If you just bought or are about to buy a graphics card, test software is how you confirm you got what you paid for. I use it for three things: checking a new card boosts correctly and stays cool, spotting a used card with degraded memory or a flashed BIOS, and proving an overclock or undervolt is actually stable.
No single program does all three well. Synthetic stress tests find heat and power problems fast, but they do not look like games. Benchmarks give you a comparable score, but a good score can hide stutter. Monitoring tools show you what is really happening. Most bad diagnoses I see come from running only FurMark and calling it done.
As an Amazon Associate we earn from qualifying purchases at no extra cost to you.
Top 3 picks at a glance
What you actually need to test
For a new AMD, Intel, or NVIDIA card, you need a repeatable benchmark, a 20-30 minute stability load, and temperature, hotspot, memory temperature, clocks, fan RPM, and power draw logged while it runs. For a used card, you need the same plus an artifact scan and a VRAM-specific test. Artifacts look like flashing triangles, checkerboards, black boxes, or green/purple dots. One crash in a game can be a driver. Artifacts plus driver crashes in multiple tests usually means the hardware.
Do not test with the side panel off and a desk fan blowing in unless you always run that way. You want realistic case airflow. Close background recording, updaters, and browsers, set Windows power to Balanced or High Performance, and install the current driver clean before you compare scores.
Quick comparison
This is the short list I return to. All of these have usable free versions. Paid versions mostly add automation and no ads, they do not test better.
| Software | Best for | Cost | Watch out for |
|---|---|---|---|
| 3DMark Time Spy / Steel Nomad | Comparable score for new GPU shopping | Free demo on Steam, paid for full suite | Score varies with CPU, RAM, and driver; compare same preset only |
| Unigine Superposition | Free game-like load + stability check | Free Basic | Old engine, does not stress ray tracing or upscaling |
| OCCT GPU + VRAM test | Finding errors on used cards | Free with limits | Very heavy; stop if hotspot passes 105-110C on AMD or 100C+ sustained elsewhere |
| FurMark 2 | Quick heat / cooler / fan check | Free | Unrealistic power virus load; some cards throttle by design |
| GPU-Z + HWiNFO + CapFrameX | Monitoring clocks, temps, 1% lows | Free | Sensors differ by brand; memory temp missing on some models |
Best all-around benchmark: 3DMark
If you are comparing cards before you buy, 3DMark is still the common language. Time Spy is useful for older midrange cards, Steel Nomad is better for modern cards like an RTX 4070 and above because it is heavier and does not need ray tracing hardware to run well. Port Royal and Speed Way are only worth running if you specifically care about ray tracing performance.
Trade-off: it costs money for the full suite outside Steam sales, and the free demo only lets you run once per preset without custom loops. Scores also move 2-5% with background tasks, Resizable BAR on/off, and driver versions, so do not return a card over a 300-point difference. Who it suits: buyers deciding between two tiers, reviewers, and anyone who wants a before/after number for an undervolt. Failure mode I see often: people compare Time Spy Extreme Graphics scores to regular Time Spy scores and panic. They are different presets.
Best for used cards and crashes: OCCT
OCCT is my first pick when someone asks me to check a used card. The GPU 3D Adaptive test with error detection enabled will log errors even if you do not see artifacts on screen, and the dedicated VRAM test fills memory in a way games rarely do for long stretches. That matters because failing VRAM is the most common used-card killer, especially on ex-mining cards.
Run 30 minutes for confidence, 10 minutes for a quick filter. If you get thousands of errors in the first minute, stop. The card is bad or severely unstable. If you get zero errors but the PC reboots, suspect power supply or riser cable before the GPU. Trade-off: OCCT is deliberately brutal. High hotspot delta, for example 65C core but 105C hotspot on AMD, usually means bad paste or mounting pressure, not necessarily a dead chip. Who it suits: Facebook Marketplace, eBay, and trade-in buyers. The free version is fine; you do not need Pro for a one-time check.
