Requesting a GPU in a Pod Spec

Why this matters

All the infrastructure exists so that a workload can ask for a GPU, and the way it asks is a single line under resources. But GPUs behave differently from CPU and memory in one critical way: they are not compressible or fractional by default — you request whole GPUs, and the scheduler hands you exclusive use of that device. Getting the request syntax right, and understanding that a request of 1 means one entire physical GPU, is the difference between a pod that schedules and one that sits Pending or, worse, silently runs on CPU. This is the most-used five lines of YAML in the entire course.

Demo

A GPU request goes under resources.limits with the key nvidia.com/gpu. The pod below asks for one GPU and runs a CUDA workload; the scheduler will only place it on a node with a free GPU.

Try it yourself

1. Apply the pod and confirm its logs print the CUDA sample's success line.
2. Set nvidia.com/gpu to a number larger than any node has and confirm the pod stays Pending with an 'Insufficient nvidia.com/gpu' event.
3. Remove the GPU request entirely and confirm the same image runs without GPU access (it will fail or fall back to CPU).
4. Schedule two single-GPU pods on a 1-GPU node and confirm the second stays Pending — GPUs are exclusive.

References

• Kubernetes — Schedule GPUs
• NVIDIA — k8s-device-plugin examples