NVIDIA GPU: High Utilization in an Initial nvidia-smi Sample

Symptom

On Linux, a single nvidia-smi invocation showed high GPU utilization despite no visible running compute application. The source image uses Tesla K20m devices and driver 319.82; it is a historical example.

Original English nvidia-smi output showing one Tesla K20m at 80 percent utilization and no running compute processes

Case explanation

The source attributes the initial reading to initialization work during the query. NVIDIA also documents that ECC memory scrubbing during driver initialization can temporarily produce high GPU and memory-utilization readings. An initial utilization sample is not, by itself, evidence of sustained application load.

Compare further samples

nvidia-smi -l

This keeps querying every five seconds by default; press Ctrl+C to stop. In the source case the reading returned to normal after several refreshes. The loop observes the GPU; it does not reset it or guarantee that persistent utilization will disappear. If readings remain high, continue investigating the actual workloads, driver state and available GPU diagnostics.

Can't find an answer?

Upload a log for AI diagnosis, or ask the AI assistant directly in the console.