From pod.yaml to /dev/nvidia0: How Kubernetes Actually Allocates GPUs to Containers

What happens between applying a Kubernetes pod spec requesting a GPU and actually getting hardware access in your container? This technical deep-dive traces the complete GPU allocation journey from YAML to running hardware. Akhil Mohan, Broadcom software engineer and containerd maintainer, walks through every layer of the stack: API server scheduling decisions, device plugins, Dynamic Resource Allocation (DRA), Container Device Interface (CDI), and the CRI handoff to container runtimes. Learn how nvidia.com/gpu:1 in your resource limits becomes /dev/nvidia0 in your running container, including GPU Feature Discovery, the evolution from opaque integer resources to attribute-based scheduling with CEL expressions, and how DRA drivers hand CDI specifications to the kubelet. Presented at Cloud Native Summit Kerala 2026 by an active contributor to Kubernetes SIG Node and the container runtime ecosystem. Covers device plugin architecture, resource slicing, node labels, and the transition from legacy GPU allocation to modern DRA with custom resources. Links: https://cnskerala.in #Kubernetes #GPU #ContainerRuntime #DevicePlugins #DRA #CDI #CloudNative