I think it is a matter of perception. When Llano was launched, the savior became Trinity. When Trinity bombed and Richland failed to generate interest with system builders, hopes were put on Kaveri. Kaveri, its half-baked HSA infrastructure and the launch of their FX-series mobile processors didn't really help matters either. I believe the real hope for AMD is in the LV and ULV ranges where Mullins and Beema saw quite good success. Progression within mainstream APUs has simply taken too long, to the point where Intel is roughly three generations, if not more, ahead.
One of the main problems is that AMD will only be able to beat the HSA drum for so long before someone calls their bluff. Granted, there are some very hopeful features built into these architectures but even first-generational marketing points like hUMA (
launched two years ago) still remain nothing more than interesting on-paper ideas rather than something that's used on a regular basis. AMD's hopes in this area have been talked about for years. What they don't have is a way to effectively leverage those key selfsame functions. HSA and its functionality was hailed as the raison d'etre for APUs but thus far all we have are years-old APUs which
still can't use portions of their feature sets.
This runs us towards the whole "what comes first, the chicken or the egg?" scenario but it is abundantly obvious that x86 performance in every architecture after Bulldozer has been sacrificed to achieve HSA goals that have yet to come to even partial fruition. Folks are buying processors for their performance NOW, not what their output COULD BE if AMD and their parters get around to actively utilizing what's built into the architecture.
I personally think that HBM will be a non-starter for AMD on the system memory front. With Intel effectively moving to DDR4 all the way to the Cannonlake generation (ie: 2016 / 2017) and possibly beyond, AMD will be in no position to use it as the main system memory interface since no one will be producing add-in modules for it. Rather, HBM will likely remain an embedded memory solution for graphics cards alone. On the flip side, it is looking increasingly like the Hybrid Memory Cube architecture will be picked up by SoC manufacturers since it is backed by Samsung, ARM and quite a few others. In addition, APUs are certainly not starved for memory performance so HBM won't be a silver bullet in that respect.
