I read it like this too - no drop in weights or model quality just optimizing the lower boundaries of performance when you are splitting from vram to ram to disk (or network).