NVIDIA has a problem. It's got a bad chip launch on its hands: it's hot and uses a lot of power, it's huge and yields suck, and it hasn't nearly hit performance targets. Worst of all, they're stuck with that for a whole year before the option for a die shrink shows up to fix all their woes.
TSMC has been promising that 28nm would be well ready by this year. But as reported by The Inq and a very good news article by Xbit Labs, they might barely make it out by the end of this year, with GPU-suitable versions only being ready by early next year. Also mentioned in the Xbit article (and here as well), 32nm has been scrapped, so that leaves NVIDIA no other alternative but to wait, and thus so must we.
Both NVIDIA and AMD have already been burned by the problems with their 40nm production. As explained in the excellent article by AnandTech, AMD managed to stave off much of the issues they might've otherwise had with the chip itself using some clever tactics, but they were/are still not immune to the issues of yields, whereas for NVIDIA that's just one of their problems. The last time NVIDIA had an issue like this was (and the comparisons continue) the FX series, and they reacted by pulling production for their next generation and giving it to IBM. But IBM isn't an option anymore, so despite all the problems having to depend on a single manufacturer, NVIDIA and AMD have no choice but to suck it up, right?
Well as much as TSMC would like to think so, this latest major foul-up has cost them dearly in the technology race, leaving them open to competition by an emerging up-and-comer called GlobalFoundries. Not only do they appear to be caught up in terms of schedule, their particular 28nm process seems to be superior to TSMC's. Obviously a company jointly-owned by their chief rival makes NVIDIA a little skittish at the prospect of joining under their wing, but for that same reason that means AMD is primed to start utilizing their own facilities at the nearest opportunity, and I imagine then that we might start seeing budget and mainstream 6000 derivatives churning out of a brand new GPU fabricator as early as Q2 2011.
Of course the more AMD starts pulling their GPU production in-house, the more that will leave NVIDIA greater share of TSMC's focus and capacity. So the incentive is certainly there for NVIDIA to stay where they are, but there are plenty of reasons not to as well, and if GF100 is going teach them anything, it's that it's good to leave your options open. A little sweat on TSMC's brow probably won't hurt NVIDIA's bottom line any.
In short, this is the worst possible time for TSMC to be having problems. On top of that, chip designers probably don't enjoy being lied to, especially if it causes them to completely reshuffle their plans because certain roadmaps turned out to be a little less accurate than they were claimed. No one gets hurt worse than NVIDIA this generation though, and while their trial-by-fire experiences might help them learn the 40nm process as well as anyone in time for a formidable mainstream Fermi derivative, the damage is already done. They and the rest of the market are ready for a new kid on the block, and he might grow to be bigger than anything that has come before.
An R600-based architecture with a shader clock might be interesting to see, though.
