The CU's are invisible to the system beyond the GPU providing an access point. The use of CUs is handled by GPU scheduling. Enable an extra CU or two and the GPU will use it for whatever work it as to do; there's no means in GCN AFAIK to wall a CU off from the rest of the GPU and handle it manually. The video encoding hardware will also be perfectly up to the task of the sharing requirements, unless it's a poop design.
PSP's RAM was different. You can easily limit games to seeing only half the RAM, and the OS has full access to all the RAM. You could certainly add more RAM to a console and wall it off from games. You can't do that with CUs are described in the current GCN architecture though.
