We don't actually want to force sampling at 32-bit, the precision only makes sense for the float type.
