For cross-chip communications it does not use CUDA. It is a by-product from the chip prohibition ... and memory bandwidth limitations. I expect to see more of these kinds of optimizations coming out of a number of players ...