NVIDIA Data Center GPU Manager

What NVIDIA Data Center GPU Manager is

NVIDIA Data Center GPU Manager (DCGM) provides tools to manage and monitor NVIDIA data center GPUs in cluster environments. It collects telemetry that can help teams understand job behavior, utilization, and application performance, and includes health checks, diagnostics, alerts, and power and clock policies. Its active monitoring is described as low-overhead and non-invasive while jobs run. DCGM can run as a HostEngine service or be embedded in a standalone application, and DCGM-Exporter exposes GPU telemetry in Kubernetes environments with out-of-the-box Prometheus integration. It also lists integrations with Bright Cluster Manager and IBM Spectrum LSF. The free, open-source foundational libraries and building blocks are available for Linux on x86_64 and aarch64 systems; some diagnostics and tests are proprietary. GPU support varies by class, with the guide listing Kepler and newer data-center GPUs and Maxwell and newer non-data-center GPUs. The guide calls for at least 16 GB of host memory and a CPU core per GPU. Its TCP protocol does not provide authentication, authorization, or encryption, so NVIDIA advises limiting access to trusted management networks.