Why it matters

Ops-scale GPU management needs proper tooling. DCGM is the standard.

Advertisement

The architecture

dcgm daemon on each host collects metrics.

Exporter feeds Prometheus.

Diagnostics run health checks.

DCGM stackdcgm daemonper hostMetrics + exporterPrometheusDiagnosticshealth + faultsGPU utilization ≠ MFU (Model FLOPs Utilization); often lower than headline number
DCGM components.
Advertisement

How it works end to end

Metrics: SM utilization, memory bandwidth, tensor core utilization, thermals, power draw, ECC errors.

Health checks: GPU sanity tests, memory checks, PCIe link health.

Integration: NVIDIA GPU Operator for Kubernetes uses DCGM.