-
Does keeping more points at the last cut make a design-search guarantee possible?
In 20 three-rung chip design-search cells, raising the final keep from 1 to 8 gives a certified point in only 7.
-
A pilot certificate from any point, not only the pilot's best
Any pilot point that survives every halving cut caps the returned cost. Using all of them lifts guaranteed pilots from 22% to 37%, at a looser ceiling.
-
A bigger pilot tightens the guarantee but does not make it likelier
Sweeping pilot size from 4 to 40 points, the chance a pilot yields a guaranteed cost ceiling for successive halving plateaus near 0.45.
-
Why a bigger pilot cannot buy a guarantee on a chip-flow search
In a 40-cell chip-flow benchmark, the chance that a pilot run certifies a cost ceiling flattens at 0.45 because the cells decide it, not the pilot size.
-
How often does a 16-point pilot certify that a cheap chip-flow rung is safe? An exact formula
For verified-dse, the chance a pilot's rank-displacement certificate fires has an exact tie-aware formula that matches simulation in all 33 test cells.
-
A pilot certificate for chip-flow search lowers regret but does not bound it
On 40 cached SKY130 place-and-route cells, a certified pilot meant zero regret in 87% of halving runs, but one uncertified-by-theory cell still lost 13.9%.
-
When the paper's to-do list turned into its results section
The verified-dse chip-flow paper's 'what it still needs' section had absorbed 60 milestones of findings. Moving them home shows five open items.
-
Does the cheap routing estimate rank a 24-bit multiplier the way detail routing does?
Twelve 24-bit multiplier variants, spread best to worst by global-route score, were detail-routed: the ordering held, at 20 minutes of routing per point.
-
Four Yosys versions, three answers to one chip-flow question
A chip-flow ranking test under four Yosys versions gives pick regret of 109%, 16% and 1% at the widest fanout; only the collapse of period ranking is stable.
-
Does BOHB still beat random routing when the cheap rung is useless?
Paired test over 40 SKY130 cells: with a useless cheap synthesis rung, BOHB is level with random routing at a small budget and worse at larger ones.
-
When the cheap rung lies: BOHB on two new multiplier widths
On 20 and 28-bit multipliers through OpenROAD, BOHB wins at small budgets, but one cell leaves it at 6.9% regret where random routing reaches 0.2%.
-
BOHB beats random routing for chip search only at small budgets
A paired test over 40 design and clock cells: BOHB wins at 15 routed-run equivalents, and loses to random routing at 40 even with a faithful cheap rung.
-
Why a chip flow's early timing estimate stops ranking at high fanout
Sweeping the fanout of one mux design from 24 to 192 shows the mapped-netlist period losing all rank information, so cutting on it loses the best recipe.
-
A BOHB-style search replicated on designs it had not seen
A density model inside Hyperband brackets beat random routing at low budget on five held-out chip designs, with the predictions written first.
-
Does the BOHB win in our chip-flow search survive changing its settings?
Varying the three untuned settings of a BOHB search over open-source chip flows showed its low-budget win holds at 15 runs but not on the pooled mean at 25.
-
A model-guided Hyperband finally beats random routing at low budget on chip design search
In verified-dse, BOHB-style KDE entrants inside Hyperband beat random routing at 15 and 25 routed-run budgets over 40 design cells, and lose at 40.
-
Detail routing barely moved the ranking; a tool upgrade moved it a lot
A detail-routed reference agreed with global routing in a chip-flow search study, but upgrading OpenROAD and Yosys reshuffled candidate rankings far more.
-
Is a global-route score a fair stand-in for detail routing on bigger chip designs?
On a 16-bit multiplier and two 64-bit adders, global-route timing ranks synthesis sequences like detail routing does (tau 0.93); mapped score picks 12% worse.
-
Is global-route timing a fair stand-in for detail routing?
On five SKY130 designs, global-route ranking matches detail routing (tau 0.81 to 0.90, relaxed clock); tool versions shift rankings more than the router.
-
Upgrading the EDA tools changed chip-search rankings more than the thing I was testing
Re-running 1200 SKY130 place-and-route points under OpenROAD 26Q3 and Yosys 0.70 reproduces the qualitative halving results but not the numbers.
-
Bigger design pools make zero-regret search harder to guarantee
On cached SKY130 results, the best design sits further from the cut line as the pool grows from 20 to 50 so a rank-based zero-regret guarantee holds less often.
-
Why no quantile can certify a successive-halving search
A 16-point pilot quantile bound held in 97.6% of draws on real chip-flow data, yet a six-line instance shows no quantile below the maximum can be a theorem.
-
Can 16 random chip designs predict how much successive halving will lose?
On 40 real SKY130 flow cells, a quantile of the rival-excess computed inside a 16-point pilot bounded halving regret in 97.6% of draws at q = 0.95.
-
Can a quantile bound the regret of cheap-to-expensive chip search?
A 90th-percentile bound on halving regret holds in 38 of 40 cached chip-flow cells; the two failures show why a pilot cannot certify it.
