Directory Storage & Hardware Overhead Comparison

Per-entry storage breakdown — isometric comparison (all protocols normalized to same LLC capacity)

bits/entry 70 60 50 40 30 20 10 Dir₂B 23b V+D+B 2×10b TLH-Dir₄B 23b ctrl 4×5b TLH-Dir₃B 21b ctrl 3×6b ⭐ Dorado 24b V+D+B 2×6b UpperBound 66b Full ctrl 2.75× 1.36× speedup ~1.37× speedup Within 1% perf! RTL-Synthesized Hardware Overhead (Dorado vs Dir₂B) Area: +0.44% | Leakage: +2.3% | Dyn. Power (typ): +1.8% | Dyn. Power (worst): +4.2%
🔵 Pointer Width
Clustering cuts pointer from 10b (1024 IDs) → 5b (32 IDs). +1 type bit = 6b. More pointers per budget.
🟢 SetOverflow
2×6b per way + 12×6b shared across 12 ways = avg 3 ptrs/way. Only 2% of sets overflow → B bit.
🟠 2.75× Gap
Dorado (24b) uses 2.75× less storage than full bit vector (66b) while staying within 1% of its performance.