Temperature, fans and power
Both controls on this page need the agent installed with full privilege.
With read-only you see every number but change nothing — see
installing the agent.
Fan policy
Section titled “Fan policy”Three modes, per machine:
Driver
Section titled “Driver”The card manages its own fans and stops them when it is cool. This is the default behaviour of a GPU nobody has told otherwise.
Fine for a machine that is idle a lot. Less good for sustained rental load, where the card’s own curve is tuned for quiet rather than for memory temperature.
Fixed speed
Section titled “Fixed speed”One duty cycle regardless of load — predictable noise and predictable airflow.
Below roughly 30% most cards are not controllable, and the panel says so at the point where you drag the slider rather than after you save.
Target temperature
Section titled “Target temperature”You name a temperature; the agent picks the speed to hold it. This is the mode worth using on a rented machine: load varies with whoever booked it, and a fixed duty cycle is either too loud at idle or too weak under a long-running job.
Pick which sensor the target applies to:
| Sensor | Hold it when |
|---|---|
| Hottest | you want whichever of the two is closer to its limit — the safe default |
| Core | the card has no separate memory sensor, or memory is not the constraint |
| Memory | GDDR6X, where the memory junction runs about 20 °C hotter than the core |
A card without the sensor you picked is held by its hottest one instead of being left uncontrolled.
Power limit
Section titled “Power limit”Sets nvidia-smi -pl in watts, for all cards on the machine.
A lower limit cuts draw and heat — and throughput. That trade is usually worth it, because the top of a card’s power range buys very little performance for a lot of watts, but it is a trade and the panel shows you both sides before you commit: what the electricity costs at that limit, and how far it is from the limit currently in force.
Two cases where the panel stops you rather than guessing:
- The cards have not reported their power range. No range means no safe bounds for the slider. Check telemetry and the agent version.
- The cards have different ranges. Per-card limits are not supported yet, and applying one number to mixed cards would silently clamp some of them.
Reading the result
Section titled “Reading the result”After a change, the Overview tab is where you check it worked: the draw should move within a couple of telemetry cycles, and the temperature follows more slowly.
If neither moves, the Activity tab will show whether the command was accepted by the machine at all.