Can cheap word-level RTL features predict synthesised cell counts well enough to screen a design search, and does that survive a new design family?

What it found

A surrogate for Yosys generic-gate cell counts, trained on 24 word-level features (multiplier, adder, mux and register sizes read after proc; opt, about 30× cheaper than synthesis) over 133 designs in four generator families I wrote: MAC engines, FIR filters, ALUs and sorting networks. On families it saw, a non-negative linear cost model has a 3.9% median error and within-family Spearman 0.995, beating boosted trees (8.5%). When a whole family is held out, median errors run from 10.5% to 210.7% and the best model changes with the family, but rank agreement stays at Spearman 0.75–1.00. Ordering a 36-design pool by predicted cells×cycles found a design within 2% of the best after 1–5 syntheses against 12.4 (MAC) and 18.6 (FIR) for random order, both optima sitting at an edge of the space, so the test is weak. Stylised: synthesis-only target, no place and route, small designs.

A figure from its own output

Paper, code and results

• README and method
• Experiments and results
• Code
• Tests

More in Algorithmic chip-design loops

• How many synthesised designs from a new family does a cell-count surrogate need before it is usable, and is it better to pool them with old families or train on them alone?Few-shot adaptation of the eda-surrogate cell-count model to an unseen design family: one synthesised design from the new family cuts the linear model's median error from 49% to 18%, training on the new family alone beats pooling from about three designs, boosted trees need about 12 (stylised, synthesis-only target, no place and route).
• Which few designs of a new family should you synthesise to adapt a cell-count surrogate, and should they count for more?Which few designs of a new family to synthesise when adapting a cell-count surrogate: a farthest-point spread in feature space roughly halves the linear model's error against random draws at 5 designs (5.3% against 8.5%), the smallest designs are worst, and upweighting new points helps linear and ridge but not boosted trees (stylised, synthesis-only target).
• Does the farthest-point spread still beat random draws when every chooser is scored on the same held-out designs?eda-adaptpick's adaptation-design choosers re-scored on one fixed test set per family (4 families x 5 splits): the farthest-point spread still beats random draws for the linear model at every k (4.8% against 6.0% at 5 designs, w = 5) but wins in only 65% of cells, the smallest-first rule stays worst, and the earlier 'halves the error' margin shrinks to 20-45% at 5 to 8 designs.
• Does a cell-count surrogate used as a screen find a chip design’s Pareto front with fewer syntheses than random search or TPE?A cell-count surrogate used as a screen in the design-search loop: 4 spread syntheses plus a log-linear fit reach 0.990 of the true front's hypervolume at 12 syntheses against 0.860 for random and 0.878 for TPE, of which about +0.05 comes from free cycles/error objectives (stylised, one 80-design space).
• How much error in a cheap objective can design-search proposers tolerate before they stop beating random search?With Gaussian error on the objectives a proposer sees, TPE's budget-30 lead over random falls from +0.037 to +0.006 (gone at error 0.8 of the objective range), NSGA-II and the genetic method lose theirs by 0.1, annealing is never better (stylised, synthetic noise, one 80-design space).
• Does the ranking of design-search proposers survive a change of hypervolume reference point or quality indicator?TPE stays best under five hypervolume reference points, recall, IGD+ and epsilon, but its lead over random at budget 30 ranges from +0.045 to +0.002 with the reference point (stylised, one 80-design space).

All 42 Algorithmic chip-design loops projects →

Other research topics

Proper scoring & elicitation28Effort & peer prediction9Verification games & slashing19Detection & tolerances14Markets as routers13Prediction-market design9Paying contributors8Decentralised training systems23Digital twins & sim-to-real164LLM chip design loops40Robot manufacturing loops53Ambient AI assistant46Self-replicating manufacturing34Simulation-first design64Engineered clothing45Task-specific AI accelerators40Algorithmic chip-design loops42Open-source purpose-built robots44The software-engineering loop43Adversarial human imitation41Open-source residential construction41Autonomous multi-angle capture34Sensing, imaging and 3D31Energy supply and dense energy storage39Robot societies among people31Agentic co-design of low-cost robots27Laundry-hanging robots25Machine learning on qualitative data23Robots on the building site27Civilizations & group psychology1History & human sciences1

Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.