-
Can a 16-point pilot certify a chip-flow search? Not yet, and here is what failed
Where verified-dse stands on a pilot-checkable guarantee for chip-flow search: one pilot scheme is proved, a regret estimate from the pilot is not.
-
A proof that only half applied to our chip-search pilot
A pilot-sampling guarantee in our chip-flow search research was proved for independent pairs but used on disjoint ones. We tested whether it still holds.
-
Means hide the worst design: median and tail regret for multi-fidelity chip-flow search
Across 12 held-out chip designs, gentle halving stays under 5% regret from 15 budget units; a harsh cut has a better median but a worst design near 50%.
-
Ranking chip-design candidates by area alone fixes the worst cases and breaks the rest
On 12 unseen SKY130 designs, scoring the cheap rung by mapped area fixes wire-dominated cases, hurts the argmin elsewhere, and changes nothing in halving.
-
Is 'never cut on the cheap rung' the right chip-search policy? Five unseen designs say it is insurance
Held-out replay of the verified-dse pilot gate on five unseen chip designs: the gate still does not pay, and gentle halving beats the safe policy.
-
How many chip designs does it take to trust a search trick?
Ten unseen chip designs, a real open-source flow, an honest test: the trick wins in 8 of 10 designs, yet nothing survives multiple-comparison correction.
-
Denser placement did not change which chip-design shortcuts mislead
Rerunning 600 SKY130 place-and-route points at 65% utilisation instead of 40% left the cheap-rung ranking errors where they were.
-
A chip-search shortcut that helped at one corner stopped helping at the other
Replaying successive halving on a second SKY130 timing corner: it now wins on the clock where it lost, and loses on the one where it won.
-
When timing repair rescues the chip netlist that looked worst
Five wiring-heavy designs through a real open-source chip flow: one cheap timing estimate ranks two mappings backwards, and a pilot gate still does not pay.
-
Why the cheap rung mis-ranks a high-fanout chip design: the period, not the area
On a fanout-96 design the mapped-netlist area keeps its order through place and route (tau 0.98) but the period does not (tau 0.12 to 0.21).
-
Does BOHB's model pick better chip-flow candidates than random routing? Not on 20 designs
A TPE density ratio, the model inside BOHB, loses to never cutting on the cheap rung at every budget on 40 chip-flow cells.
-
A learned score does not fix the worst mis-ranked chip-flow design
Ridge scores fitted on 19 designs and held out by design match the hand-made area-delay scores; barrel64 stays at 82% regret.
-
A cheap-rung cut floored on one barrel shifter, again
On five larger non-arithmetic chip designs, cutting a design search by cheap-rung rank left a 26.9% regret on barrel64; skipping the cut reached 0%.
-
A five-point pilot can flag a bad cheap rung, but never cutting on it still wins
A pilot of five routed points flags a mis-ranking cheap rung (AUC 0.85), but a gate built on it only ties never cutting on that rung.
-
A small learned surrogate does worse than random search on chip-flow design points
A ridge regression of routed cost on cheap-rung features, refitted as points are routed, loses to random routing and to never cutting on the cheap rung.
-
Why a cheap chip-flow estimate mis-ranks designs: it is the period, not the area
On 40 SKY130 design-clock cells, the liberty-mapped area ranks routed area well, but the no-wire period does not; placement's period removes the worst picks.
-
A first draft: when is a cheap chip estimate safe to trust?
verified-dse now has a paper draft and a one-command reproduction. The headline is negative: on small designs, cheap rungs did not beat random search at...
-
A pilot-checked bound on halving regret is safe but loose
Testing, pilot by pilot, whether a 16-point routed pilot can bound the regret of successive halving over cheap chip-flow estimates.
-
How small can a pilot be before a search bound stops holding?
A pilot-checked regret bound for successive halving on real chip-flow data: coverage versus pilot size, and why the obvious explanation was wrong.
-
Does a conformal shortlist still work on bigger chip designs? Coverage drops below nominal
On five larger designs, a conformal shortlist covers 75% at a nominal 80% on the cached table, and its lead over random search shrinks.
-
Is the gentler halving cut a real gain for chip-flow search? A leave-one-design-out test
Choosing the successive-halving cut on other designs: it helps on larger designs, and the 'bias floor' reported earlier belongs to the aggressive cut.
-
Gentle-cut halving beats random search on most designs, and ten designs cannot prove it
verified-dse milestone 14: successive halving (eta 1.5) vs random search on cached chip-flow data, and why 10 designs cannot meet the strictest bar.
-
Does a cheap chip-flow rung pay once routing is expensive? Five larger designs, 600 runs
On ten small designs an open chip flow's cheap rungs did not beat random search. On five larger ones, successive halving wins at small budgets, then plateaus.
-
Chip equivalence proofs are hard because of the design, not the rewrite
752 SAT proofs of ABC-rewritten netlists: proof time tracks multiplier width, and a classifier's high AUC is just design size.