I heard Jim Keller say on a podcast recently that many computation problems are N^2, so if unoptimized and run on silicon with 2x the transistors, it runs 4x slower.