How do I check GPU temperature and power limits on a dedicated server?
Read temperature, power and performance-limit information together with the workload. A lower clock can reflect normal idle behaviour, a configured power policy or a thermal constraint. Safe limits and supported telemetry depend on the specific GPU and server design.
Before you start
- The selected GPU identity and the vendor/server operating specifications.
Instructions
1. Collect the current telemetry
Replace index 0 with the intended device. Note current readings and any reported limiting reason. Treat N/A as unavailable information rather than a zero measurement or proof of normal temperature.
nvidia-smi -i 0 -q -d TEMPERATURE,POWER,CLOCK,PERFORMANCEReads the selected telemetry groups without setting clocks, changing power limits or resetting the GPU.
2. Compare equivalent workload periods
Observe the same task at a known load and compare with earlier healthy measurements. Power draw and clock speed naturally vary when a GPU is idle. A single reading should not be used to diagnose cooling failure without considering the device’s documented behaviour.
3. Escalate hardware evidence
If temperature or performance-limit evidence is abnormal under expected load, provide the server owner with the GPU identity, timestamps and relevant readings. Ask the hardware operator to review cooling, power delivery and platform events. Any power-policy or hardware intervention needs a model-specific plan and workload coordination.
Check the result
- Readings are consistent with the supported device limits and the observed load.
- An unexplained sustained limit has evidence and an identified hardware or workload owner.
Troubleshooting
- Performance is reduced while power readings look normal
- Inspect reported limiting reasons, temperature, utilisation and application bottlenecks. Power draw alone cannot distinguish thermal throttling, idle behaviour and software limits.
Official references
Use the documentation for your installed version when a flag, service name or platform feature differs.