Best free game-like load: Unigine Superposition
Superposition 1080p Extreme and 4K Optimized are excellent free stability loops. Unlike FurMark, it loads the core, memory controller, and fans in a pattern closer to an actual game, and any stutter, flicker, or texture corruption is easy to see. I loop it for 30 minutes after any overclock.
It is also useful for Intel Arc owners. Arc drivers have improved a lot, but older DirectX 9/11 engines can still behave oddly. If Superposition is clean but one specific game crashes, that points to a game/driver issue, not a faulty card. Trade-off: scores cannot be compared across different quality presets, and ray tracing / DLSS / FSR / XeSS performance is not tested. Pair it with a real game capture if you are choosing between an RX 7800 XT-class card and its NVIDIA equivalent.
For heat and fans: FurMark 2, used carefully
FurMark 2 is not a buying benchmark, it is a cooler test. Five to ten minutes at your monitor’s native resolution will show max core temp, hotspot, memory junction, fan noise, and whether the card holds its rated boost. I still use it to find rattling fans, dried paste, and cases with no exhaust.
Do not fail a card just because it power-throttles in FurMark. Modern NVIDIA and AMD cards intentionally limit clocks in that workload. If FurMark hits 85C with loud fans but Superposition and games sit at 70C quiet, the card is probably fine and your case airflow is adequate for real use. Stop the test immediately if you smell hot plastic, see temps climb past 90C core / 110C hotspot and keep climbing, or the screen goes black and does not recover. Who it suits: checking repastes, SFF builds, and triple-fan vs dual-fan cooler differences.
Do not skip monitoring: GPU-Z, HWiNFO, CapFrameX
Scores without sensors mislead. GPU-Z validates the BIOS, bus width, memory type, and PCIe link speed. That catches the classic scam: a relabeled old card flashed to report as a newer model. HWiNFO logs per-sensor power, hotspot, memory temperature, and thermal throttling flags. CapFrameX or PresentMon measures average FPS plus 1% and 0.2% lows, which is what you actually feel.
A practical pass for a new card: run Steel Nomad or Superposition while logging with HWiNFO, then play one demanding game for 20 minutes with CapFrameX. You want steady clocks within 100-150 MHz of reviews for your model, fans under 1800-2000 RPM in a normal case, no throttling flags, and 1% lows within 60-70% of average FPS. If average FPS is high but 1% lows collapse, try a different driver or check CPU/RAM before blaming the GPU.
A simple 45-minute routine for new and used GPUs
For a new card in a mid-tower with decent airflow, I do: 1) GPU-Z check and 5-minute FurMark 2 for temps and noise, 2) one run of 3DMark Steel Nomad or Time Spy for a baseline score, 3) 30-minute Superposition or OCCT 3D loop for stability while logging HWiNFO. For a used card, add OCCT VRAM for 20 minutes. Any artifacts, errors, or black screens across two different tests means send it back.
If you are still shopping and need a test platform, it helps to buy around a known quantity. Builders testing 1440p cards often keep something like an Intel Arc A770-class GPU or its AMD/NVIDIA competitor as a baseline, then compare scores on the same CPU, RAM, and driver branch. Changing two parts at once makes results useless.
FAQ
What is the single best GPU test for a beginner?
Run Unigine Superposition 1080p Extreme once, then loop it for 30 minutes while watching temperature and fan noise. It is free, hard to misconfigure, and close enough to gaming to catch most bad cards and bad overclocks.
Is FurMark bad for your GPU?
It will not kill a healthy modern card because drivers throttle to protect it, but it creates unrealistic heat and power draw. Use it for 5-10 minutes to check cooling, not for hours, and stop if temperatures keep climbing past safe limits.
Why does my benchmark score not match YouTube reviews?
Different CPU, RAM speed, Resizable BAR setting, driver version, Windows power plan, and background apps easily shift results 3-8%. Match the exact preset and resolution, update drivers, close other software, then compare again.
How do I test a used GPU before the return window closes?
Verify it in GPU-Z, run OCCT GPU error check plus VRAM test for 20-30 minutes, then play your heaviest game for 30 minutes. Reject the card if you see artifacts, logged errors, repeated driver crashes, or hotspot/memory temperatures far above normal for that model.
