What it does
NVIDIA DCGM Exporter uses the Data Center GPU Manager stack to expose GPU telemetry in Prometheus format. It can run as a package-managed service, a container, or a Kubernetes DaemonSet, and publishes selected fields such as utilization, memory, and other device metrics. It is useful for GPU infrastructure monitoring but requires supported NVIDIA hardware and compatible driver and DCGM versions.
Start with a supported GPU host
DCGM Exporter requires an NVIDIA GPU and a driver/DCGM combination supported by the installed exporter release. The official guide describes package-managed, container, and Kubernetes DaemonSet deployments. For a one-host evaluation, the container example publishes the exporter on port 9400; Kubernetes users can deploy it with the documented Helm chart or NVIDIA GPU Operator. These setup paths are examples from the documentation, not a deployed service.
Scrape the metrics
The exporter serves a Prometheus metrics endpoint at /metrics on port 9400 by default. Add the reachable GPU host to a Prometheus scrape job, then use the metrics reference to select fields supported by the GPU and runtime. The output can help track device utilization, memory, temperature, power, and profiling data where hardware and configuration allow.
DCGM Exporter focuses on NVIDIA GPU telemetry, not general host monitoring. Its operating requirements include compatible NVIDIA software and access to the GPUs; some profiling fields need additional capability settings. Prometheus or another compatible scraper and a separate backend are still required for retention, queries, dashboards, and alerting. The repository is Apache-2.0 licensed.
Strengths
- Exposes DCGM GPU fields as Prometheus metrics for device-level monitoring.
- Supports package, container, and Kubernetes DaemonSet deployment paths.
- The current documentation describes scraping `/metrics` on port 9400 and provides a reference for shipped metric fields.
- The official repository is licensed under Apache License 2.0.
Watch out for
- Requires supported NVIDIA GPU hardware and a compatible driver/DCGM/exporter combination.
- Some profiling metrics need additional privileges, and the available fields depend on GPU and runtime support.
- It exports metrics but does not provide time-series storage, dashboards, or alert delivery itself.
Best for
- Platform teams monitoring GPU utilization and device health on NVIDIA compute nodes.
- Kubernetes operators adding NVIDIA GPU telemetry to an existing Prometheus stack.
- ML infrastructure teams checking GPU workloads with a metrics backend rather than relying only on `nvidia-smi` snapshots.
Sources
- NVIDIA/dcgm-exporter (Oct 26, 2026)
- Install DCGM Exporter - NVIDIA DCGM Documentation (Oct 26, 2026)
- Configure Prometheus for DCGM Exporter - NVIDIA DCGM Documentation (Oct 26, 2026)
- DCGM Exporter Metrics - NVIDIA DCGM Documentation (Oct 26, 2026)
- LICENSE (Oct 26, 2026)