Passthru or time sharing? The latter is difficult because you need something to manage the timeslices and enforce process isolation. I'm no expert but I understand it to be somewhere between nontrivial and not realistic without GPU vendor cooperation.
Note that the GPU vendors all deliberately include this feature as part of their market segmentation.
