Same eight GPUs, thirty-three points more

A team has published measurements from swapping a first-in-first-out GPU scheduler for a constraint-aware allocator. Nothing about the hardware or the jobs changed — only the order in which allocation decisions are made.

On a training-heavy workload across 8 GPUs, utilisation rose from 53.6% to 87.0%, a gain of 33 points, and priority-weighted output more than doubled at +105.1%. A mixed control scenario on the same 8 GPUs moved from 51.6% to 72.4%, with value up 54.8%.

Where it stops helping

The 64-GPU scale test is the honest row in the table: utilisation did not move at all, staying at 44.9%, while value rose 15.9%. Whatever the allocator fixes at eight GPUs is not the binding constraint at sixty-four, and the write-up publishes that alongside the flattering numbers rather than instead of them.

Decision latency stays small — 1 to 2 milliseconds in the smaller scenarios, 15 milliseconds at 64 GPUs with 30 jobs — so the scheduler is not itself the bottleneck at these sizes.

The mechanism is worth stating plainly, because it applies well beyond one cluster. Instead of reserving capacity all day against peak demand, the allocator treats real-time traffic as a curve rather than a ceiling, matching allocation to actual demand at each timestep and fitting batch work around it in priority order.

For anyone renting compute rather than owning it, the same arithmetic runs backwards: a provider quoting a price per GPU-hour on a cluster running near 50% utilisation is selling the idle half to somebody.