That cost reduction seems to stem from cheaper cache reads, mostly.