"CUDA" is not a particular application. You could very well be doing something with CUDA that does extremely little data transfer to/from the host, but a lot of computation on the GPU. Like mining :D

Also, host to device can be bottlenecked by memory speed on either side.

CUDA meaning GPGPU. All I was pointing out is that there is definitely a bandwidth doubling going from 8 to 16 lanes. It's used extremely often by people using the cards for compute. Mining may not have that issue, but deep learning and other applications do.

And even for games, the only reason it's not a difference is because game developers target less bandwidth between CPU and GPU so they don't use it all.

If we move to a future where every GPU has a lot more bandwidth then games would most likely start taking advantage of it.