Namely AMD would have to publish the hardware documentation (for instance command packets) of anything specific to inference (unless, this is just their GPU doing only compute).