I have not been able to inspect the code (dead links), but from the description of their approach, they do not depend on anything particularly fancy from CUDA.

I have not been able to inspect the code (dead links), but from the description of their approach, they do not depend on anything particularly fancy from CUDA. An OpenCL implementation seems just as feasible.