The slowing down on Intel platforms has always frustrated me because denorms provide nice smoothing around 0.
At the same time it was nice only having to consider normal floating point when trying to get more accuracy out of calculations, etc.