I have rarely observed that much difference between CUDA and OpenCL. I have, however, noticed that CUDA (at least by default) is more aggressive about picking

I have rarely observed that much difference between CUDA and OpenCL. I have, however, noticed that CUDA (at least by default) is more aggressive about picking faster but less accurate instructions for transcendental functions like inverse square root. You can ask for the same kind of stuff by passing extra OpenCL compiler flags, but it seems that CUDA optimises more aggressively by default (maybe analogous to the kind of stuff -ffast-math does).