Wider multipliers, relaxed clock: why cutting on a cheap rung gets stuck

On SKY130 place-and-route, BOHB and successive halving lose the best design for good on wide multipliers at a relaxed clock; random routing recovers.

The verified-dse programme asks how much of a chip design search can run on cheap rungs (generic-gate synthesis, liberty-mapped synthesis, placement) before paying for routing. BOHB, a strong multi-fidelity baseline, wins at a small budget. This run asked where it breaks.

We ran 240 real Yosys and OpenROAD (26Q3) flows each on 36-bit and 40-bit multipliers at a tight and a relaxed clock, 60 ABC pass sequences per cell. The earlier run on 20 and 28 bits used the same protocol.

  • At 15 routed-cost equivalents BOHB beats random routing in all four new cells (1.83% mean regret against 2.68%).
  • At the relaxed clock, BOHB and successive halving from the placement rung stay at 6.5% (36-bit) and 8.5% (40-bit) regret at 25 and 40 equivalents. More budget does not help, because the cheap rung already discarded the best design.
  • Random routing reaches 0.1% and 0.9% at 40 equivalents, and "route first, then cut" recovers on 36 bits.
  • Together with the 28-bit cell, three of four relaxed cells in this tool version fail; the 20-bit one does not.

Four cells and one design family: this is a count, not a significance test, and the tool version differs from older tables. The tables, the flow results and the replay are in research/verified-dse/experiments/.

All writing · More in Research findings · All tools and research