The verified-dse programme asks how much of an open-source chip design search (Yosys, ABC and OpenROAD on SKY130) can be skipped by using cheap estimates. Its working rule, the clipped ladder (SWC), maps every ABC pass sequence, ranks the mapped netlists by a cheap score, and routes only the best 8. The baseline is routing 8 sequences picked at random, which costs the same.
The last run brought the blind evidence to 73 designs: the ladder wins 55 and loses 15. Every one of those designs was placed at 40% utilisation, the operating point where the rule's two settings were chosen. A practitioner packing a block tighter would want to know whether the result was a property of the rule or of that one operating point.
What this run did
It reran the 52 one-clock blind designs at 65% placement utilisation, with the same clocks and pass sequences and nothing retuned. A seeded run order, seven predictions and the rule for a time-limited stop were committed before any point ran. The flow stopped at its two-hour limit with 20 designs complete (2,468 routed points), so those 20 are the analysed set.
What came out
All seven predictions held.
| operating point | SWC mean regret | random | SWC better / worse | sign p |
|---|---|---|---|---|
| 65% utilisation | 2.52% | 7.13% | 16 / 4 | 0.012 |
| 40%, same 20 designs | 2.57% | 6.70% | 17 / 3 | 0.003 |
The new operating point is a real change, not a copy of the old one. Routing costs more relative to mapping (the median ratio goes from 5.3 to 7.6), the routed ranking of the same netlists shifts (median Kendall tau-b 0.74 between the two tables), and the best sequence changes on 8 of the 20 designs. Even so, the ladder's edge over random routing moved by -0.47 points (95% interval -1.82 to +1.06), and its sign agrees on 17 of the 20 designs.
The losses are the three 256-bit restoring dividers. That family also lost the most at 40%: a chain of compares and subtracts is still where the cheap score misleads.
What went wrong, and what it does not show
One 256-bit design lost 82 of its 120 points at 65% to a pin placement error: the smaller die no longer had enough perimeter for its I/O pins. The points that survived are the larger netlists, so that design's cell is a selected pool. We found this after the run and report it separately. Without that design the result is 15 to 4 (p = 0.019).
This is 20 of 52 planned designs, one more utilisation rather than a second library corner, one tool version, and global-route estimates rather than detailed routing.
Code, rows and the full table: experiments/util65_blind/.