// the find
NVIDIA/gpu-operator
NVIDIA GPU Operator creates, configures, and manages GPUs in Kubernetes
The GPU Operator is NVIDIA's official Kubernetes operator for managing the full GPU software stack — drivers, container runtime, device plugin, DCGM monitoring — on GPU nodes. It's aimed at platform teams running Kubernetes clusters where GPU nodes need to scale dynamically without custom OS images. This is the blessed path if you're running NVIDIA hardware on Kubernetes at any serious scale.
The containerized driver model is genuinely useful: you can update drivers by restarting a container rather than reimaging nodes, which matters when you have a mix of driver version requirements. DCGM integration for GPU metrics is first-class — Prometheus scraping via ServiceMonitor is wired up out of the box, not an afterthought. The state machine architecture (each component gets its own state directory with numbered YAML manifests) makes it predictable to debug — you can see exactly what DaemonSets and RBAC rules the operator will apply. MIG manager support means multi-instance GPU partitioning is handled the same way as everything else, not a separate bolt-on.
The CRD surface area is large and the ClusterPolicy type has grown into a catch-all configuration blob — expect to spend time with the docs before you understand what half the fields do. The NVIDIADriver CRD is still in v1alpha1 despite being a central piece of the stack, which means the API can change under you. Community contributions have to go through both GitHub and GitLab CI (the pipeline badge points to GitLab), which is confusing and suggests internal development doesn't fully happen in the open. No AMD or Intel GPU support — this is NVIDIA-only, so if you're running a mixed-vendor cluster you're managing two separate operator lifecycles.