Rules, not rankings: interpretable AI for processor microarchitecture design analysis

Abstract

Expert-driven microarchitecture analysis provides valuable insights but fails to scale to modern design spaces, while black-box AI offers scalability at the cost of interpretability. We present PTAH, a data-driven methodology that combines codified expert knowledge with an interpretable conditional-density model. PTAH derives human-readable rules that link processor mechanism configurations to the full IPC probability distribution. Clustering these rules yields design principles that highlight key interactions, trade-offs, and behavioral regimes. A case study on SPEC CPU workloads shows that PTAH uncovers design insights while preserving full interpretability.