Well remember the description for small FFT "data fits into l2 cache, not much ram used", so i'd assume that means it doesn't go into the l3 cache either, which is the largest cache. Blend will put load on the uncore aswell so not surprising to see temps higher.