How do I identify which process is using a dedicated server GPU?
Use the GPU driver’s process view, then inspect the matching operating-system process. Confirm the owner and workload before taking any action. A process identifier can disappear or be reused, and container process namespaces can make host and container IDs differ.
Before you start
- Permission to inspect the GPU and the relevant operating-system processes.
Instructions
1. Find the reported process
Select the intended GPU and note the PID, process type and reported memory where available. Some graphics, virtualised or partitioned configurations do not expose identical process accounting.
nvidia-smi -i 0Shows the selected GPU summary and supported process information. Index 0 must be replaced if another device is intended.
2. Map the PID to a host identity
Replace 1234 with a PID observed moments earlier. The selected output avoids printing full command arguments, which may contain secrets. A missing row can mean the process exited or is not visible in your current namespace.
ps -p 1234 -o pid,ppid,user,commReports the selected PID, parent PID, user and executable name without terminating the process.
3. Confirm the owning workload
Compare the process with the job scheduler, application service or container inventory. Ask the owner whether its memory reservation or runtime is expected. If a job needs to stop, use that application’s approved cancellation path after checking the impact and current process identity.
Check the result
- The GPU process is linked to an accountable application, job or container owner.
- Any proposed action targets the current verified workload rather than an assumed PID.
Troubleshooting
- The reported PID cannot be found inside the container
- Inspect from the authorised host namespace or map the container’s process namespace. Host and container PID values need not be identical.
Official references
Use the documentation for your installed version when a flag, service name or platform feature differs.