Complete nvidia-smi commands cheat sheet for GPU monitoring, management, and diagnostics on Linux. Essential reference for ML engineers, data scientists, and GPU server administrators.
nvidia-smi is the first command on any machine with an NVIDIA GPU, and this page splits its uses in three: watching, changing settings, and querying for scripts. The notes explain what the default screen actually shows, which management commands need root and persist, and how the query flags turn nvidia-smi into a monitoring source.
Search all nvidia-smi (GPU) shortcuts interactively
Interactive Shortcut Finder| Shortcut | Action | Description |
|---|---|---|
| nvidia-smi | GPU status summary | Show GPU utilization, temperature, memory, and processes. |
| nvidia-smi -l 1 | Monitor every 1s | Refresh GPU status every 1 second. |
| watch -n 1 nvidia-smi | Real-time monitor | Real-time GPU monitoring with watch command. |
| nvidia-smi -q | Detailed info | Show all detailed GPU information. |
| nvidia-smi -L | List GPUs | List all GPUs with UUIDs. |
| nvidia-smi pmon | Process monitor | Monitor per-process GPU usage. |
| nvidia-smi dmon | Device monitor | Monitor device metrics every second. |
| nvidia-smi topo -m | Topology | Show GPU NVLink/PCIe topology. |
| nvidia-smi nvlink -s | NVLink status | Check NVLink connection status and bandwidth. |
| Shortcut | Action | Description |
|---|---|---|
| nvidia-smi -pm 1 | Persistence Mode ON | Keep driver loaded to reduce GPU init latency. |
| nvidia-smi -i 0 -pl 250 | Set power limit | Limit GPU 0 max power to 250W. |
| nvidia-smi -i 0 -ac 1215,1410 | Set clocks | Set memory/graphics clock speeds. |
| nvidia-smi -rgc | Reset clocks | Reset GPU clocks to default. |
| nvidia-smi -r -i 0 | Reset GPU | Reset GPU 0 (resolves ECC errors etc). |
| nvidia-smi -e 1 | Enable ECC | Enable ECC memory correction. |
| nvidia-smi -q -d POWER | Power details | Show detailed GPU power information. |
| Shortcut | Action | Description |
|---|---|---|
| nvidia-smi --query-gpu=name,memory.total,memory.used --format=csv | CSV query | Output GPU info in CSV format. |
| nvidia-smi --query-gpu=utilization.gpu,temperature.gpu --format=csv -l 5 | 5s CSV logging | Log utilization and temp every 5 seconds. |
| nvidia-smi --query-compute-apps=pid,used_memory --format=csv | Per-process memory | Show memory usage per GPU process. |
→ Related guide: GPU cluster management: the full command guide
Run 'nvidia-smi' to see a summary of all GPUs including utilization, temperature, memory usage, and running processes.
Use 'watch -n 1 nvidia-smi' or 'nvidia-smi -l 1' to refresh GPU status every second.
Run 'nvidia-smi --query-compute-apps=pid,used_memory --format=csv' to see memory usage per process.
Run 'nvidia-smi -pm 1' as root to keep the NVIDIA driver loaded, reducing GPU initialization latency.
Use 'nvidia-smi topo -m' to see GPU interconnect topology including NVLink and PCIe connections.
Yes — use My Stack to combine nvidia-smi (GPU) shortcuts with any other platform on this site into one printable reference, which is useful if your daily workflow spans several tools.
Browse shortcuts for 268 platforms
Explore Allnvidia-smi prints driver and CUDA versions, then one row per GPU with temperature, power draw against limit, memory used and utilisation, and a process list below. nvidia-smi -l 1 repeats every second and watch -n 1 nvidia-smi does the same with a cleared screen. nvidia-smi dmon streams a compact per-GPU line of clocks, utilisation and memory, and nvidia-smi pmon the same per process — better than the default screen for spotting a job that holds memory but does no work. nvidia-smi -q dumps every detail including ECC counts and retired pages, and nvidia-smi -L lists GPUs with UUIDs. nvidia-smi topo -m shows GPU interconnects and nvidia-smi nvlink -s NVLink status.
nvidia-smi -pm 1 enables persistence mode so the driver stays loaded between jobs, removing several seconds of startup latency; it needs root and resets on reboot unless the persistence daemon runs. nvidia-smi -i 0 -pl 250 sets a power limit in watts on GPU 0, the usual lever for a hot rack, and nvidia-smi -i 0 -ac 1215,1410 pins memory and graphics clocks for reproducible benchmarks, with nvidia-smi -rgc resetting them. nvidia-smi -e 1 enables ECC (reboot required) and nvidia-smi -r -i 0 resets a GPU that is wedged, if no process holds it. nvidia-smi -q -d POWER shows the power section alone.
nvidia-smi --query-gpu=name,memory.total,memory.used --format=csv prints selected fields as CSV; --format=csv,noheader,nounits makes it parseable, and nvidia-smi --query-gpu=utilization.gpu,temperature.gpu --format=csv -l 5 logs every five seconds into a file for a run's duration. nvidia-smi --query-compute-apps=pid,used_memory --format=csv lists processes with their memory, which is how a job's footprint is recorded. nvidia-smi --help-query-gpu lists every available field.
Open your assistant with this page preloaded as the source — great for follow-up questions like "which of these work in other apps?"