juggernautxtr :
juan is right and wrong at the same time, yes APU will become the more powerful processing unit. and more cost efficient both in power and production.
yes the apu's will be taking over, AMD will shift all production to an apu, as they get smaller they will soon add cores to both cpu and gpu. and as on kaveri they will work more closely together on same processes, which will increase processing done. fx will not last more than 3-4 more years before they have the apu's just annihilating the separated components.
the only thing i see them continuing is dgpu for gaming, where you will need the dedicated processing power.
the GPU/PPU in kaveri alone is capable of i think i read 3 Tbytes of processing power and nowhere near being utilized fully for what it can do. yeah it's not the greatest for gaming but is a much more powerful processor than being let on at the moment. the software capable of being used on them hasn't hit.
the change is coming guys, and it's going to be a whole new world.
We agree on everything except on dGPU for gaming. You make a very good point, but this is what I think. The principle of locality identified on computing at exascale level also applies to gaming, even if you use the dGPU only for graphics, because the identified 10x barrier of moving data at exascale level remains.
Second, dGPUs for gaming are not designed in isolation. Both AMD and Nvidia reuse designs aimed at compute. Nvidia GeForce cards for gaming are derived from Tesla cards for HPC. Once Nvidia and AMD stop designing and selling cards for compute (both Nvidia and AMD agree that APUs will be much faster), the costs of the new design will be transferred entirely to the gaming cards which, in my opinion, will kill the market. The cards will be so expensive that nobody will purchase them.
This is similar to AMD selling FX Piledriver when the cost of the design was split between Opterons and FX (with most part of the cost being included in the more expensive Opterons). Once AMD stopped developing Steamroller Opterons it also stopped developing Steamroller FX.
The people who is signing the petition to AMD to develop FX Steamroller seems unaware of these questions. Even if AMD did accept the petition and could design a new FX chip in less than one year (which is unlikely) all the cost of the design would be transfered to the FX Steamroller chips, which couldn't be sold so cheap as the Piledriver FX were.
Third, dGPUs for gaming are also attacked by the low end side. Current APUs are killing the low-end dGPUs. When APUs achieve enough performance for most people, dGPUs for gaming will be killed because the enthusiast market will be so small that the development cost of anew design will be prohibitive. I already gave before the link with Intel plans to kill gaming dGPUs for 2015 or so.
Cazalan :
It's a good summation Juanrga but I think people here are more concerned (and rightfully so) with what they can buy today or tomorrow for a desktop PC. Those future ultra high end APUs are going to continue to cost a fortune for the next 3-4 years. Those Xeon Phi CPUs are in the $3500-$5000 range. They will be cool but geared towards HPC not gaming.
People want to know what they'll be buying for BF5/BF6/BF7. If they're running a 290x now they'll still likely be buying a discrete for the next 2-4 years.
There's no magic here when it comes to APUs. It's all about transistor counts, density, yields and process nodes. TSMC/GF are just now getting their teeth wet with FinFET. Intel stumbled a bit at the start on their FinFET roll out. They're just now getting the yields high enough to make their massive 15 core server parts with gobs of cache.
We know what NVidia can do with a 7B transistor (Titan), and we know what AMD can do with a 2.4B transistor APU (Kaveri). Neither is going to change drastically without a major die shrink or much higher yields. The 16nm/14nm nodes should make for some interesting parts but how soon those will make it into consumer grade high end APU parts depends on the yields. I've heard rates as low as 10-20% for GF and 20-30% for TSMC for their FinFETs.
The point is that those future ultra high end APUs will be cheaper than a hypothetical CPU+dGPU with same performance. Not even the HPC community (which pays for superexpensive supercomputers) could afford the costs of a hypothetical CPU+dGPU with same performance. This is why AMD, Nvidia, and Intel are designing APUs for those supercomputers. Check the link that I gave above where AMD reports its plan to use a 10TFLOP APU to build the fastest supercomputer.
There is no plans to build it using a traditional CPU + GPU architecture because it is impossible, that is why I call them "hypothetical". Nobody will build them.
I agree that people will continue purchasing GPUs before they are killed. Intel claims will kill them by year 2015 or so. I think 2018 is more realistic.
Your point about transistors was replied before. It is not magic, it is what the laws of physics say. I already explained what the laws say and why the discrete card will be slower. That is why Intel, is migrating from discrete card to socket because the socket version will be offer superior performance. Intel admitted it at SC13.
I also mentioned that there is consensus among all the scientists/engineers on which is the magnitude of the performance wall at exascale level: 10x. That is why Intel, AMD and Nvidia are developing the fastest systems (1000x faster than current supercomputers) around APUs. A hypothetical 1000W CPU+dGPU would be slower than the 300W APU.
I will not spend more time on this discussion, because it is being overly-repetitive. If people here continue believing that a dGPU will be always faster,* they will be shocked about 2018 when Nvidia and AMD will release their future APUs and none dGPU (I repeat they are not designing any dGPU for that level of performance).
* This is similar to when discussing with someone who only studied Newtonian physics and believes that by adding energy an object in motion will be always faster. Newtonian laws are only an approximation to more general laws of physics and those more general laws predict a c-wall. At ordinary performance level of current desktops a high-end dGPU is fastest than an APU. At exascale level the APU is fastest than any dGPU due to the existence of a performance wall.