seller red flags and listing signals

A bad seller signal saves you from a bad card. Before clicking “Buy It Now,” read the listing and the seller’s history like you are auditing them.

Seller reputation matters. On eBay, check feedback percentage (aim for 99%+ with 100+ sales) and read the last 20 comments for patterns. Look for:

Multiple “item not as described” disputes in the last 30 days
Phrases like “sold for parts” or “unknown condition” in recent feedback
Seller with <3 months history selling GPUs (they may disappear after a bad sale)

A seller with clean feedback on 50+ electronics sales is trustworthy; a seller with 4 GPU sales and one refund dispute is a coin flip.

Listing red flags that should make you skip:

Stock photo instead of real images. If the listing shows a manufacturer’s product shot instead of a photo of this specific card, walk. You cannot inspect the cooler, heatspreader condition, or bracket integrity. Some sellers do this to hide damage.
No mention of use history. “Gently used” is not a use history. Ask via messages: “Was this card used for gaming, mining, or professional work? How many hours per day?” A seller who will not answer is hiding something.
Vague or missing specification info. Does the listing say “RTX 3090” but not confirm 24GB VRAM? Search the model number yourself and confirm specs. Counterfeits and rebadged cards are rare but not zero.
Price too low to be true. A pristine used 3090 at $300 when the market is $500–$800 is either stolen, non-functional, or about to be. Verify the price against recent eBay sold listings and the gpu-market digest; if it is 30%+ below fair value, ask why.
Seller unwilling to offer return. “No returns” on a used GPU is a red flag. Push for at least 14 days; many eBay sellers will accept 30-day returns if you message first.

Reading the card’s history: thermal abuse and mining signs

Even a good seller may not know their card’s true history. You can infer it from condition and behavior clues.

Mining cards: These ran 24/7 at stable (if hot) clocks for weeks or months. Signs:

Thermal pads are discolored or hardened (mining = low-power draw but long runtime)
The heatsink and cooler are clogged with dust
Core clock is locked or underclocked in BIOS (miners undervolt for efficiency)
Seller mentions “stable for 6+ months” (miners boast about uptime)

Mining cards are not inherently bad. They ran at consistent temperature and load, which is less punishing than gaming thermal spikes. The trade-off: you cannot inspect the card’s true age without opening it, and failed capacitors or memory controllers may show up weeks later. If the discount is steep (>20% below fair price), factor in the higher risk.

Thermal abuse and gamer cards: Gaming involves thermal cycling (on, off, on, off, heat stress, cool-down, repeat). Signs:

Cooler mounting bracket is bent or loosened
Thermal pads are cracked or missing
Heatsink has visible corrosion or is separated from the board
Core temperature is high on boot (before any load)

Gamer cards have unknown thermal history. They could have been gently used or hammered in a mining rig someone is lying about. Demand photos, inspect the cooler, and test aggressively once you own it.

Refurbished cards: Officially reconditioned cards (Newegg Refurbished, Reconditioned) have no use history — they were returned, tested in a warehouse, and reboxed. Condition is often pristine and they carry 30–90 day warranties. The downside: you do not know why they were returned in the first place, and refurb pricing is often only 10–15% cheaper than new. Good choice if you value the warranty more than the savings.

The first-48-hours test protocol: catch defects before the return closes

You have a narrow window (often 48 hours, sometimes 14 days). Use it ruthlessly. Do not assume the card “looks fine” — run a test suite that stresses the failure modes you cannot see.

Before you even power it on:

Inspect the physical card. Check for:
- Bent pins on the power connectors (even one means it may fail under load)
- Cracks in the PCB or heatspreader
- Missing capacitors or obviously desoldered components
- Bulging or leaking capacitors (rare, but catastrophic)
Clear the cooler of dust. Use compressed air on both the heatsink and the GPU die area.
Check the thermal pads. If they are hard and cracked, plan to repaste before stress testing (dried paste will thermal-throttle the card and mask real speeds).

The test suite (run in order, same day as power-on):

Stage 1: Boot and baseline (30 min)

Install the card in a stable system.
Boot and confirm it is recognized. Check GPU memory in Windows Device Manager or nvidia-smi (Linux). VRAM should match spec (24GB for a 3090, no errors reported).
Run a quick 3D benchmark (Furmark, 3DMark, or VRAM calculator at full load) for 10 minutes. Core clock should not throttle below ~2.3 GHz on a 3090 (varies by card). Temperature should stabilize around 70–80°C on the core; memory should sit 5–10°C lower.
If the card fails to boot, has memory errors, or throttles immediately, initiate a return now.

Stage 2: VRAM stress test (2+ hours) This is the stage that catches silent VRAM failures. Use memtest_vulkan (github.com/stohrendorf/memtest_vulkan), a Vulkan-based memory test that does not require CUDA.

Run memtest_vulkan for at least 2 full passes. A single pass takes ~20–40 minutes on a 3090, depending on VRAM clock.
Watch the output for “ERROR” or “FAIL” messages. Any error means the card has a defective VRAM module or a failing memory controller. Initiate a return immediately.
If you see no errors after 2 passes, the VRAM is likely healthy. (Rare: some defects only appear under specific load patterns, but 2 passes catches 95%+ of bad chips.)

