Does skipping the cheapest rung survive a smaller candidate pool?

Subsampling cached SKY130 place-and-route pools to 20, 30 and 40 points: scoring all at rung two and routing 8 stays slightly ahead of a wider schedule.

The verified-dse programme searches chip-flow design spaces by scoring candidates on cheap rungs and routing only survivors. Last milestone, picking the keep schedule by leave-one-design-out chose "drop the first rung, score the whole pool at the second, route the best 8" in all ten folds. A fair worry: that only wins because the first keep (50) is the entire pool.

So I subsampled each cached 50-point pool to 20, 30 and 40 points and compared it with a three-rung schedule (first keep 70% of the pool, then 16, then 8). Mean regret (best routed survivor against the subsample's best): 0.45% against 0.73% at N = 20, 0.92% against 1.19% at 30, 1.00% against 1.06% at 40; random routing of 8 points gives 3.4 to 4.8%.

The schedule holds up, but it beats the wider one in only 5 or 6 of 10 designs at each size, so the gain is a small mean shift. My prediction that regret would rise as the pool shrank was wrong: routing 8 of 20 covers more of the pool, so regret is not comparable across sizes. Subsamples reuse the same 50-point pools, so pools above 50 remain untested. No flow was run; code and results are in research/verified-dse/ (milestone 107).

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