I've actually a tough time believing that somebody will use "cat' type of CPU on 14/16nm but MAD won't do it and try to get into a couple windows tablet/nettop/laptop. No matter the rumors I believe the PS4 Neo will use that good old 28nm lithography we know and like.
From here I expect a pretty huge chip. Looking at the announced specs mostly GFLOPS throughput and clock speed, I believe that the power budget would be a little high even for a system bigger than the PS4.
I expect Sony to have refined its system and have now a proper power management feature from the late "cat APUs". So I think the rumors we have are about the top clock speed and not the sustained performances. The reason I believe that are the following:
To push higher frames rates on the CPU side, your are going to need higher clock on the CPU which will raise the power budget significantly by self, so before you ever render something you need extra power. You need to make room for the power.
Say you accounted for the extra power and you are in a situation to fed the GPU, how much CPU do you need to render at say twice the frame or twice the resolution (depending on how the experience is tweaked)? Twice as much sounds right-ish. Now I do believe the numbers we have are to high to do just that, especially as the system could/should be built around newer IPs (which further improve the real world performances).
So my take on all of that is that an extensive power management system will make it so the system does what it has to do within a beefier TDP yet a more reasonable TDP than what you would expect for such big GPU (+ the CPU cores) in the PC realm.
Going down further that rabbit hole (so a 28nm one), power may have been a major concern in the design, especially if Sony want system to be relatively sleak and silent and sell it at a fair margins.
Memory and memory controllers burn power, so if both act faster, the end result is easy to predict: more power, power than comes on top of the extra CPU power already burnt. So I think to would make sense to Sony to have move to a wider bus width: 320 bits along with memory acting (on average) at a slower pace.
How would that match the GPU, I could see the GPU consisting of three shader engines, each one consisting of a the appropriate fixed function hardware (geometry engine, etc.) 12 CUs, 16 ROPs. So the GPU would consist 36 CUs and 48 ROPS.
I could see that 36 CUs figures around being the total number of CUs on die, not the total number of CUs activated (which may depend on what comes back from the foundry). That is already big on 28nm, either way it has more than 36 CUs...
I would not shock me if the system end up anywhere in the 30ish (I pass on a random bet).
While the system is in PS4, one Shader engine would be shut off, and the two left would run at slower speed to emulate the two available in nowadays PS4 system (/I've no idea about the level of control AMD has here, disabling an engine sounds reasonable, cluster of CUs? I don't know).
I'm going back to the memory, I also believe a wider bus make sense because I can't explain myself why the new system would have more memory available for the OS or games (unclear at this point to me at least, yet more memory has to come from somewhere). A 320 bits bus would explain that:
Sony may have linked the SOC to 9GB of GDDR5 (mixing different capacity chips operating at the same speed), in a case akin to what happened to a Nvidia GPU, may 512MB may end up slower and SOny prefers to hide it away (or for themselves who now).
So that it the system could end being pretty reasonable in power draw, Sony may have traded power for an increase in silicon area and proper turbo mechanics.
