Buying a used GPU for mining: how to avoid a dead one
The second-hand market is where farm GPUs come from, and it breeds more myths than facts. "Mined on — throw it away" and "a gaming card is always better" are equally wrong. The truth is duller: a GPU does not wear out like mechanics; what dies is fans and thermal interfaces — and those are consumables. Here is what to check when buying, and how to tell a card that was cared for from one that was abused.
The short version (for the no-time-to-read)
- "Ex-mining" is not a verdict. A card run at a lowered power limit with stable temperatures is less worn than a gaming card doing heat-up/cool-down cycles every evening.
- Check fans and thermal pads, not the dust on the shroud. Dust blows off in a minute; a dying bearing and baked thermal paste cost time and money.
- The memory vendor matters more than the card brand. The same model ships with Samsung, Micron or Hynix memory and overclocks differently — for mining that outweighs the logo on the shroud.
- An on-the-spot test is mandatory: 30 minutes under load, core and memory temperatures, zero artifacts. A seller who refuses testing is a seller to walk away from.
- A price a third below market is not a discount, it is a diagnosis. Used cards have a market price, and a suspiciously cheap one always has a reason.
The "killed by mining" myth
Silicon and memory degrade from heat and current, not from the name of the workload. Mining done properly is round-the-clock work at a reduced power limit, stable temperature and constant fan speed. Physics says this regime is gentler than gaming: there are no thermal cycles, the very thing that fatigues solder joints and capacitors (details in our article on temperature and lifespan, the Coffin–Manson law).
What actually dies in mining:
- fans — mechanics with a finite lifespan; they spin 24/7 and honestly use it up. They are cheap to replace, question closed;
- thermal paste and pads — dry out over 2–3 years under heat. Replacement is routine maintenance, not repair;
- the card's power delivery — only suffers if the card was run at maximum watts in hot ambient. That shows on the power connectors.
The conclusion neither camp likes: judge a used card not by "mined or not", but by the state of its consumables and its test results.
Inspection and test: a 15-minute checklist
- Fans. Spin them by hand: quiet, no wobble, no grinding. Ask for a load test and listen at full speed — rattling or pulsing means a dying bearing (how a fan predicts failure).
- Power connectors. Darkened plastic and discolored contacts mean the card lived hot at high wattage. That is a price cut or a pass.
- Signs of disassembly. Scuffed screws and broken warranty seals mean someone has been inside. Not a problem if the paste and pads were replaced — ask directly. A problem if "nobody touched it" and the seals are gone.
- A 30-minute load test. Any stress test or miner. Watch: core and memory temperatures (on hot GDDR6X — a dedicated article), no screen artifacts, stable clocks without sudden drops.
- An identification utility. Verify the model, memory size and memory vendor — sometimes a different revision is sold under a familiar name.
Brands, revisions and memory
The same NVIDIA chip ships from dozens of board partners — ASUS, MSI, Gigabyte, Palit, Zotac, Colorful and others — with different power delivery and coolers. For mining, what matters most is invisible on the box: how the memory is cooled and how thick the heatsink is. A three-slot cooler with proper memory contact beats the name on the shroud.
The second invisible factor is the memory chip vendor. The same card model ships with Samsung, Micron or Hynix memory, and their overclocking headroom differs. On memory-bound algorithms that is a visible hashrate gap at the same watts. The vendor is identified by utilities in seconds — ask for a screenshot or check during the test. Our catalog stores presets separately per memory vendor — it also shows what to expect from a specific model.
One line on firmware and modded BIOS: a card with non-stock firmware is somebody else's experiment with an unknown history. For a farm, predictability is worth more than promised percentages.
Where automation fits
A used card almost never matches the "average from the forum": fresh pads and replaced fans give it their own thermal profile, and the specific memory gives its own OC ceiling. Forum presets made for someone else's card work poorly here. OCBot tunes a stable overclock for the specific card and verifies it over long runs — and a good preset is remembered by model and memory vendor, so the second identical card gets configured faster than the first.
This article is general information summarizing public practice; a specific card's condition is determined only by inspection and testing. OCBot is an independent project: not affiliated with NVIDIA, HiveOS or Hiveon. Use at your own risk — see the disclaimer.