Factorial attribution of prefill and decode energy savings

Determine the separate efficiency contributions of the prefill clock-window floor, the prefill in-window runtime policy, and the decode power cap by conducting a factorial per-lane ablation against the uncontrolled baseline on both workloads, including a floor-only prefill condition with the in-window controller disabled.

Background

The reported controller combines multiple mechanisms: a prefill clock-window floor, a proprietary in-window prefill controller, and a decode power cap. Because the in-window controller was active in every experimental arm, the measured prefill savings cannot be attributed to the window floor alone. The paper also did not run the planned factorial study separating prefill-only and decode-only effects, so the individual mechanism contributions remain unresolved.