AMD has released Spur, a modern GPU job scheduler designed to streamline the management of GPU clusters for high-performance computing (HPC) and artificial intelligence (AI) workloads. The tool is built with a focus on GPU-first scheduling, offering features such as topology-aware resource allocation and native Kubernetes integration. Spur is developed by AMD and is open-sourced under the Apache 2.0 license as part of the company’s Open Ecosystem initiative. It provides Slurm-compatible CLI and APIs while delivering features tailored for today’s AI infrastructure, including embedded high availability through Raft consensus and vendor-agnostic device management via the Container Device Interface (CDI).
Spur was created to overcome the limitations of traditional HPC schedulers, which were built in an era dominated by CPUs and lacked native GPU support. The new scheduler addresses the complexities of managing GPU clusters, such as scheduling large language model training jobs that may require 64 MI300X GPUs across 8 nodes, or computer vision pipelines that need fractional GPU allocation. It also handles inference workloads that demand rapid placement and teardown across heterogeneous hardware. The scheduler is designed to meet these needs with first-class GPU support, rather than as an afterthought. The tool is intended to simplify operational complexity while maintaining the batch job model that HPC users rely on.
Spur consists of modular components that can be deployed either standalone or within Kubernetes. These include the controller daemon (spurctld), agent daemon (spurd), REST API server (spurrestd), Kubernetes operator (spur-k8s), and a command-line interface (spur-cli). Cluster scheduling state—such as jobs, nodes, and partitions—is replicated through embedded Raft consensus in spurctld, eliminating the need for external databases. Spur-Cloud extends the platform into a complete GPU-as-a-Service solution by adding its own PostgreSQL database for billing and session metadata.
Source: amd
