Skip to main content
For GPU runs, Lyceum Cloud records DCGM-sourced GPU metrics and system telemetry into Prometheus. The metrics endpoint queries that data per execution, so you can attribute time, debug stalls, and confirm a job actually used the GPU it was billed for.

​
Endpoint

By default the endpoint returns the entire execution at 15-second resolution. Narrow the window or change the step when you’re zooming into a specific phase.

​
Available series

​
GPU (via DCGM)

​
System

​
Example

​
What it’s good for

  • Confirming GPU utilisation, if gpuUtilizationPercent is consistently low during the heavy phase of a run, you’re likely bottlenecked on data loading or CPU pre-processing
  • Diagnosing OOMs, gpuMemoryUtilizationPercent climbing to 100% just before a crash points to a memory issue rather than a code bug
  • Power and thermal investigations, gpuPowerWatt against gpuPowerLimitWatt and gpuTemperatureCelsius show whether the GPU is throttling
  • Cost attribution, combined with the run’s wall-clock time, the metrics let you compute cost per GPU-hour-of-actual-work