The problem with AMD APUs is that they fall into a very awkward niche. For standard desktop apps, the extra GPU power is wasted; they fail to offer any compelling value proposition to make up for the lack of single-threaded performance. For gaming, they're just not good enough, largely due to the use of slow DDR3. A current AMD APU will be lucky to run modern AAA titles at 720p at 30 fps. That just won't cut it.
In order for an APU to provide real value, it has to let players run modern games at 1080p medium settings and get around 60 fps. This is the basic minimum benchmark for a gaming PC. That's going to mean, at a minimum, something like Pitcairn (~1280 GCN shaders) with either shared GDDR5 (which is what the PS4 did) or shared HBM. And the CPU portion of the APU needs to not be bottlenecked by modern loads, which means it's going to need a better CPU architecture than Bulldozer and its derivatives.
The idea of the APU was sound, but the execution just wasn't (and isn't) there. With any luck, 2016 will change that. I believe that the primary reason AMD is playing around with HBM on its new graphics card flagship is because they want to get the experience for future APUs. If it were just the R9 390X, they'd do fine with GDDR5, which hasn't proven to be a bottleneck on Nvidia's Titan X. The eventual goal is Zen (4-8 cores) + GCN (2048 shaders or so) + shared HBM (8GB+) on a 16nm FinFET process, for a good midrange gaming PC on a single APU chip. This will also be the kind of thing that Sony and Microsoft want for the Playstation 5 and Xbox Two.
