The verified-dse programme asks when a cheap estimate of a chip design's area-times-delay can safely steer a search that ends in an expensive routed run. Three recent steps, all on a cached table of real SKY130 flow runs (ten small designs, about 50 pass sequences each):
- A certificate is not the optimum. A schedule that certifies a good point can still miss the routed optimum. The wide three-rung schedule has mean regret 1.18% against 2.27% for the two-rung ladder; random routing of the same count is 5 to 8%. One design (a barrel shifter) is worse than random, because its cheap scores point the wrong way.
- Pick the schedule without seeing the design. Leave-one-design-out selection chose the same schedule in all ten folds: skip the generic-gate rung, score everything at the mapped rung, keep 8, route 8. Held-out regret 0.95% against 1.17% for the hand-chosen one.
- It survives smaller pools, barely. Against the hand-chosen schedule the edge is within noise, and better in only 5 to 6 of 10 designs.
One prediction was wrong: regret falls as the pool shrinks, because routing 8 of 20 points covers more of the pool. Limits: small designs, one library corner, subsamples of the same pools. The next test is detail-routed 24 and 32-bit multipliers, where the routed rung costs more. Code, tests and the full log are in the project directory, MIT licensed.