Juan, for what you are saying to happen, we'd basically end up seeing Intel and AMD walk away from Gaming PCs entirely and just basically cancel the platform.
I don't see that happening. Gaming PCs are the fastest growing segment in desktop computing, and most surveys only include sales data from OEMs selling pre-built desktops. They don't count the DIY yourself market at all. And a quick survey of anyone you know of who built their own gaming rig compared to who bought a pre-built gaming rig would (and I say this with absolute confidence) show up as DIY absolutely destroying pre-built purchases.
To put all of this into perspective, the reports that only count OEM sales are saying that gaming PC growth is just a little smaller than shrinkage of overall OEM desktop sales.
The model of AMD making server chips and then selling them as desktop chips is over. It doesn't work anymore, because x86 Opteron is dead. Obviously ARM Opteron will win in the long run, x86 Opteron is under 5% market share. Beating a product with less than 5% marketshare and calling it a win is basically AMD ARM plans.
You keep confusing that model dying for AMD x86 dying completely. If AMD was going to abandon x86 market entirely, they wouldn't care about releasing Mantle because Intel CPUs are not massively bottlenecking GPUs. AMD CPUs are. x86 AMD CPUs to be exact.
So I don't get what you're playing at by saying that AMD will go ARM only, or at least significantly ARM. AMD is going to spend lots of resources making APIs efficient on lower end CPUs to play games that only exist as x86 (it's hard enough to get Linux ports of games, let alone ARM Linux ports) and then abandon x86 computing completely?
Mantle seems like a lot of work to just sell GPUs for a few years and then give up, don't you think so?
If you ask me, AMD plans on keeping x86 around for a while, but keeping it completely out of HPC and servers. Meaning that AMD continuing to sell mobile x86 and desktop x86 CPUs and APUs lines up very well with their Mantle plans.
I keep thinking back to that guy who claimed to work for IBM and the things we discussed when he was drunk. 22nm SOI being a disaster at IBM was one of the claims he made, and now IBM is selling their fabs. Another one was that there's going to be a platform where the problem of making HSA work across multiple devices is solved. There is also HPC discussion of HSA working across APUs in add-in boards in official AMD presentations for HPC.
So while APUs are being pushed right now, they only solve the problem of getting GPU and CPU to cooperate well. Once that problem is solved (which I admit, is really difficult) there will no longer be a need for APU. In fact, I think that when this problem is solved, we will see separation of the two as it makes things easier to cool and improves yields by reducing die sizes of both products. Kind of like how Intel moves some things from CPU to NB because it reduces CPU power consumption, power draw, etc.