The verified-dse programme asks when a cheap fidelity rung (generic gates, liberty mapping, placement) can stand in for a routed SKY130 run when searching over ABC pass sequences. On ten small designs the answer was "not at equal wall-clock": a routed point cost 1.3 s against 0.15 s for the cheapest rung, so there was little to save.
This run adds five larger designs (16, 24 and 32-bit multipliers, 64-bit adders), the same 60 pass sequences and two clock targets: 600 runs with Yosys 0.33 and OpenROAD 26Q2 (the nixpkgs binary cache). A routed point now costs 19.4 s against 0.77 s, 25 times the cheapest rung.
What came out, replayed at equal wall-clock on the cached table:
- At the tight clock, successive halving from the mapped rung reaches 1.34% mean routed regret at 100 s against 4.44% for random routed search, and 0.33% against 1.85% at 200 s.
- Random search overtakes it from about 600 s, because halving stops at a 0.30% floor of rung bias. At the relaxed clock it wins only at 100 s.
- Halving from the generic rung loses everywhere.
- Two of my three predictions were wrong in detail: the generic rung is weakly informative on multipliers, and mapped rungs mis-rank the big multipliers at the relaxed clock, the opposite of the small-design result.
Five designs, pools of 50 and one corner, so this is a first sign, not a result. The log with the predictions made before the run is in LOG.md.