Stage 3: Sustained load and thermal stability (1+ hour)

Run a sustained benchmark (llama.cpp with a large model, Furmark, or a professional stress test like OCCT).
Let it run for 60+ minutes at full load. Watch for:
- Clock throttling (core or memory clocks dropping unexpectedly)
- Temperature creep (temps climbing past 85°C and not stabilizing)
- Thermal shutdown (rare, but the system will crash if the card hits the thermal limit)
- Artifacts or visual glitches (a sign of overclocking instability or VRAM corruption under load)
Note the stable power draw. A 3090 should sit around 300–350W under full load; if it spikes to 400W+, the cooler may be failing or the card may be unstable.

Stage 4: Cool-down and final boot (1 hour)

Let the card cool naturally (do not use extreme cooling).
Boot one more time and run a quick benchmark. If thermal throttling appears now (when the card was stable before), you may have a heatspreader separation or failing cooler mounting. Return it.

If all four stages pass, the card is likely healthy. Defective VRAM modules, thermal damage, and clock instability almost always surface within this protocol. Silent VRAM degradation (a chip that fails in months, not days) is possible but rare, and longer-term failure is a different return-window problem.

Return-window discipline: know when to walk

The return window is not a suggestion; it is your insurance policy. Use it.

If you find any error in stages 1–3, initiate a return immediately. Do not wait, do not hope it is a one-time glitch. eBay and reputable sellers understand: “Card failed stress test, returning.”
If you suspect the card is slower than advertised (you are getting 60 tok/s instead of the expected 80–110 on a 7B model), check against official specs and community benchmarks. Thermal throttling, memory clock throttling, or power limiting can all cause real-world slowdown. If you see throttling, test with the buying framework and your quantization in mind. A slower card might still be fit for your use case — do not return it reflexively unless it is outside acceptable bounds.
If the seller offers a return period shorter than 7 days, push back. Message: “Can you extend to 14 days? I need time to test stability.” Many will; if they refuse, reconsider the purchase.
Document everything. Take screenshots of stress-test results, temperature logs, GPU-Z readings, and any errors. If you do return, include these in your dispute. “Card failed memtest_vulkan with memory errors” is much harder for a seller to contest than “card is slow.”

Refurb vs. mining vs. gamer: a quick comparison

Card source	Typical price	Condition	History known?	Warranty	Risk
Refurbished (official)	-10% to +5% vs. new	Excellent	No (returned goods)	30–90 days	Low risk, but highest cost
Ex-mining	-20% to -30% vs. new	Fair (dusty, thermal pads hard)	Partially (stable clocks)	None	Medium: stable use, but age unknown
Gamer/private	-15% to -25% vs. new	Variable (good to poor)	Unknown	None	Medium-to-high: thermal cycling, unknown load
Refurb (third-party marketplace)	-20% vs. new	Variable	Minimal	Limited or none	High: no oversight, no recourse

The decision hinges on your risk tolerance and return-window length. Refurbished at retail prices is safe but expensive. Ex-mining at a steep discount is a bet that you can test well in 48 hours and the seller will honor a return. Private sales are cheapest but offer no platform arbitration — your only recourse is a chargeback, which takes weeks.

Where to buy and price anchors

eBay is the standard. Most used 3090s trade here; seller feedback is verifiable, and eBay’s Buyer Protection covers most disputes (card not as described, failure to ship, etc.).

Browse used RTX 3090 24GB on eBay →

Browse used RTX 3060 12GB on eBay →

Price anchors (as of 2026-06-29, observed on eBay and community reports): RTX 3090 used is running $500–$800 depending on condition and listing freshness; RTX 3060 used is $200–$350; refurbished official cards are 85–95% of new price. If you see a 3090 at $400, either the listing is about to close and has no bids, or the card has an obvious issue. Verify against the gpu-market digest (community-tracked pricing) before bidding.

Facebook Marketplace and Craigslist are cheaper because there is no platform protection. You inspect in person, pay cash, and leave. If the card fails the next day, you have no recourse. Only use these if you have the technical confidence to test on-site (bring a laptop and test rig if you need to).

Newegg Refurbished and Overclockers UK carry officially refurbished cards at a premium over eBay prices, but with warranties. Worth it if you value the guarantee over the savings.

If you need to repaste or replace thermal pads

Most used cards over 2 years old benefit from a repaste. If the card is thermally throttling and memtest passes, the thermal interface is likely degraded. Plan for:

Thermal paste: 5+ W/m·K conductivity. Thermal Grizzly Kryonaut on Amazon →
Thermal pads (5mm, 1–2 W/m·K): thermal pad kits on Amazon →
Small screwdrivers and patience. Most RTX cards have 6–8 screws on the cooler.

Do not skip the pads on memory modules. Paste alone will not transfer heat; the pads bridge the gap between the VRAM chips and the cooler. Apply paste thinly (a 2mm bead per die) and pads evenly. Let the paste cure for 1–2 hours before powered testing.

Bottom line

A used GPU is a good buy when the seller is verifiable, you test the card hard in the first 48 hours, and you are willing to return it without hesitation if something fails. The first-48-hours protocol (VRAM stress test, sustained load, thermal stability) catches 95%+ of defects. Do not skip it. Do not assume “the listing looked honest” means the card is honest. Test, document, and walk if something does not add up.

The used market is where the RTX 3090’s value comes from — and that value is real only when you avoid the defects that turn a $600 investment into a $600 lesson. Buy from sellers with clean feedback, demand photos, test aggressively, and use the return window.