I think this is only true if using x87 floating point, which anything computationally intensive is generally avoiding these days in favor of SSE/AVX floats. In the latter case, for a given vector width, the cpu can process twice as many 32 bit floats as 64 bit floats per clock cycle.