Skip to content

Ten engine experiments on small boards: faster nodes, no smaller trees

Ten small preregistered engine experiments on generated boards from 6x6 to 9x9 kept five and dropped five. The one measured speed-up is a node that costs 5.3 to 6.1 times less for the identical search; none of the three methods that change the tree (blocker backjumping, value order by impact, Hall matching) shrank it against its baseline. Nothing here is Eternity II progress.

by Raphaël Anjou

experimentmeasurementmeasuredBacktrackingReduce the searchGo fasterUpdated 2026-11-14
Reproduceseeded: reproduces with the given seedreruns the searchBudget: about 1 h for the first experiment, 27.37 core-hours for the 9x9 pilot, 19.1 summed per-run hours for the tournament
Hardware & run
Native runCPU only
2.0core-hours
Cores
2
GPUs
0
CPU
Apple M4 (4 performance and 6 efficiency cores)
Machine
10-core Apple M4 laptop, shared; two cores per experiment (one for the timing runs). Some experiments also ran on an 8-core Apple M1 and a 14-thread x86 laptop.
Budget
about 1 h for the first experiment, 27.37 core-hours for the 9x9 pilot, 19.1 summed per-run hours for the tournament
Start
seeded; fixed seed lists on generated boards, node counts identical across the three machines wherever compared
Reproduce this result

seeded: reproduces with the given seed

just experiments mrv-node-cost

Rebuilds the cheaper-node solver, re-checks that its node sequence equals the original on the 140 gate instances and re-times it; it does not re-run the other nine experiments, whose commands are listed below.

Code & data on GitHub

Keep exploring

Referenced by

Page sourceView as Markdown