A better question would be how it will compare to CUDA implementations when it's released?

Moore's law killed all such initiatives in the past, however, perhaps this time it will be different?