Same Cluster, 33 Points More Utilization: What Changed Was the Order

A constraint-aware GPU allocator was benchmarked against a FIFO scheduler across seven scenarios with identical hardware and workloads. On the same cluster, GPU utilization rose by as much as 33 percentage points and priority-weighted output increased by as much as 105% simply by changing the order in which allocation decisions were made rather than using arrival-order placement. The core problem is that different workload types, including training, real-time inference, batch inference, and quantization, have incompatible resource requirements and compete for the same GPUs in the same timesteps. FIFO scheduling wastes capacity through fixed reservations held for peak demand periods and by committing resources to lower-priority jobs that arrived first, whereas the allocator treats real-time demand as a curve allocated at each timestep and places batch-like jobs by priority across the entire horizon.