Earlier today we wrote that skipping the cheapest estimate was the schedule our verified-dse programme picks when it chooses without peeking, and that its edge was "small and within noise". All of that came from one cached table of about 50 ABC pass sequences per design. So we ran new data.
What we ran. 120 ABC pass sequences that the original table never used, on the same ten small designs and two clock targets, through four rungs: generic-gate synthesis, SKY130 liberty mapping, placement, and global route with timing repair (Yosys 0.70, OpenROAD 26Q3). That is 2,400 flow runs, about 14 minutes on 4 cores. 17 of the sequences are rejected by ABC on every design (a resynthesis pass after dch), leaving 103 or 104 points per cell. The predictions were written into the analysis script before the flow ran.
What held. All five predictions. Routed area-times-delay regret, averaged over 20 design-clock cells:
| schedule | mean regret |
|---|---|
| skip the generic-gate rung, score all at the mapped rung, keep 8, route 8 | 1.04% |
| hand-chosen three-rung schedule (keep 70%, then 16, then 8) | 5.49% |
| route 8 random points | 8.00% |
Why, and the caveat. Almost the whole gap is one design. The 16-bit ripple adder has only four distinct generic-gate scores, and 87 of its 104 points share the lowest one. The best routed design sits in the second group. Keeping the best 70% by the cheap score keeps 73 points, all from the wrong group, so the search cannot find the optimum: 34 to 40% regret, worse than routing random points. Leave out that adder and one barrel-shifter cell, and the two schedules are equal (1.30% against 1.29%); cell by cell it is 7 wins, 9 ties and 4 losses.
So the practical advice is narrower than "skip the cheap rung, it is better". It is: a cheap score with many ties can silently cut the optimum, and skipping it costs nothing on average. If you run successive halving over an open chip flow, check how many distinct values your first rung produces before you trust a cut on it.
Limits: the same ten small designs (new sequences, not new designs), a tool-version change from the first table, and the cost of scoring every point at the mapped rung is not yet re-priced. Code, raw results and tests are in the project directory, MIT licensed.