80 Million Candidate Buildings in 42 Minutes: A GPU Genetic Algorithm for Generative Design with NVIDIA Warp
GD–AI Loop 2.0 · Lab Notes #5 · the pipeline behind Lab Notes #6–#8 Before the evacuation experiments of Lab Notes #6, we rebuilt the layout generator of our small testbed so that one GPU thread builds one candidate building: it decodes a genome into rooms, doors, corridors and stairs, checks the rules and scores the result. With that pipeline a genetic algorithm evaluated 80 million candidates across 480 design conditions in 42 minutes. The 6,334 designs it kept then let us ask — with the decision rule written down first — whether a learned model is needed to propose designs for a new condition. Oct 25, 2026 (English edition Nov 12) · Next: Lab Notes #6 — evacuation simulation as the fitness function · Lab Notes #7 — early-stage site feasibility · Lab Notes #8 — a verified crowd model Read this first — the crowd model shown in this post has been replaced. The crowd evacuation in this post uses the first crowd model of the series (called v1 in the later notes): one person per 1....
