Ai2 replaces priority-based GPU scheduling with time budgets, fair-share allocation and time-slicing

Ai2 says its new scheduler gives research projects GPU-time budgets, uses fair-share scheduling to allocate access and time-slices long-running jobs. During a 30-day test period, it says teams received 98% of the GPU hours they were owed, while cluster occupancy held steady at 98%. Ai2 also says preemption disrupted some interactive sessions, prompting plans for restorable CPU-only sessions.