Chip design search: hedging a failed spot-check barely helps, because the check is the problem

On 90 replayed Yosys and OpenROAD cells, keeping three ranked candidates when a spot-check fails moves mean regret only from 4.39% to 4.37%.

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 ranks mapped candidates by a cheap clipped score and pays for place-and-route only on the best 8. On some wide shifters that ranking misleads.

The last step added a free spot-check: compare the best of a random 32 mapped candidates against the one random candidate the rule already routes to time the flow. If the random one wins, stop trusting the ranking and route randomly. That beat the unguarded rule on mean regret (4.39% against 4.77%), but on cells where the ranking was perfect, the check fired falsely in about 12% of draws and cost 1.15 points.

Hedge instead of abandon

The obvious fix is to soften the response: when the check fails, still route the subset's next three ranked candidates, then go random with the rest. Those three are already mapped, so they cost three random routes out of about 30. We wrote the rule and six predictions into the repository before computing the table (commit a1f0fd90a), and replayed it on the same 90 cells, draw for draw against the previous rule.

What happened

rulemean regret, 90 cellswhere the ranking is perfect (40)where it fails (17)
no check4.77%0.00%16.63%
spot-check, random fall-back4.39%1.15%11.83%
spot-check, hedged fall-back4.37%1.11%11.79%
spot-check, map all then hedge with global ranks4.69%0.53%14.94%
32-gate size rule3.91%1.35%8.82%
random5.44%6.56%4.25%

All six predictions held, and the gain is 0.02 points (interval -0.04 to -0.01). It is better on 24 cells and worse on 1, but it barely moves anything. The reason is in the subset: on a false alarm, the subset's next three ranks hold the cell's best candidate only 19% of the time, because 32 random candidates rarely include the global winner. The hedge that could recover it uses global ranks, which means mapping everything first. That halves the false-alarm cost but leaves too little budget for the random fall-back on the cells where the ranking really fails.

What it means

The fall-back is not where the loss comes from. To get past the simple 32-gate rule, this check has to raise fewer false alarms in the first place: a margin on how badly the random point must beat the ranked one, or a second probe pair. That is the next step, priced by replay first, then tested blind with real flow runs on new designs.

All numbers are from replaying recorded flow results (no new place-and-route run), on cells earlier steps have already seen, so this prices the rule rather than testing it blind. Code and table: experiments/hedge_probe/.

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