How Kubernetes Sees a GPU

Why this matters

Kubernetes was built to schedule CPU and memory, which it understands natively. A GPU is none of those things — the kubelet has no idea a GPU exists until something tells it. That something is the device plugin framework, which lets vendors advertise hardware as a named extended resource like nvidia.com/gpu. Until you grasp that a GPU is just a counted, opaque resource the scheduler matches against pod requests, every GPU-scheduling decision later in this course will feel arbitrary. This task plants the single mental model the rest of the course extends: a GPU is an extended resource a node advertises and a pod requests.

Demo

Inspect a GPU node and you'll see nvidia.com/gpu listed under Capacity and Allocatable, right next to cpu and memory. That line is the device plugin doing its job — the node is telling the scheduler how many GPUs it has to hand out.

Try it yourself

1. Run kubectl describe node and confirm nvidia.com/gpu appears under both Capacity and Allocatable.
2. Compare Capacity vs Allocatable for nvidia.com/gpu — if they differ, a system pod is already holding a GPU.
3. On a node with no device plugin installed, confirm nvidia.com/gpu is absent entirely (not zero — missing).
4. List all nodes with the custom-columns command and verify the per-node GPU counts match your hardware.

References

• Kubernetes — Schedule GPUs
• Kubernetes — Device plugins