AMD Open-Sources Spur GPU Scheduler for HPC

AMD open-sourced a new GPU job scheduler called Spur on Apache 2.0, built as part of its Open Ecosystem initiative and written in Rust. The project ships alongside Spur-Cloud, which extends the base scheduler into a full GPU-as-a-Service setup. AMD's pitch is straightforward: Slurm was built for CPU-era HPC clusters and bolted on GPU support later, while Kubernetes offers strong orchestration but needs a pile of extra operators and queueing layers before it handles batch GPU jobs the way Slurm shops expect. Spur tries to close that gap directly, with a Slurm-compatible CLI and APIs, topology-aware scheduling for interconnects like XGMI, built-in high availability through Raft consensus, native Kubernetes integration, and vendor-agnostic device handling via the Container Device Interface.

The stakes are concrete. A large language model training run might need 64 MI300X GPUs spread across 8 nodes with exact locality requirements. A vision pipeline might want to split one 8-GPU node across 8 separate jobs.

Inference workloads need fast placement and teardown on hardware that isn't uniform. That's the terrain AMD says Spur was built to handle.