Installation#

Choose the installation guide for the component that owns the task. Install DCGM to operate its host engine, command-line tools, diagnostics, or APIs. Install DCGM Exporter to expose selected DCGM telemetry to Prometheus.

Choose What to Install#

Installation guides#

Component

Provides

Use it when

Guide

DCGM

The DCGM host engine, dcgmi, libraries, diagnostics, field definitions, and development files.

A host needs DCGM management, health, diagnostics, telemetry, or API access.

Install DCGM

DCGM Exporter

A Prometheus /metrics endpoint backed by selected DCGM fields.

Prometheus needs GPU, NVSwitch, NVIDIA CPU, or workload-attributed telemetry from a host or Kubernetes cluster.

Install DCGM Exporter

How the Components Relate#

DCGM Exporter uses DCGM through libdcgm. By default, it initializes an embedded DCGM host engine inside the exporter process. It can instead connect to a separately managed DCGM host engine over a supported remote transport.

Exporter releases pair a tested DCGM version with an exporter version. The installation method determines who supplies and manages that DCGM dependency:

  • With packages, the system package manager updates DCGM and DCGM Exporter, and systemd manages their services. Keep the package versions paired during upgrades.

  • A standalone exporter container carries the DCGM libraries selected for its release. The container image tag owns the pairing and can use embedded or remote operation.

  • A Helm release manages the exporter DaemonSet and related Kubernetes resources. Its chart values select the exporter image and version pair.

  • NVIDIA GPU Operator manages DCGM, DCGM Exporter, and the surrounding GPU software stack as one Kubernetes lifecycle.

Follow the selected child guide instead of assuming that every exporter deployment requires a separately installed DCGM host engine.

Shared Prerequisites#

Both products require a supported NVIDIA GPU, driver, operating environment, and sufficient access to the GPU devices. Package installation and service changes require administrative privileges. Container deployments also require the NVIDIA Container Toolkit and permission to use the container runtime.

Use the DCGM product support matrix and the detailed requirements in the selected installation guide before deploying across a fleet.

Next Steps#