Testing a fix that follows from the diagnosis: a delay check on the chip-flow ladder's top picks fails

Routing the ladder's top 3 and checking whether the smaller netlist is slower flags 47% of cells where the ladder is perfect, and 45% where it fails.

The verified-dse programme searches over ABC logic-synthesis pass sequences in an open chip flow (Yosys, ABC and OpenROAD on SKY130). Its working rule, the clipped ladder, maps every candidate, ranks the candidates by mapped area times a clipped delay, and routes the best 8. The previous milestone's diagnosis found that the rule fails where the smaller netlists place and route slower, which no statistic from the mapped netlists shows. It proposed a guard: route the ladder's top few points first, which costs nothing extra because the ladder routes them anyway, and check whether the smallest of them is also slower.

This run tested that guard. If the smallest-area point among the top 3 routes more than 2% slower than the fastest, the guard spends the remaining 5 routes on points at least as large as the fast one (or, in a variant, on random mapped points). Every branch routes 8 points, the same budget as the ladder. Before replaying it on the 90 cells of seven earlier flow tables, we committed six predictions.

The check fires as often where the ladder is right as where it is wrong. It alarms on 45% of the 17 cells where the ladder loses to random routing and on 47% of the 40 cells where the ladder finds the optimum. Among the top few points, a smaller netlist routing slower than a slightly larger one is normal. It is not a sign of failure. What decides the ladder's regret is whether the routed optimum is inside the top 8 at all. On failing cells that optimum sits around the 26th percentile of the ranking, far below anything a check on the top 3 can see.

So it costs about what it saves. Mean regret is 5.01% against 4.77% for no guard and 3.91% for the 32-gate rule on these seen cells. On the 40 cells where the ladder is perfect, the alarm usually changes nothing, because the next ranks already sit in the larger-area band. On two of them it costs 19% to 20%. On the failing cells it saves one point (15.64% against 16.63%), almost all of it on one 128-bit shifter. Sending the alarmed routes to random points does no better. Four of six predictions were refuted.

The cells behind the idea made it look better than it is. On the 42 blind cells that the diagnosis came from, the alarm separated the groups a little (50% on failing cells against 28% on winning ones). That separation did not carry over to the 90 seen cells. A rule suggested by a set of cells needs testing on different ones, which is why this one was replayed on the 90 seen cells.

This closes the search for a probe-based guard on this rule. Random probes, top-against-uniform probes, and now probes within the top have each failed to beat having no guard. The unguarded clipped ladder is the rule the paper will report. Its failure mode, a routed optimum outside the mapped top 8 on designs where smaller netlists route slower, becomes a stated limitation and an open problem.

Limits: this is a replay of committed flow tables, not a new flow run. It tests one alarm statistic with two fall-backs, 200 draws per cell, and splits the cells into groups using the ladder's own outcome. Code, per-cell table and tests are in experiments/delay_check/.

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