Install DCGM#

Confirm that the target platform is supported, satisfy the applicable requirements, configure the NVIDIA package repository, and then use the package procedure for the installed Linux distribution.

Supported Platforms#

Hardware and Environments#

DCGM supports the following products and environments:

All NVIDIA Kepler™ (K80) and newer NVIDIA datacenter (previously, NVIDIA Tesla®) GPUs
NVIDIA® NVSwitch™ on NVIDIA DGX™ A100, NVIDIA HGX™ A100, and newer systems
All NVIDIA Maxwell™ and newer non-datacenter GPUs, including NVIDIA® GeForce® and NVIDIA® Quadro® GPUs
CUDA® 7.5 and later and NVIDIA Driver R450 and later
Bare-metal systems and virtual machines with full GPU passthrough

On non-datacenter GPUs, some DCGM functionality is unavailable. The following table summarizes category-level availability. For the module organization and feature-specific documentation, see Functionality by module.

Feature Availability by GPU Class#

Feature availability by GPU class#
Feature area	Datacenter GPUs	Non-datacenter GPUs	Qualification
Field watches and GPU telemetry	Available	Available	Individual fields remain hardware- and driver-dependent.
Configuration management	Available	Available	Individual settings require corresponding GPU and driver support.
Passive health monitoring	Available	Available	Monitored subsystems depend on available fields and hardware.
Process and job statistics	Available	Available	Particular statistics can be unavailable when their source fields are unsupported.
Topology and link state	Available	Available	Reported PCIe, NVLink, and fabric details depend on system topology.
Host-engine introspection	Available	Available	Describes host-engine resource use rather than a GPU capability.
Policy notifications	Available	Not supported	Policy registration and notification are restricted to datacenter GPUs.
Single-node diagnostics	All levels supported for the applicable product	Level 1; higher levels only where explicitly supported	Plugin, SKU, driver, privilege, and installation requirements still apply.

Available means that the feature area is exposed for the GPU class. It does not guarantee that every field, setting, or diagnostic is supported on every GPU.

Linux Distributions#

Linux Distributions and Architectures#
Linux Distribution	x86_64	aarch64
Amazon Linux 2023	✓	✓
Azure Linux 3.0	✓	✓
Debian 12	✓	✓
Debian 13	✓	✓
Fedora 43	✓	✓
RHEL 8.y/Rocky Linux 8.y	✓	✓
RHEL 9.y/Rocky Linux 9.y	✓	✓
RHEL 10.y/Rocky Linux 10.y	✓	✓
SLES/OpenSUSE 15.y	✓	✓
SLES/OpenSUSE 16.y	✓	✓
Ubuntu 22.04 LTS	✓	✓
Ubuntu 24.04 LTS	✓	✓
Ubuntu 26.04 LTS	✓	✓

Requirements#

The system-resource and NVIDIA software requirements apply when DCGM is run. They are not prerequisites for installing the development package.

System Resources#

These requirements apply to bare-metal hosts, virtual machines, and containers.

Resource	Requirement
Minimum System Memory (Host RAM)	>= 16GB
Minimum CPU Cores	>= Number of GPUs

NVIDIA Software#

Install a supported NVIDIA Datacenter Driver. DCGM is tested with NVIDIA Datacenter Drivers; other driver types can omit functionality. The NVIDIA driver documentation describes the available branches and support timelines.

On systems with NVSwitch™ hardware, also install:

Fabric Manager
NVSwitch™ Configuration & Query (NSCQ) for Hopper and earlier GPUs
NVSwitch™ Device Monitoring (NVSDM) for Blackwell and later GPUs

Fabric Manager, NSCQ, and NVSDM package major versions must match the NVIDIA driver major version. For example, driver version 565 uses libnvidia-nscq-565 and libnvsdm-565. Use nvidia-smi to identify the installed driver version.

See the Fabric Manager installation instructions, the NSCQ instructions in the HGX Software Guide, and the NVSDM installation instructions.

Package Repository and Privileges#

Configure the system package manager to use the NVIDIA repository for the installed Linux distribution. If the system uses a local CUDA repository package, update that package before installing DCGM. See the CUDA Installation Guide for Linux for repository setup instructions.

Package installation and system-service changes require root privileges. The examples use sudo to elevate only the command that requires them.

Commands are written for a Bash-compatible shell. A leading $ is the prompt, not part of the command; backslashes continue a command on the next line.

