Stock NVIDIA fan profiles are tuned for a gamer who plays for three hours and cares about noise. A rental GPU runs flat out for three days and nobody is in the room to hear it. Left on the default curve, your cards will idle their fans until the die is already hot, then chase the temperature instead of preventing it. A custom curve fixes that in about five minutes.
Why stock fan curves fail at 24/7 hosting
The default profile optimizes for acoustics. Fans stay near idle until the core passes roughly 60 C, ramp gently, and settle wherever the card's target temperature lands, often 75 to 83 C under sustained load. For a desktop under a desk, fine. For a hosting rig, three problems:
- Sustained load is the normal case. A training job holds the card at 100% for days. The stock curve lets temperatures sit near the throttle point the entire time, which means lower boost clocks and slower results for your renter.
- VRAM runs hotter than the core. Memory junction temperatures on GDDR6X and GDDR7 cards can run 15 to 20 C above the core reading. A curve keyed to a cooler core lets memory bake.
- Heat is cumulative in a rig. Four cards in one chassis feed each other exhaust. The curve that held one card at 70 C holds four cards at 85 C.
Long-term, sustained high temperature is also the main thing that ages fans, thermal paste, and pads. You bought the hardware to earn for years, not eighteen months.
What a good hosting curve looks like
The goal is a temperature target, not a fixed fan speed. Pick the core temperature you want the card to hold under full load and let the fans do whatever it takes to stay there. For 24/7 duty, 65 C is a solid target: far from throttle, easy on VRAM, and reachable for most air-cooled cards without running fans at 100%. If your ambient is high or the chassis is dense, 70 C is still a big improvement over stock.
Noise is not a real constraint in a garage or basement, so err on the aggressive side. Fans are the cheapest replaceable part on the card.
Set the curve with one script
On a headless Linux host the practical tool is the open source GPU_FAN_OC_Manager script, which drives nvidia-settings for you and targets a max temperature across every card in the machine:
wget https://github.com/jjziets/GPU_FAN_OC_Manager/raw/main/set_fan_curve
chmod +x set_fan_curve
./set_fan_curve 65The argument is your target temperature. The script checks temperatures every second and adjusts fan speed to hold the target. It needs the NVIDIA driver, nvidia-smi, and nvidia-settings installed, which you already have if the machine is listed on a marketplace.
To survive reboots, run it inside a screen session started from a @reboot crontab entry. The exact crontab line, plus commands for checking status and stopping the script, are in the custom fan curve guide in our open source Linux tools collection.
Verify after every driver update
ps aux | grep set_fan_curve and watch temperatures under load before you trust the machine again.Prove it under load
A curve you have not tested is a guess. Put the cards under real load and watch:
nvidia-smi -l 5You want the temperature to climb, then plateau at your target and stay there for hours, with clocks holding steady and fan speed settling somewhere below 100%, so there is headroom left for a hot summer day. If the fans are pinned at maximum and the card still drifts past your target, the curve has done all it can: the fix is airflow, spacing between cards, or a lower power limit, not a more aggressive script. The best time to find all this out is during a proper burn-in test before the machine goes on the market, when a cooling problem costs you nothing but time.
Fan curves and power limits work together
A fan curve attacks temperature from the cooling side. A power limit attacks it from the heat-generation side, and it cuts your electricity bill at the same time, usually for a performance cost of only a few percent. Most serious hosts run both: the power limit sets how much heat exists, the fan curve decides where it goes. Next step: GPU power limits for hosting.