GPUs have always had amazing floating point performance for single precision floats, that's basically all graphics is. The problem is that double precision support was traditionally non-existent, and has been slow until recently and it is still much slower than single precision float performance.
Double precision performance is now just half speed, which is the same as the CPU. Many applications are bandwidth bound, so running in double vs. single doesn't end up with a 2x difference anyway.
Comments
GPUs have always had amazing floating point performance for single precision floats, that's basically all graphics is. The problem is that double precision support was traditionally non-existent, and has been slow until recently and it is still much slower than single precision float performance.
Double precision performance is now just half speed, which is the same as the CPU. Many applications are bandwidth bound, so running in double vs. single doesn't end up with a 2x difference anyway.