NVIDIA DCGM

NVIDIA DCGM monitors and manages NVIDIA data-center GPUs for infrastructure teams running GPU clusters. It helps teams investigate slow jobs, detect hardware problems and control GPU power and clock settings.

• Health checks and diagnostics: Low-overhead monitoring checks GPU health while jobs run, while active and passive diagnostics help identify failures, performance degradation and power inefficiencies.
• Performance telemetry: GPU performance and utilization measurements help teams understand job behavior, find underused resources and investigate application performance problems.
• Power and clock policies: Administrators can set policies that govern GPU power use and clock settings.
• Kubernetes monitoring: DCGM-Exporter provides GPU metrics and health data for monitoring and alerting in Kubernetes, with integrations including Prometheus, Bright Cluster Manager and IBM Spectrum LSF.

DCGM runs on your own Linux servers with NVIDIA data-center GPUs and supports x86_64 and aarch64 (SBSA) systems. Teams can use it independently or connect it to cluster management and resource scheduling software. C and Python APIs, plus Go bindings, support custom management applications.

DCGM is open core. Its foundational libraries and building blocks are open source under Apache 2.0; some diagnostics and tests remain proprietary.