Package Installation#

Package Relationships#

DCGM packages#
Package	Contents and relationships
`datacenter-gpu-manager-4-core`	CUDA-independent runtime components. Both the CUDA runtime package and the development package require the same version of this package.
`datacenter-gpu-manager-4-cuda<major>`	CUDA-major-specific runtime components. This package requires `datacenter-gpu-manager-4-core`.
`datacenter-gpu-manager-4-proprietary*`	The subset of DCGM functionality that is not included in the DCGM open-source project. The relevant packages are installed as recommendations.
`datacenter-gpu-manager-4-multinode-cuda<major>`	Optional multi-node diagnostic components, available for CUDA 12 and later.
`datacenter-gpu-manager-4-dev` or `datacenter-gpu-manager-4-devel`	Headers, linker artifacts, CMake files, and samples. This package requires `datacenter-gpu-manager-4-core`; it does not require a CUDA-major-specific runtime package.

The runtime-install commands below enable package recommendations. In addition to the proprietary package family, the recommendation set contains non-DCGM packages such as the system NUMA library. The alternatives shown below omit all recommendations.

Select a CUDA Runtime Package#

DCGM’s CUDA-dependent binaries are divided into packages by CUDA user-mode driver major version. The package suffix identifies the target version; for example, datacenter-gpu-manager-4-cuda12 targets CUDA 12.

Use nvidia-smi to identify the CUDA version supported by the installed driver:

$ nvidia-smi -q | grep -E 'Driver Version|CUDA Version'
Driver Version                            : 575.57.08
CUDA Version                              : 12.9

Set CUDA_VERSION to the major version shown in the output. For the example above:

$ CUDA_VERSION=12

Warning

On Maxwell, Pascal, and Volta systems using driver version 580, set CUDA_VERSION=12. Although that driver supports the CUDA 13.0 user-mode API, CUDA Toolkit 13.0 does not support those GPU generations.

Stop an Existing DCGM Service#

Stop DCGM before replacing an existing installation:

$ sudo systemctl list-unit-files nvidia-dcgm.service > /dev/null && \
  sudo systemctl stop nvidia-dcgm

Install Packages#

Use the section for the installed Linux distribution. Each procedure shows the CUDA-major runtime package, the development package, and the optional multi-node package. Install the packages needed on the system. After installing the runtime package, continue with Start and Verify DCGM.

Amazon Linux 2023#

Install the runtime package for the selected CUDA major version and its recommended packages.

$ CUDA_VERSION=<major version of CUDA user-mode driver>
$ sudo dnf install --assumeyes \
                   --setopt=install_weak_deps=True \
                   datacenter-gpu-manager-4-cuda${CUDA_VERSION}

To omit all weak dependencies, including recommended packages, replace --setopt=install_weak_deps=True with --setopt=install_weak_deps=False.

If needed, install the optional multi-node diagnostic plugin.

$ sudo dnf install --assumeyes datacenter-gpu-manager-4-multinode-cuda${CUDA_VERSION}

If needed, install the development files.

$ sudo dnf install --assumeyes datacenter-gpu-manager-4-dev

Azure Linux 3.0#

Install the runtime package for the selected CUDA major version and its recommended packages.

$ CUDA_VERSION=<major version of CUDA user-mode driver>
$ sudo tdnf install --assumeyes \
                    --setopt=install_weak_deps=True \
                    datacenter-gpu-manager-4-cuda${CUDA_VERSION}

To omit all weak dependencies, including recommended packages, replace --setopt=install_weak_deps=True with --setopt=install_weak_deps=False.

If needed, install the optional multi-node diagnostic plugin.

$ sudo tdnf install --assumeyes datacenter-gpu-manager-4-multinode-cuda${CUDA_VERSION}

If needed, install the development files.

$ sudo tdnf install --assumeyes datacenter-gpu-manager-4-dev

Ubuntu LTS and Debian#

Remove any installations of the datacenter-gpu-manager and datacenter-gpu-manager-config packages

$ sudo dpkg --list datacenter-gpu-manager &> /dev/null && \
  sudo apt purge --yes datacenter-gpu-manager

$ sudo dpkg --list datacenter-gpu-manager-config &> /dev/null && \
  sudo apt purge --yes datacenter-gpu-manager-config

Update the package registry cache
```
$ sudo apt-get update
```

Install the runtime package for the selected CUDA major version and its recommended packages.

$ CUDA_VERSION=<major version of CUDA user-mode driver>
$ sudo apt-get install --yes \
                       --install-recommends \
                       datacenter-gpu-manager-4-cuda${CUDA_VERSION}

To omit all recommended packages, replace --install-recommends with --no-install-recommends.

If needed, install the optional multi-node diagnostic plugin.

$ sudo apt install --yes datacenter-gpu-manager-4-multinode-cuda${CUDA_VERSION}

If needed, install the development files.

