GPU Tuner: safety, limits and results
GPU Tuner is built to be trustworthy: it asks the driver first, validates, previews, applies in a way that can be undone, and never leaves you stuck.
The safety model#
- Ask the driver first. A control is offered only when reading its value back works and the numbers are plausible. A card that answers with nonsense (a laptop GPU reporting a 5 kW power range, seen in testing) is treated as unsupported.
- Validate. Values outside the card's reported range are refused. Fixed fan speeds cannot go below 20 %. A fan curve must keep speeds from dropping as the card gets hotter and must reach 60 % by 85 °C.
- Preview. Every change is shown as a list of steps, each with the value it will restore.
- Transactional. Steps run in order. If one fails, the earlier ones are undone in reverse. If an undo fails too, you are told exactly which setting is stuck.
- Confirm or revert. The change is applied by a root helper that waits for your Keep. Nothing else keeps it: a click on Revert, the countdown ending, the app closing, or the app crashing all undo the change.
- Startup guard. A saved profile is applied by
voltkit-gpu.service. It leaves a marker before applying and removes it after the computer has stayed up for a minute. If the marker is still there at the next start, the profile is skipped and disabled until you turn it back on, so a profile that freezes the machine cannot do it again and again. Open GPU Tuner to turn it back on. - Fans go back to the card. When the service stops, fans return to the card's own control. With no temperature reading, a software curve runs the fan at full speed.
What is tested#
Detection, planning, rollback, the confirm protocol, the startup guard and the fan governor are covered by automated tests that simulate a card failing at any step. No one can test every graphics card, driver and firmware combination, so how a control behaves on your card may differ from what has been seen so far. Start with small changes, one setting at a time.
Monitoring was checked against nvidia-smi on one NVIDIA card: temperature, utilization and clocks matched, and monitoring uses about 0.04 % of one processor core at a reading every 2 seconds.
Help improve it#
Reports from users are very welcome and are the best way to learn how GPU Tuner works across more hardware. Please send a GPU Tuner hardware report telling me which GPU you tested and how it worked, whether everything went well or something went wrong. Include the output of voltkit gpu, your distribution, kernel and driver, and what you tried.
Not built yet#
- A stability scanner that raises clocks under load and backs off on errors: it needs a load generator and real crashes to test safely.
- Comparative testing against other tools on AMD and Intel hardware.
- A separate fan curve per fan or sensor, and matching programs by window instead of process name.