are __fadd and __fmul faster than add and mul?

Accelerated Computing CUDA CUDA Programming and Performance
1

Or do they just return rounded towards zero results?