Stability test · GPU
GPU stress test.
Every chunk of work is calculated twice and compared bit for bit, and the card’s memory is checked with patterns. The test looks for errors, not a temperature record. Useful after overclocking, after an undervolt and for games that crash the driver.
- NVIDIA, AMD and Intel
- 1 min to 1 h for free
- No time limit in PRO
VRAM buffers, 64 MB each
An illustration. The real window shows the VRAM buffer map, card sensors and counters of checked results.
What the test does
Calculates twice, compares bit for bit.
Compute work in Direct3D 11, with no vendor specific paths. The card calculates the next chunk while the CPU checks the previous one.
Compute compared bit for bit
2 million threads, a loop on integers and floating point numbers. Every chunk is calculated twice, and the integer part is also checked on the CPU on random samples.
Video memory
Half of the VRAM, from 256 MB to 4 GB, in 64 MB buffers. Four patterns in turn: random, address as data, inverted random and a walking bit. The data sits for seconds under full load.
Short chunks
One chunk takes about 40 ms of GPU time, and a weaker card gets fewer threads. That way Windows never resets the driver for no reason.
When to use it
A game crashes the driver? Check the card.
An unstable card often does not freeze the whole PC. It shows artifacts, a frozen image or a driver reset message, and only in some games.
After overclocking
After changing the core or memory clock. On NVIDIA you can set them in the GPU tab in PRO.
After an undervolt
Lower voltage at the same clock means less heat. Too low and the card starts calculating wrong or loses the driver.
Crashes and artifacts in games
Driver resets, black screens, coloured dots or a frozen image. The test helps check whether the card is to blame.
A new or used card
A few minutes of testing right after installing it or buying second hand, before the return window closes.
After cleaning or new paste
Live core and hotspot readings during the test show whether cooling works as well as before you took the card apart.
After a driver update
A new driver can treat an overclocked card differently. A short test after updating shows whether your settings still hold.
How to read the result
Passed, failed or stopped.
Both passes of every chunk gave the same result, the VRAM patterns matched and there was no driver reset and no WHEA event.
A difference between passes, a bad read from video memory, a driver reset or a WHEA event. A driver reset ends the test right away.
You stopped the test, the core hit 90 °C or the hotspot 105 °C in three readings in a row, or a laptop switched to battery.
- A difference in calculationsThe GPU core calculates wrong: most often the core clock is too high or the voltage too low. Undo your last change.
- An error in video memoryUsually the VRAM is overclocked too far. Lower the memory clock offset.
- A driver resetThe card stopped responding. At factory settings, also check power, the driver and temperatures.
- Stopped by temperatureThis is not a stability error, it means the cooling cannot keep up. Check the fans and airflow.
To be honest: the test checks compute and video memory. It does not yet include a looping 3D scene, so a problem that only shows up while rendering a specific game can slip past it. For smoothness in a scene, use the GPU arena.
Questions
FAQ
How is it different from just loading the card?
Every chunk is calculated twice and compared bit for bit, and the card’s memory is checked with patterns. The test does not chase a temperature record, it looks for calculation errors.
What does “graphics driver reset” mean?
Windows reset the driver because the card stopped responding. The test counts it as an error and ends right away. On an overclocked card it usually means the clock is too high or the voltage too low.
Why are the work chunks so short?
One chunk takes about 40 ms of GPU time. Windows resets the driver when a card works on one job for too long, so short chunks never trigger a false reset.
How much video memory is checked?
Half of the VRAM, from 256 MB to 4 GB, in 64 MB buffers. Four patterns in turn: random, address as data, inverted random and a walking bit.
When does the test stop by itself?
When the GPU core hits 90 °C or the hotspot 105 °C in three readings in a row, when a laptop switches to battery, and on a WHEA hardware error event.
Does it work on AMD and Intel cards?
Yes. The test uses Direct3D 11 with no vendor specific paths. Temperature readings depend on the card’s sensors.
Overclocked card? Check that it calculates right.
The GPU stress test is in the Benchmark tab in SmoothWizard. 1 minute to 1 hour in every version, with no time limit in PRO.
