What Does a Local RL Controller Add to a System-Level Prefetch Coordinator?

Athena coordinates a data prefetcher and off-chip predictor at the system level using global signals and Q-value-driven external aggressiveness control. We study when an independent local RL controller has a useful performance lever inside the selected prefetcher. We adapt sBerti, a DPC4 2026 prefetcher, to Athena's L2C interface, add a direct cache-side prefetch-quality term to Athena's reward, and embed a local SARSA degree controller inside sBerti. In a matched final-build milc control, the reward term alone changes IPC by only +0.06% with overlapping three-run ranges. The local controller, however, improves IPC by +1.9% on milc and +5.8% on leslie3d at the primary K=5 setting; leslie3d reaches +8.3% at K=20 in sensitivity analysis. Three other workloads improve local prefetch accuracy without a robust IPC gain for distinct reasons. The central lesson is that improving a component-level quality metric does not guarantee system-level improvement; the active bottleneck determines whether the local controller has a useful lever.