Thank you for your earlier response.
I am wondering if the tech is not remotely ready by E3 2013 then would AMD opt (on Sony's behest) to emulate the future tech via more conventional hardware setup and if so then what effects would it have on the devs that would have to present the first glimpses of their games during the début of the hardware?
Your post ties into bgassassin's above so I'll answer both at the same time.
The 2014 AMD
Fabric computing memory model with FULL HSA support requires minor changes to the AMD GPU in the APU and probably the second GPU is getting changes also. I've already posted cites that mention this for the APU (In RED) and some of the efficiencies can be applied to the GP GPU outside the APU also so it's logical that it would receive changes also.
The
fabric memory model works with as few as 2 CPUs (X86 and GP GPU) or more, including because of the Fabric model Memory controller in the APU and common address scheme, other CPUs on the memory buss or nodes connected with special Network cards or optical connections. This
was the same vision for the Heterogeneous Cell, the patent for 4 SPUs connected to a memory buss etc.
Memory controller and OS that supports Fabric computing and Heterogeneous CPUs each with enough cache to allow concurrent operation but passing memory pointers rather than data between CPUs. All connected nodes must have a common memory pool (address scheme) or understand how to access (local translation) a common memory pool (it's transparent to applications). Understood in this is a VERY fast memory buss or transport and Very Very fast memory which is now possible with 3D stacked memory.
Memory inside the APU is probably going to be less than 100 meg of 3D stacked DDR3 with wide I/O....VERY fast (faster than XDR2) with a second memory pool outside the APU all with a common address scheme. 2014 and later embedded, handheld and high performance platforms are going to be using 3D stacked memory both inside SOCs and outside not GDDR5 (3D stacked memory will be Cheaper, faster and more energy efficient).
The Second GPU could be connected to the APU with a PCIe BUSS just like PCs do and would require it's own memory or it could be
(Fabric model) connected to a common memory BUSS and pointers could be passed from APU to second GPU with DMA. 2 gigs or more of very fast 3D stacked memory can be on the main board on a common memory buss that has APU and second GPU connected to it. Both models have advantages.
The APU with CPU and GPU takes maximum advantage of the fabric computing model. The second GPU if it has enough cache can do the same. Since video output is from INSIDE the APU, it does not make sense to have video memory in the second GPU, common memory makes more sense.
This is speculation based on hours of reading with clues here and there in articles.
i-LO, with OpenCL yes, developer platforms can have all the features of a final chipset even if missing the hardware as Sony can and probably does provide support using the other CPUs using the same OpenCL routines. Since Developers will be using AMD openCL and HSA libraries (JIT engine in chipset) when HSA efficiencies are available they will automatically be used. So developer platforms will run hotter and cost much more than the final PS4 chipset....this is nothing new (just much easier this generation to create a developer platform due to OpenCL and same AMD chips but of older generation). The features available in the PS4 will also be nothing new as they are available now in top end AMD PC platforms.
AMD needs to kickstart HSA (they are betting the farm on HSA) and NEEDS Sony and Microsoft to use the new chipset and educate a generation of developers in how to use HSA. This is why they made it Open Source and are cooperating with Khronos. HSA and fabric computing is also going to be seen in super computers. Moore, of Moore's Law, recently gave a lecture on this.
Mentioned (Hints) is this HSA and Fabric Computing model supporting super computers and cloud computing. Cloud computing can be the Cell's vision of multiple Consumer electronics devices in the home sharing a common memory address scheme and sharing data, information and CPU loads among the various platforms using the home network. Think about this and the PS4 being designed for a 10 year life....... Then that AMD has said the model scales from handheld to super computer.
bgassassin said:
Hmmm. Still didn't catch a reference to having multiple CPUs even after looking at those articles again. Unless things have changed the target specs said 2GB of GDDR5. I guess they could stack that, but it doesn't sound like what you're proposing is what Sony is pursuing.
What do you consider multiple CPUs; 2 X86 + 1 GPGPU + GPU that may be GPGPU + DSP (some part of AMDs I/O) + FPGA?
Game developers do not have the final hardware, they have what's available now that has similar performance. GDDR5 is the min memory target that's available now and may not be the final memory seen in the PS4. Game developers don't need to know about the 10 year plan for the PS4 and the OS changes and new SDKs that are going to be developed. Just like the PS3 but most likely with fewer delays, the OS will evolve. It will parallel PC development this time as AMD and possibly Microsoft will be supporting it.
I'm reading allot into this and it should be considered my opinion at this time.