Denser placement, more designs: the routing shortlist wins 38 of 46 at 65% utilisation

26 more blind SKY130 designs at 65% placement utilisation: the cheap shortlist beats random routing 22 to 4, and its edge matches the 40% one.

The verified-dse programme asks 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, ranks the mapped netlists with a cheap score, and routes only the best 8. The baseline routes 8 sequences picked at random, which costs the same.

The last run reran the blind designs at 65% placement utilisation instead of 40%, but its flow stopped after 20 of the 52 designs it had planned. Three design families had no design in those 20. A result from a prefix like that could be luck, so this run picked up the same seeded order where it stopped.

What this run did

Same flow, designs, clocks and pass sequences, nothing retuned. Eight predictions were committed before any point ran. The flow again stopped at its two-hour limit, this time with 26 of the 32 remaining designs complete (3,069 routed points).

What came out

All eight predictions held.

designsSWC mean regret at 65%randomSWC better / worsesign p
the 26 new ones2.41%5.75%22 / 40.0005
all 46 at 65%2.46%6.35%38 / 8< 0.0001
the same 46 at 40%2.77%6.05%37 / 9< 0.0001

At 65% routing costs more relative to mapping, and the routed ranking moves: the best sequence is the same at both densities on only 24 of the 46 designs. The ladder's edge over random routing still barely changes. The shift is -0.61 points (95% interval -1.60 to +0.37), and the design-level sign agrees on 39 of the 46.

Restoring division is the exception. It is the only family where the ladder is not ahead of random on the mean (+0.24 points, 4 of 8 designs lost). It was the weakest family at 40% too. A chain of compares and subtracts is where the cheap score misleads, at both densities.

What it does not show

A second 256-bit design lost about half of its points to OpenROAD pin placement at 65%, because the smaller die had too little perimeter for its pins. Without the two affected designs the result is 36 to 8. Six designs are still unrun. This is one more utilisation, not a second library corner, on one tool version, and it uses global-route estimates rather than detailed routing.

Code, rows and the full table: experiments/util65_rest/.

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