$ sudo apt install --yes datacenter-gpu-manager-4-dev

RHEL, Rocky Linux, and Fedora#

Remove any installations of the datacenter-gpu-manager and datacenter-gpu-manager-config packages.

$ sudo dnf list --installed datacenter-gpu-manager &> /dev/null && \
  sudo dnf remove --assumeyes datacenter-gpu-manager

$ sudo dnf list --installed datacenter-gpu-manager-config &> /dev/null && \
  sudo dnf remove --assumeyes datacenter-gpu-manager-config

Update the package registry cache.
```
$ sudo dnf clean expire-cache
```

Install the runtime package for the selected CUDA major version and its recommended packages.

$ CUDA_VERSION=<major version of CUDA user-mode driver>
$ sudo dnf install --assumeyes \
                   --setopt=install_weak_deps=True \
                   datacenter-gpu-manager-4-cuda${CUDA_VERSION}

To omit all weak dependencies, including recommended packages, replace --setopt=install_weak_deps=True with --setopt=install_weak_deps=False.

If needed, install the optional multi-node diagnostic plugin.

$ sudo dnf install --assumeyes datacenter-gpu-manager-4-multinode-cuda${CUDA_VERSION}

If needed, install the development files.

$ sudo dnf install --assumeyes datacenter-gpu-manager-4-devel

SUSE SLES / OpenSUSE#

Remove any installations of the datacenter-gpu-manager and datacenter-gpu-manager-config packages

$ sudo zypper search --installed-only --match-exact datacenter-gpu-manager &> /dev/null && \
  sudo zypper --non-interactive remove datacenter-gpu-manager

$ sudo zypper search --installed-only --match-exact datacenter-gpu-manager-config && \
  sudo zypper --non-interactive remove datacenter-gpu-manager-config

Update the package registry cache
```
$ sudo zypper refresh
```

Install the runtime package for the selected CUDA major version and its recommended packages.

$ CUDA_VERSION=<major version of CUDA user-mode driver>
$ sudo zypper install --no-confirm \
                      --recommends \
                      datacenter-gpu-manager-4-cuda${CUDA_VERSION}

To omit all recommended packages, replace --recommends with --no-recommends.

If needed, install the optional multi-node diagnostic plugin.

$ sudo zypper install --no-confirm datacenter-gpu-manager-4-multinode-cuda${CUDA_VERSION}

If needed, install the development files.

$ sudo zypper install --no-confirm datacenter-gpu-manager-4-devel

Start and Verify DCGM#

Enable the packaged systemd service at boot and start it now:

$ sudo systemctl enable --now nvidia-dcgm
$ systemctl is-active nvidia-dcgm
active

Use dcgmi to confirm that the host engine discovered the supported GPUs and NVSwitches in the system:

$ dcgmi discovery -l
8 GPUs found (Active).
...
6 NvSwitches found.

For local service options, environment changes, full unit replacement, and the multi-node diagnostics systemd generator, see Customizing the systemd Service.

Continue with the learning overview to learn how the host engine represents and monitors the system.

Installed Components#

Runtime Components#

The runtime packages provide these principal components:

Runtime components#
Component	Installed item	Further information
DCGM shared library	`libdcgm.so.4`	Learn and the C/C++ API
DCGM modules	Loadable feature-module libraries	Modules and dcgmi modules
Diagnostic plugins	Single-node diagnostic plugin libraries and supporting executables	Diagnostics and Diagnostic Plugins
NVIDIA Host Engine	`nv-hostengine`	nv-hostengine
systemd service	`nvidia-dcgm.service`	Customizing the systemd Service
Command-line interface	`dcgmi`	dcgmi
Profiling test utility	`dcgmproftester11`, `dcgmproftester12`, or `dcgmproftester13`, according to the installed CUDA-major package	dcgmproftester
Python bindings	`/usr/share/datacenter-gpu-manager-4/bindings/python3/`	Installed Python modules and examples
Multi-node diagnostics	Components supplied by the optional `datacenter-gpu-manager-4-multinode-cuda<major>` package	Multi-Node Diagnostics

Development Components#

The development package adds the following files to the runtime components installed by its datacenter-gpu-manager-4-core dependency:

Development files#
Component	Installed item
Public C/C++ headers	`/usr/include/datacenter-gpu-manager-4/`
Shared-library linker name	The unversioned `libdcgm.so` symbolic link
CMake package files	`/usr/share/cmake/DCGM/`, including the `DCGM::dcgm` imported target
DCGM stub library	`libdcgm_stub.a`
SDK samples	`/usr/src/datacenter-gpu-manager-4/sdk_samples/`

For C/C++ linking behavior and the complete interface reference, see the C/C++ API.