Look at:

Do you know how to use TLBs with OpenGL?

He suggests that graphics cards should use a MMU to implement virtual memory. No one does this right now, so you can’t program for it and even if someone did implement it you would’nt program for it. The whole point of virtual memory is that it’s transparent to the application programmer.

he’s talkig about adding MMU to GPU so that
an application can have a virtual view of
physical card memory and have control over
its usage. Sorta like processes have
virtual view of physical memory through
CPU MMU (paging).

It would still be transparent to the
application programmer.