nvidia_gpu_exporter
A Prometheus exporter for NVIDIA GPU metrics that reads from the nvidia-smi binary or the NVML driver library, running on Linux, Windows and macOS, on bare metal, Docker or Kubernetes.
nvidia_gpu_exporter collects, parses and exports NVIDIA GPU metrics for Prometheus. It primarily uses the nvidia-smi binary, which works on Windows as well as Linux without needing Docker, and can optionally read metrics directly from the NVML driver library instead of parsing nvidia-smi output. A demo backend serves realistic synthetic metrics on a machine with no GPU at all, for testing.
It auto-discovers whatever metric fields the installed nvidia-smi exposes, so it stays compatible as fields change, and can optionally report per-process GPU memory usage. It targets consumer and prosumer GPUs where datacenter tooling exposes little, small Kubernetes clusters, edge boxes and homelabs, virtualized or restricted setups such as vGPU guests or MIG slices, and mixed fleets of old and new cards; for datacenter cards already running the NVIDIA GPU Operator, the project suggests DCGM-exporter instead.
It requires no C bindings and can be configured to execute nvidia-smi remotely rather than on the monitored machine itself.