Can it visualize PyTorch Tensors without an additional memory copy? I.e. mapping it directly to a texture which is then displayed?

I know this is not much of a concern on a system with unified memory (all recent apple computers).

Not yet. Some internal prototypes did try using IOSurfaces on macOS to go with the zero copy route, but there were a fair number of limitations.

That said, the IPC minimizes copies and is actually fairly efficient at handling large numerical arrays.