A certified chip-flow search is not a zero-regret one

On 20 SKY130 design-clock cells, a wider three-rung screening schedule loses 1.18% routed area-delay, against 2.27% for two rungs and 5.38% for random.

In verified-dse we ask when cheap pre-routing scores can be trusted to shortlist designs for place-and-route. Earlier milestones showed a three-rung schedule (keep 35 of 50, then 16, then route 8) certifies 18 of 20 design-clock cells and costs about 9% less wall-clock than a two-rung ladder (keep 16, route 5).

Milestone 105 measures what that costs in quality, replaying both schedules on the cached 1,000-run table with 200 random tie-breaks per cell. Mean routed regret over cells: 1.18% (three-rung), 2.27% (two-rung), 5.38% and 7.86% for random routing of 8 and 5 points. The three-rung schedule is at least as good in 18 of 20 cells.

The caveat: a certificate says a safe point exists, not that the optimum is found. The two-rung ladder misses the routed optimum in 5 cells, the three-rung one in 3, and on barrel16 both do worse than routing 8 random points (6.8% and 11.8% against 2.6% and 5.7%). The choice of schedules was made on this same table, so the comparison is in-sample.

Files: research/verified-dse/experiments/theory/pilot_cert_sched_regret.py, its results file and LOG.md. No flow was run; detail-routed mult24/mult32 cells remain the next experiment.

All writing · More in Research findings · More on verified-dse · All tools and research