No, APU is a value solution rather than a professional solution, focusing on the cost rather than the performance, so your solution tends to be meaningless. In Sony PlayStation 4 (Pro) the system memory is replaced with GDDR5, so your idea is a cloned one!
Lol, looking at your pictures, you seem to be suggesting placing your discrete GPu into the DDR3 ram sockets.. I wouldn't do that my friend. EVER!
Lord knows what will happen.
Pins have different assignments and positions for especially voltages, not just data for a start, doing that you would likely blow your system up, secondly, it won't fit, Thirdly, it's barking madness.
But keep at it, your thinking outside the square, you seem interested, I'd suggest you study hardware for a while, different buses, and take an IT coarse or three and keep up the interest, it's a fun hobby.
janagewen ...mmm okay, just didnt know, that something similar using in consoles...
ifixit thank you for answer. "bad idea"
There's two main reason for this to be not a good idea
1- The first is already answered by janagewen, APU's are mainly aimed as a value solution, adding GDDR5 Memory will be costly.
2- Even If we used the already exist resources, like the PCIe x16 slot, this slot is very slow compared to System RAM, the fastest PCIe v4 x16 slot provides almost 32GB/s, compare this to simple dual channel DDR4-2133, these can reach 34GB/s, knowing that current APU's support faster than that, and next gen APU's will support faster memory also... PCIe will be the bottleneck in such configuration, not benefiting from GDDR5 not to mention that GDDR5 are more expensive than DDR4.
You need to integrate the memory controller directly in the APU to gain any noticeable speed, this will mean more pins in the socket and more cost with that
AMD actually have some ideas in this regard, it's rumoured that they're planning an APU with built in HBM2 memory, while the first rumour actually hinted this as an APU for HPC applications, we don't know if such concept will be available for consumer, it will be interesting idea for high-performance SFF systems and also high-end laptops where there's no space to add a discrete GPU, so they can integrate a mid range GPU ( like mobile RX480 equivalent ) with 4C/8T Ryzen APU along with 8GB HBM2 in the same die with 45-60W TDP target
The idea seems sound on some levels of motherboard design.Having the GPU directly linked to the other GPU would double its performance but it still needs to be translated an updated for continuous use.So a ROM or with preprogrammed A.I with different driver types in between the two would give it an edge.
It could draw data an filter instructions clearer,as it would have fast draw from the ram directly too the CPU.This would be more constant.I am talking about Image 4. by the way.
Sorry had to reedit this.