hcl123 :
I think Kaveri on FM2+ will be hUMA only on the APU side, put there a discrete card and it will not have that discrete GPU cache, coherent with the CPU cache. There are IOMMU tricks to make the GPU card see the same system virtual memory (VM) , will be good for virtualization solutions (all APUs have an IOMMU, which makes all of them better VMware players as example, than any intel chip... in this respect for the FX, only the top 990FX chipset have an IOMMU but old version by now), but will not improve performance not nearly as much as GPU card cache coherent with CPU cache, if this solution is to account for.
Don't take me wrong, kaveri seems a very good solution, already better than what is similar out there, better than Intel for sure for graphics and compute... a lot depends on how the drivers deal with the "crossfiring", if they address the discrete adapter only for graphics and video (which intel will not dispense for real gaming), then only the APU GPU will do compute, compute jobs where the lack of cache coherence is more penalizing... which is a very good interim solution... and those platforms are not for the top of "enthusiast gamers", but they can manage very well.
Kaveri will remain 4 cores top i suspect (good bet since Windows doesn't really take advantage of more now), but this time will NOT lose anything of the already ridiculous small difference for intel core to core... quite possible gain to hasfail core to core lol (of course bentmarks are already being "optimized" to show me wrong in next editions, is one of the easiest bets ever lol)
The best of the APUs is they are the best performance for the buck ever, top of those chips are around $100, is a sweat deal.
In the future is quite possible the APU platforms having HTX+PCIe combo slots ( like this: __________ x16 main slot + _____ smaller sideband on the same row, only HTX uses the sideband ) and so the APUs having also proper interfaces "on die" ... and 4 cores/threads per module (8 total)... and nearly has good for top "enthusiast gamers", but is not the case for kaveri for FM2+.
This is close to a pair of things that I wrote in this thread some days ago:
The first is if a new version of dual graphics for HSA APUs could use the full iGPU for compute whereas the dGPU is used for graphics.
The second is if the iGPU of an APU HSA could be fused with a HSA enabled dGPU for dynamic split into compute or graphics according to workloads.
The former is technologically simpler, cheaper, and gives enough performance: the APU would work as a 'CPU' (HSA enabled soft.) with about 4x the performance of the FX-8350. The problem is that goes against the HUMA architecture and returns, basically, to a traditional CPU+dGPU architecture.
The second is technologically complex, expensive, maintains HUMA, and has the potential to give still more performance: the dGPU could help the APU in some heavy compute workloads. The main problem would be on the interface.