GPU low-utilization monitoring
Rolling-window observability
for dark, dim, and underfed GPUs
Signals in
  • NVML or DCGM exporter
  • Low-util policy time
  • Idle-state sampling
  • Power & energy telemetry
Rolling windows
  • Configurable short & long
  • Low-util percentage
  • Idle sampled % & entries
  • Power corroboration
Operator view
  • Long-window low-util KPI
  • Dark / dim / bright signal
  • Underfeeding or burstiness
  • Not a root-cause engine
At a glance
Standard monitoring can show a GPU as moderately busy while rolling windows reveal high low-util time, repeated idle entries, and dim power.
Practical observability for H100, H200, and similar datacenter GPUs — not omniscient truth.