The verified-dse programme searches over synthesis pass sequences and uses cheap stages (global route plus timing repair, "R4") to rank candidates instead of fully routing each one. The question for a larger design: does that cheap ranking agree with the expensive one (detail routing plus parasitic extraction)?
We took 50 variants of a 24-bit multiplier at a relaxed clock, picked 12 evenly spaced from best to worst by the cheap score, and detail-routed all 12 (Yosys 0.51, OpenROAD 2025-03, SKY130).
- Kendall tau between the cheap score and the detail-routed score: 0.73 (placement-only scores: 0.52).
- The best detail-routed point was the cheap score's best point. Picking by the cheap score lost 0.0%.
- Detail routing took a median 20 minutes per point and left no DRC violations.
Twelve hand-spread points give no confidence interval, and this is one design, one clock and one tool build. It supports using the cheap score as the reference for these cells; it does not settle a 32-bit multiplier.
Along the way, three launches produced nothing because a newer OpenROAD prints areas as u^2 and slack without the word max. The parsers now accept both.
Log and data: research/verified-dse/LOG.md, milestone 79.