I'm more of an enthusiast than an engineer when it comes to the technical specifics with gpu architecture/production, so I'm asking you "Veterans" here. Seeing this article makes me wonder two things that I'd like everyone's opinion on:
1. Would AMD/ATI stop using TSMC completely or would they just have one or the other company make up for a lack of supply (like what we saw earlier this year) ?
2. If they indeed do switch to GF for all chips, would there be a big enough difference in fab process quality to possibly give ATI a(n) advantage/disadvantage compared to Nvidia, assuming they stick with TSMC?
Fab processes are different. Sometimes extremely dramatically. Even if TSMC and GF were on the same X nm process (both at 28nm, both at 22nm, whatever) ATI's design would have to be significantly reworked to fab chips at both. It is unlikely they will do this sort of double-designing just to have the same chips fabbed at both.
What is significantly more likely, that if they do use both companies, one chip will be made at GF and another could be made at TSMC. (ie. RV970 chips made at GF while RV930 chips are made at TSMC.) But that's not very likely, either (because a large part of the chips is "the same." Shaders are the same whichever chips they're in, and would have to be reworked for the other process).
If ATI switches, they will probably switch all new production (RV8xx will all be fabbed at TSMC, RV9xx will all be fabbed at GF).
Global Foundries process tech. is very likely significantly better than TSMC's. GF comes off of AMDs process tech, which has been leading edge (just behind Intel) for many years, whereas TSMC does/did a lot more business in higher processes. GF had to be top-notch to compete with Intel, whereas TSMC just had to compete with other, less performance necessary fabs like Charter).
2) Yes that's entirely possible, but if that's the case NVIDIA could be forced to use Global Foundries if GF yeilds are much higher than TSMC.
A couple of things to consider going forward is that TSMC will be using a gate-last approach going forward while GF would be using a gate-first. That means that if you design a 28nm chip for one fab, you have to redo it for the other as the two processes are not compatible. Since the two processes are different it could be that one makes more sense for bigger faster chips, while the other makes more sense for smaller slower chips. It could then be that you have higher end chips at one foundry and another line of chips at the second one. Just depends on what makes financial sense at the time.