APU's are in a bad position where they're not really enough for gaming, and too expensive for anything less than gaming, especially since even the cheapest motherboards include an onboard GPU that would do just fine in an HTPC on a budget, or even an ultra cheap discrete videocard would be more suitable and probably cost effective too. No one really needs that in-between performance. A quad core with an APU is just a joke.....in the sense that if an APU is enough video performance for someone, then a quad core anything is most likely too much processing power for their needs.
The performance of a Kaveri chip is still superb, there's only the Iris Pro from Intel which can beat it and that chip costs a whack. So for someone who needs a quiet space or something with decent
onboard video without having to spend money on a video card (think power consumption) it's a very good choice. Even the 7850Ks seem to come close to their rated TDP and tend to use around 75W on average. Now that's pretty good considering what's on the die.
Also a quad core is a joke? Your Intel Core I5s and I7s are quad cores... It's only when you get onto the 2011 socket for instance their octa cores are available, and that price range is another step up completely so that doesn't make sense. Intel simply using hyperthreading to pretend there's more than 4 cores.
And while the CPU and GPU may share memory in HSA and reduce some overhead of copying back and forth, system memory still has a limited bandwidth divided between both.
No matter how powerful of a GPU you put in an APU, memory will always be the bottleneck. GPU's need 100-200GB/s for good performance, and no system memory is remotely that close.
I see people who buy 2000mhz+ RAM at a premium, just to squeeze more performance out of their APU's. Seems kind of silly to spend something like $150-180 on a quad core APU, then spend $40 or more extra for faster memory that still bottlenecks it greatly.
Well no. Traditionally the way GPUs and CPUs communication is over the PCI-E bus meaning it has to copy everything over and back again. That's inefficient. By harnessing a common buffer (RAM) the CPU and GPU can access it simultaneously
System RAM is slower yes but the point is tasks can be split up serial and parallel accessing common cache which is far quicker than copying everything back and forth. Working on the same memory is beneficial. Yes there is a long way to come but the effects at the moment are nothing short of amazing. Yes there is latency but the process is a whole lot quicker and it's full potential hasn't been fully used yet, which is the exciting thing.
Another point is the potential of these chips. Take a look at this testing here http://benchmarkreviews.com/13238/amd-a10-7850k-performance-optimized-catalyst-14-2-driver/4/
You've got a chip rated at 95 TDP, quad core with 8 GPU cores driving some popular new titles. Mantle has improved performance, Catalyst updates to drive GCN. And the result is amazing. We've got a chip able to game some popular titles at 1080P sometimes above the minimum 30 fps.
That isn't bad, it just shows there is potential there. The thing flies along when paired with a discrete card such as the R7 250. For hard core gamers wanting absolute maximum quality will sniff at it, but the truth is it's a completely viable option for a lot of people out there wanting to game. Short on cash etc.
That's why there's been some people over here (including myself) jumping over, because it has potential. We've all seen the benchmarks where HSA is in place, and the result is an A10-7850K is much quicker than the I5, sometimes 3x as quick.
That's why AMD is going down the HSA route, it has potential.
What seems to be the logical simpler step is not an all-on-die design, but rather a system on a chip design. You can't exactly plop RAM onto the die for a variety of reasons.
But this would remove the need for third party motherboards and RAM... that would be a difficult path with all the companies out there that make motherboards and RAM. Then you'd have the problem of features such as the typical slots you'd find on a motherboard. Variety of options would be out of the question, limited to about 3-5 system on a chip configurations.
Two people may want the same CPU but different GPU's or different amounts of RAM. Companies have enough trouble with sales as it is just by having different colors of products like tablets. With moherboards and RAM, there are tons of possible combinations that would be impractical to offer in a SoC.
AMD is coming out with a SoC, and that's AM1 with Kabini. Cheap alternative for budget systems and systems you can plop in your living room for everyday stuff. Also you can't fit everything on a single die, it's difficult to do and you are not going to get extreme performance for it. There's simply not enough room or room for high end performance. Besides that's not the point with system on a chip.
APU's are a long way off from ever replacing GPU's even in the mid range segment. And I just don't see how they would get there until they consolidate high bandwidth memory onto the CPU, along with more powerful GPU's in general.
They are, but they will close the gap pretty soon I suspect. Budget builders may opt for a crossfire pair with the 7850K or even 7700K. I myself would love to remove my discrete card but it isn't going to happen yet, no.
But the point is we have seen it working in action and it'll improve over time (FM2+ for now).
AMD simply need the support, get driver support out and push ahead with it.