AFAIK it comes down to efficient FFTs for multiplying the large numbers, which are well-suited to GPUs, and this could likely have been done a lot faster on a GPU.

I'm guessing there just isn't a good CUDA implementation out there.