Not a single bus. You'd have RAM connected to the memory controller on each APU. If one APU has to access memory that was physically connected to the other it would do so via a bus connecting it's memory controller to that of the other chip.
AMD's existing MOESI schemes rely on the memory controller to be the final arbiter of snooping requests, so they themselves do not snoop.
http://www.realworldtech.com/qpi-evolved/3/
Putting a bus between memory controllers would be on the wrong side of the cache/memory divide unless the idea is that one APU's controllers are slaved to the other up to the point that not even the cores local to the slave controllers can directly access them without going through the other chip. There's currently no infrastructure to communicate to another SoC, and even it were added it would need to differ from AMD's existing architecture that uses the controllers as their home agent.
Slaving one APU's controllers to the other would wind up creating a link between controllers and require an HT interconnect to transfer snoops.
An alternate formulation of AMD's memory architecture that moves the complication of creating memory controllers with the intelligence to snoop each is a large enough endeavor on its own, and it's not clear how much the cache subsystem that underpins Jaguar would need to change. The non-trivial aspect of this is indicated by how long AMD has gone without touching this distributed system.
It would seem less disruptive to create some kind of hybrid Xfire solution, similar to AMD's laptop APUs having a matching discrete GPU. It would then require explicit management by software and would add further division in the unified memory model. Compute that relies on Onion+ might not be portable or useful, although this would explain why legacy games could not use the extra processing power or bandwidth without a patch. It would be difficult to not get some kind of unwitting benefit from a stronger GPU or CPU block.
I am curious where the chatter is about a PS4 slim model. Unless Sony thinks the PS4's success has sopped up so much demand, the slim model would be part of the life cycle where more general buyers that balk at early adopter pricing or the bulkier and hotter initial implementations buy in. Not putting that out there is taking the foot off gas in that broader market.
Also unclear, given the speculated significant graphical upgrade, is why increasing performance allows for a smaller console when the power benefit has been eaten up--unless the thing stays as loud or louder. I suppose if a PS4K is using a hybrid Xfire solution, it could be a PS4 Slim + GPU rather than designing two different APUs or forgetting about slimming the console down.
