GPU Job Scheduling in DRM - Past, Present and Future

Explore the evolution and future of GPU job scheduling in the Linux kernel's DRM subsystem through this 19-minute conference talk from XDC 2025. Discover how the DRM GPU Scheduler, which serves as the shared infrastructure for load balancing GPU job submissions across all GPU drivers, has accumulated significant problems over its decade-long development. Learn about the historical context that led to these issues and examine the solutions implemented in the past 12 months, while understanding which challenges still remain unresolved. Gain insights into proposed approaches for avoiding the most critical categories of problems that have plagued the scheduler, presented as a request for comments to the kernel development community. Understand the important shift toward firmware scheduling support, where GPU firmware handles actual scheduling instead of the kernel, transforming the DRM GPU Scheduler from an active scheduler into a load balancer. Examine the introduction of the DRM Jobqueue, a new simplified piece of DRM infrastructure designed as a more robust replacement for the current GPU Scheduler, with initial implementation planned for upcoming drivers like Nova and potential future adoption by existing GPU Scheduler users.