Open Source Summit + Embedded Linux Conference North America 2026

Kubernetes Dynamic Resource Allocation is now being adopted by hardware vendors implementing drivers for accelerators and high-speed networking. But as real-world AI workloads hit the spec—disaggregated inference across multiple nodes, topology-aware co-location of compute and network devices, per-workload hardware configuration—gaps are emerging. This Birds of a Feather session invites platform engineers, infrastructure teams, and Kubernetes contributors to compare notes on where DRA works well and where the spec leaves teams filling in the blanks themselves. We will use the AWS Trainium Neuron DRA driver as a concrete reference point to ground the discussion, then open the floor: What attributes should be standardized across vendors? How should multi-device atomic allocation be handled? What does the community need to build so that DRA delivers on its promise across all accelerator hardware?