Used RTX 3090 buying checklist

A used RTX 3090 is a card with an unknown past: it may have run a mining rig around the clock for years. The checks below catch the failures you can see before and right after you pay. They don’t prove a card will last. Nothing can.

Know what you are buying

• The reference numbers. NVIDIA lists the RTX 3090 with 24 GB of GDDR6X on a 384-bit bus, 350 W of graphics card power, a 750 W system power requirement and a 93 °C maximum GPU temperature. The Founders Edition is 313 mm long and three slots wide, and takes two 8-pin PCIe cables through the 12-pin adapter in the box. Partner cards vary in length, width and connector count, so check the exact model against your case and PSU; the PSU sizing guide covers connectors and wattage for one card or several.
• Founders Edition or partner card. A missing 12-pin adapter on a Founders Edition is a problem only if your PSU has no native 12-pin cable. Ask for it anyway.
• Half the memory sits on the back. The 3090 carries GDDR6X on both sides of the PCB. The back modules rely on the backplate for cooling, which is why VRAM temperature is the 3090’s weak spot.

Ask the seller

• Was it mined on, for how long, and at what settings? PCWorld suggests asking how long the card was mined, whether it was underclocked, whether it ran stable, and what temperatures it saw. Evasive answers are an answer.
• Has it been opened? New pads or paste are fine when done well, so ask what was replaced and with what.
• Can you see it running? A short video of the card in a working system under a stress test beats any photo.
• Can you return it? Buy where a return is possible after you have tested the card. The tests below are only worth running if a failed card can go back.

Inspect the card

• Fans first. Fans wear out on cards that ran around the clock. Spin each one by hand: it should turn freely, without grinding or wobble.
• Look for dust, corrosion and damage. Thick dust or bent fins say how the card was kept. Check the PCIe fingers and the power connectors for scorch marks or melted plastic.
• Founders Edition VRAM pads. PCWorld singles out the RTX 3080 and 3090 Founders Edition for factory thermal pads that let the VRAM run hot. On a Founders card, budget time for a repad if memory temperatures come out high.

Test it under load

Steps 1 to 3 run on a Linux box with the NVIDIA driver installed. Step 4 needs a Windows boot with HWiNFO.

1. Check what the system sees. nvidia-smi -q lists the card, its power limits and the PCIe link. Under FB Memory Usage, Total should read about 24 GB (24576 MiB); anything less is not a full RTX 3090. That total is what the VRAM sizing guide sizes models against. The current link generation and width can drop when the GPU is idle, so read them again under load.
2. Run a compute burn. gpu-burn is a multi-GPU CUDA stress test. The RTX 3090 is compute capability 8.6, so build it with make COMPUTE=86 (the default targets 7.5), then run gpu_burn -m 90% 1800 for half an hour on 90 % of the VRAM. With Docker, build with docker build --build-arg COMPUTE=86 -t gpu-burn . and run docker run --rm --gpus all gpu-burn ./gpu_burn -m 90% 1800. At the end it prints OK or FAULTY per GPU. FAULTY means the card goes back.
3. Watch temperatures and clock events. In a second terminal, run nvidia-smi -q -d TEMPERATURE,PERFORMANCE -l 60 to print a report every 60 seconds. The temperature section shows the current GPU temperature next to the slowdown and shutdown thresholds. The performance section lists clock event reasons. A card that keeps hitting HW Thermal Slowdown in a cool room needs new paste or pads.
4. Read the memory junction temperature (Windows). Boot Windows, put the card under sustained load there, and watch HWiNFO, which reads the GDDR6X memory junction temperature. Igor’s Lab cites 110 °C as the maximum set by NVIDIA and Micron. In their tests the stock Founders Edition hit that limit within seconds under a memory-heavy load and throttled the memory to stay there. It held about 100 °C in a ten-minute 3DMark loop. A card that sits at the limit in normal inference is a repad candidate.
5. Watch for artifacts and shutdowns. A stress test is what exposes screen artifacts, driver resets and sudden power-offs. Any of them is a fail.

Settle it in the box

• Cap the power. nvidia-smi -pl sets the maximum power limit in watts (it needs root). nvidia-smi -q shows the card’s Default Power Limit. If the box runs hot or the PSU is tight, setting a limit below that default is the first lever to pull. The limit resets at every reboot; the PSU sizing guide covers what a limit does and doesn’t cap, and how to reapply it at boot.
• Give the backplate air. With memory on the back of the card, airflow over the backplate matters more than on most GPUs, especially with cards stacked close together.

Parts in this guide

Store links may earn us a commission at no cost to you. As an Amazon Associate we earn from qualifying purchases.

Related

• GPU
• NVIDIA GeForce RTX 3090 (Founders Edition)
• 48 GB build

Last reviewed 2026-11-12