Think of it this way: If they were thinking of using a APU with 10 compute units, and a GPU with 20, it would be a lot cheaper in terms of design, OS work, fabricating et al to just use a GPU with 30 compute units to start with. Same overall power, no multi-GPU integration headaches. (and fewer transistors, since you wouldn't need to duplicate the common sections)
30 compute units in an apu, is such a thing even possible?