Discussion Socketed GPU solution finally?

[vov4ik_il]

I am curious, at what point one of the mainstream vendors will introduce a socketed GPU+Video Memory solution.
Every graphics card we buy has a set of input/output interfaces (display interface and PCIe) and VRM circuitry.
Things that change are video memory, memory <-> GPU interface and the GPU itself.
We have seen video and video memory on a single chip before. Even on recent products like 8809G.
It would make sense to be able to reuse the IO and the VRM and only replace the updated part, bringing down manufacturing and upgrading cost.
It can start as a card and later move towards a motherboard saddle allowing transition.
I think the first one to introduce it will pave a road to success in the long run.

 

[beers]

Eh, I think it would be hard to reach roi from all of the retooling and similar, plus a potentially convoluted support matrix for each parent card, I don’t see this being a thing at least for a while.

to abstract the socket would be bulky and not really feasible from a case volume perspective

 

[vov4ik_il]

@beers

• While investment has to be there and return would be questionable on the first wave, a solid carrier card will pay itself off on the very next iteration, saving a bunch on reference board design and prototyping. ROI is not immediate but is not too far away either.
• Convoluted support matrix is dependent on how good and future-proof the initial design would be. Those “carrier cards” should be only containing GPU supporting circuitry and would be priced accordingly. It is a step to allow transition to motherboard gpu sockets or additional gpu saddles in multi-gpu systems. Ultimately a socketed gpu on a motherboard will make a smaller system footprint possible too.
• It will open a door for standardized (and reusable) cooling solutions too.

 

[Wolfshadw]

Not sure current motherboard designs could handle the power requirements of the higher-end cards. Otherwise, why have PCIE power connectors on current graphic cards? Why not just pump 375 watts through the PCI-Ex16 expansion slot?

-Wolf sends

 

[vov4ik_il]

Why not just pump 375 watts through the PCI-Ex16 expansion slot?

Not sure what this has to do with the offered concept, but current motherboards already offer 1200A VRM solutions which are 1.2KW or more given vCore is 1v or more.

Nothing prevents having power connectors on the carrier card and having dedicated VRM by the GPU socket if it is on the motherboard itself.

There is a bunch of graphics board real estate being allocated to provide interface to the memory ICs, all that will become available.