GPU scheduling in Kubernetes is mostly count-based: request one device, land on a node, and let driver logic handle the rest. That fails when partitioning, topology, memory, and capability differences decide whether a job runs well, poorly, or not at all.

This session explains how Dynamic Resource Allocation changes scheduling. Speakers show how DRA gives Kubernetes a richer picture of devices and workload needs, and why partial-device capacity, preferred fallbacks, and partitioned hardware matter for placement.

Using upstream Kubernetes and the open source NVIDIA DRA driver, the talk traces workload request to scheduler decision to kubelet preparation, and what remains hard. Attendees leave knowing when device-aware scheduling is worth the complexity and what Kubernetes can decide today that count-based scheduling could not.