I believe it’s the depth mapping node that requires CUDA.
I think there are some other nodes that have the (advanced?) option to do CPU or GPU (CUDA).
It looks like a lot of work to move to another GPU-compute library, and there doesn’t seem to be a maintained, cross GPU, cross OS system yet. Eg I had a quick look (again) yesterday and found SYCL: AMD, Intel, ARM, Linux, Windows… but no OS X, no Nvidia.
So the only way (right now) would be to create an alternate version of that step. Which at least is possible (thanks to the steps being done as a pipeline) but requires someone to step forward and do the work, and for the main devs to be willing to accept it. They’d have to learn the other API(s) to be able to maintain it long run. It’s possibly not too complicated in hind-sight, but probably a fair amount of work up front.
Julian