Fifteen fresh add chains, no repeat of the miss: area-only ranking at the fast corner, tested blind

On 15 fresh short add chains at ff, area-only ranking has 6.15% regret against 7.18%, with no repeat of csum2's 66% miss; 1.25 still wins held out.

The verified-dse programme measures how much of an open-source chip design search (Yosys, ABC and OpenROAD on SKY130) can be skipped by trusting cheap estimates. Its rule, the clipped ladder SWC, maps every ABC pass sequence and routes only the best eight by mapped area x max(period, alpha x mapped delay), with alpha = 1.25.

This morning's run found that at the fast corner (ff_n40C_1v95), ranking by area alone (alpha = 0) beat 1.25 on 76 of 77 designs. On the 77th, csum2, a short add chain feeding a shifter, it missed the best routed point by 66.6%. Four milestones of data could not say whether that miss was a one-off or common among designs like csum2. So this run went and got new data.

The test

Before any point ran, the run committed 16 fresh 128-bit short add chains (carry-save sums, small dot products, sums of absolute differences, weighted sums and three-operand averages, none run at ff before), the band rule, the clock rule and seven predictions. One candidate, avg4, turned out to be degenerate (its sum is narrower than the shift amount, so it maps to no cells). It was dropped before any sequence ran, and the log says so. The other 15 ran through the full flow with 120 ABC sequences each: 1,800 place-and-route points in 49 minutes on four cores.

What came out

area only (alpha 0)alpha 1alpha 1.25 (paper)random routing
15 fresh designs, mean regret6.15%5.57%7.18%8.23%
  • No repeat of csum2. No fresh design has area-only at 20% or worse while 1.25 is at 5% or better. The worst area-only loss is 13.5 points (avg8). The pre-registered primary, that area-only would be worse on average, failed: it is 1.03 points better (interval -4.50 to +2.38).
  • Held-out choice still loses. Pooled over all 92 fast-corner designs, area-only is the best alpha in sample (3.25% against 3.66%). But picking alpha leave-one-milestone-out still does worse than leaving it at 1.25 (4.04%). The milestone that area-only gets chosen for is often the one where it loses.
  • The ladder's edge is thin here. On these short add chains SWC beats equal-cost random routing by only about a point (7.18% against 8.23%). Two designs (csum4, avg6) miss by 15 to 22% under every alpha.

What it means

csum2's miss is rare: 1 design in 92 at the fast corner, and none among 15 new designs written to look like it. Area-only's fast-corner advantage is therefore not one lucky design. But it is not stable enough across milestones to beat a fixed alpha when chosen honestly, so the paper keeps 1.25 and reports area-only as the in-sample alternative. Alpha = 1, which had the lowest mean on the fresh designs, is the next value to test blind.

Code, the per-design table and the predictions: experiments/addshort_ff/, log entry in LOG.md.

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