Workarounds wouldn't have to be all that difficult. It's not like GPU HSFs are the most elegant solution imagineable.
Where is it going to go?
You can ship it out a couple of slots like a 1900XT or out the case door like the Lian Li 60plus if you want to use existing kinds of solutions. It would be possible to push it towards a blowhole. Problem is, to come to any CPU HSF kind of solution, you'd either need to target the mobo layout to a specific kind of case design or get a case maker to work with you. That's not been done before that I know of, but moving hot air around a metal box is simple tech.
So sell the DIMM with the socketed GPU or put the memory on-die.
Its going to be tough to have a 256bit bus with DIMM slots, on-die or substrate.
OK, difficult. Do you think either is possible?
What if AMD has bought into Physix in a big way and are looking to have a more efficient means of CPU to GPU to PCP communication?
Its not a bottleneck on AGP so whats there to do?
That's part of my question. If for example Physix is a focus, is a multi-socket approach part of the solution? I'm not claiming to have the answers, just wondering if we might be mising something. Sure, the talk may be baseless rumors, but what if it isn't. Just brainstorming, that's all.
If graphics memory capacity or speed is decoupled from the GPU
????[/quote]
Meaning that if you can update graphics memory independantly of GPU or vice versa, as new tech arrives, that's a potential performance advantage. Look at it another way - if one presupposes that game programmers will continue their quest for more realistic games, what GPU, memory, coprocessor, etc., changes will be needed to keep up? How close are we to AGP/PCI-e-based designs from hitting the wall? Are there multisocket approaches that will help for projected needs? For my own personal needs, current hardware and software are almost there, but that doesn't mean that the industry will stop progressing.