GPUs on Kubernetes: drivers, device plugin, sharing and health
How a GPU becomes schedulable: drivers, the container toolkit, the NVIDIA device plugin and GPU Operator, extended resources, node labels and taints, whole-GPU versus MIG versus time-slicing, GPU memory pitfalls, health and XID errors, and a verification checklist.
17 min·Verified 2026-11Pro+50 XP
A GPU is not a Kubernetes resource until some software on the node makes it one. Between "a machine with an H100 in it" and "a pod that can call CUDA" sit a driver, a container runtime hook, a device plugin and a set of labels. Knowing this stack turns mysterious Pending pods and "no CUDA-capable device" errors into five-minute fixes.
Pro content
Keep reading with Inference Academy Pro
Unlock the rest of this lesson, plus every advanced lesson, hands-on lab, the larger playground models and completion certificates.