Of particular note in the world of data analytics, there are some exciting things on the horizon for using GPU cores to accelerate database workloads[0][1]. Of particular interest here is removing the communication overhead between CPU and GPU and avoiding the memory translations necessary to move between GPU VRAM to system RAM.
AMD's success in the iGPU market and their HSA design (along with supporting APIs) could be enough to win AMD some market share.
Aren't the economics of this are based on gamers buying large quantities of the chips, thereby spreading out the non-recurring engineering (NRE) costs of making the chips? Is AMD/ATI doing well in this separate from CPUs market? All I can remember is people still complaining about poor device drivers.
The goal of HSA is to break this down. If you have GPU and CPU on one chip with a single pool of RAM that both can access a lot of communications complexity and overhead goes away.
All of this is still very new tech, so what I am saying is conjecture and possibility, but there is a lot of potential in this area.
I am not able to say whether the magnitude of this is similar, but you might think of it analogously to AMD64 architecture. AMD was first to the table with 64 bit processors. They are also first to the table with this. Intel may catch up and blow them away, as has now happened with 64 bit processors, but right now AMDs integrated graphics far beyond Intel's offerings.
[0]http://developer.amd.com/resources/heterogeneous-computing/w...
One possible fly in that ointment: how are AMD's ... arrangements with Intel structured? AMD is playing some of these games because Intel chose them to second source the 8086, 8088, and 286 and that resulted in their getting various rights to future chips (https://en.wikipedia.org/wiki/Advanced_Micro_Devices#IBM_PC_...). If these don't hold if AMD is acquired, it wouldn't likely be attractive. Or companies that only want to play the ARM game, might not be interested in joining all the drama of AMD, including the remaining ATI graphics stuff, which are irrelevant to server chips.
Are you sure ARM at 28nm is that losing of a proposition for servers? Don't ARM designs have quite a bit of power margin to waste? Although that of course doesn't help the performance/power ratio.