How to Test the GPU in a Used Laptop Before Buying (15-Minute Protocol)
A used laptop’s GPU is the component you’re actually paying for on this site — and the one where sellers’ listings are most often wrong, sometimes innocently, sometimes not. This guide is the GPU sibling of our battery health check: what to demand before you travel, and a 15-minute on-the-spot protocol that catches the expensive problems.
Before you go: three things to demand from the listing
1. The exact GPU model and VRAM size, in writing. “RTX 3080” is not a spec — the Laptop GPU exists in 8 GB and 16 GB variants (our Legion 7 review explains why that’s a different machine). “NVIDIA workstation graphics” spans a 4 GB T500 to a 16 GB A5000. Ask for a GPU-Z screenshot showing the Memory Size field, or the machine’s service tag / type number so you can look up the factory configuration yourself (Dell and Lenovo both offer this).
2. Photos of the actual unit, vents included. Dust-caked vents and yellowed fan stickers tell you about the machine’s life. A missing bottom-panel screw or two suggests amateur surgery.
3. A straight answer about history. “Was this machine used for mining or 24/7 rendering?” A seller who dodges the question has answered it. 16 GB laptop GPUs of the 2021–22 era were mining candidates; treat all of them as suspects until load-tested.
The 15-minute in-person protocol
Bring a USB stick with GPU-Z, HWiNFO64, FurMark and (ideally) OCCT. Ask to plug in the charger — all tests happen on mains power.
Minute 0–2: identity check. GPU-Z: does the GPU name match the listing? Does Memory Size match (4096/6144/8192/16384 MB)? Does the Bus Interface show the expected width? Two minutes, and it catches the single most common used-laptop misrepresentation outright.
Minute 2–5: idle sanity. HWiNFO sensors open: idle GPU temperature should sit near 40–55 °C. Fans at idle should be quiet or off — grinding or rattling means worn bearings (£15–£30 per fan, plus your time). Check Device Manager for a yellow triangle on the display adapter.
Minute 5–15: load test. Run FurMark (1080p, no AA) with HWiNFO on the second half of the screen. You’re watching four numbers:
- Power draw should climb to the card’s expected TGP and stay there — a 3060 at ~115 W, an A5000 at ~100–110 W, an XPS-style 45 W-capped card at ~45 W. Sawtoothing far below spec = cooling failure.
- Temperature should plateau, typically 75–87 °C depending on chassis class. A sprint to 90 °C+ within two minutes means dried paste or blocked fins — fixable (£20–£40 DIY repaste) but a negotiating fact.
- Clock speed should stabilise, not stair-step downward for ten straight minutes.
- VRAM errors: FurMark artefacts (coloured confetti, flashing geometry) or a driver reset are disqualifying. If you have time, OCCT’s VRAM test for 5 minutes catches memory faults FurMark misses — degraded VRAM is the classic mining scar, and it’s not economically repairable on a laptop.
While it runs: listen (steady fan ramp is good; clicking is not), and rest a hand on the keyboard deck — scalding hot at the WASD keys after five minutes bodes ill for sustained AI workloads.
If buying remotely: ask the seller to run the same protocol and send phone video of HWiNFO during minute 10 of FurMark. Refusal is information. Buy only with returns accepted, and repeat the protocol yourself within the return window.
Reading the results: rough healthy baselines
Expected steady-state FurMark behaviour for GPUs we review (estimates — exact figures vary by chassis and power mode):
| GPU | Sustained power | Healthy plateau temp |
|---|---|---|
| NVIDIA T500 | ~25 W | under 75 °C |
| RTX 3050 Ti (45 W cap, e.g. XPS 15) | ~45 W | under 80 °C |
| RTX 3050 Ti (full fat, X1 Extreme) | ~60–75 W | under 85 °C |
| RTX A2000 / A3000 | ~50–60 W | under 80 °C |
| RTX 3060 (Legion 5) | ~115 W | under 87 °C |
| RTX 3070 Ti / A4000 | ~100–140 W | under 87 °C |
| RTX 3080 16 GB / A5000 | ~110–165 W | under 87 °C |
A machine that holds these numbers for 15 minutes will hold them for a six-hour fine-tune. That’s the entire point of the test: AI workloads are sustained loads, and a laptop that only survives benchmarks in bursts will spend its life throttling on you.
After purchase: the first-week checklist
Repeat the load test at home for a full hour (FurMark or an actual SDXL batch). Run OCCT VRAM for 30 minutes. Check nvidia-smi reports the full VRAM amount under Linux/WSL2 if that’s your working environment. If anything degrades from the shop test, that’s what the return window is for — use it, don’t rationalise it.
Summary
- Verify the exact VRAM variant with GPU-Z or a factory-config lookup before paying — “RTX 3080” alone tells you nothing on the used market.
- The 15-minute protocol: identity (GPU-Z) → idle sanity (HWiNFO) → 10-minute FurMark watching power, temperature, clocks and artefacts.
- Steady plateau = healthy; sawtoothing power or stair-stepping clocks = cooling work needed; visual artefacts or driver resets = walk away.
- Treat every high-VRAM 2021–22 card as an ex-miner until it passes a load and VRAM test.
- Buy with returns, and re-test harder at home during the window.