About

Overview

NVIDIA Data Center GPU Manager (DCGM) is a suite of libraries and tools for managing and observing NVIDIA data-center accelerators. It runs on each host and provides a common entity, telemetry, health, configuration, policy, diagnostic, and accounting model for individual systems and larger cluster- management integrations.

DCGM can run as a shared library embedded in another management agent or in a standalone host-engine process used by command-line and remote clients. DCGM’s full feature surface is intended to be available through the C/C++ APIs, the installed Python bindings, NVIDIA’s separately distributed Go bindings, and the dcgmi command-line tool. Individual language-binding releases can temporarily lag the canonical C API.

Focus Areas

Online health and diagnostics

DCGM supports passive background health monitoring while workloads run, short readiness checks before work is admitted, and longer active diagnostics for investigating suspected hardware or integration problems. These online checks operate in the installed host environment; they do not replace NVIDIA offline field diagnostics or the hardware RMA process.

Job statistics and continuous telemetry

Schedulers and resource managers can collect low-overhead telemetry and associate execution, utilization, energy, error, and health information with a process or scheduler job. Raw fields also support continuous monitoring and integration with a site’s observability system.

Collections of related resources

DCGM groups let clients apply a common operation to the GPUs or other entities that make up a node, partition, tenant, or job. Groups can overlap and can be short-lived or node-wide. Isolation and allocation remain responsibilities of the operating system, scheduler, or virtualization platform.

Fabric and adapter observability

On supported NVSwitch systems, DCGM discovers switches and ports and exposes link state, traffic, error, temperature, and health information. Fabric Manager is a separate component responsible for configuring and managing the fabric itself. Current DCGM releases can also expose health, PCIe state, errors, and temperature for supported ConnectX adapters.

GPU configuration state

DCGM can apply and retain target settings such as compute mode, application clocks, power limits, ECC mode, and workload power profiles where the GPU supports them. This lets a node or job establish a known runtime configuration and re-enforce it after selected state changes.

Management policies

DCGM policies watch a defined set of GPU error or threshold conditions and can notify a client or request a recovery action and follow-up validation. Policy is intended for the conditions built into DCGM; broader site automation normally consumes DCGM telemetry and results from an external controller.

Target Users

DCGM serves several kinds of users:

• datacenter and cluster administrators operating GPU infrastructure;

• scheduler, monitoring, and platform developers integrating GPU management into a larger system;

• OEM and ISV developers validating and supporting GPU-enabled platforms; and

• users and support engineers investigating accelerator behavior or workload failures.

The command-line tools provide direct operational access and reproducible troubleshooting commands. The C/C++ APIs and language bindings are intended for long-running integrations and workflows that need structured results, callbacks, or finer control.

Continue with Install DCGM to set up DCGM, Learn for concept and workflow introductions, or Reference for exhaustive interface details.