This was in part an insurance policy against AMD in case their competing Southern Islands architecture's performance outshone that of Kepler. Had that happened and had manufacturing not yielded so many duds we may have seen a prioritised GK-110 based GTX 680 (potentially with some parts disabled). However the opposite situation occurred; Kepler is faster blow-for-blow than AMD's architecture and has low production yields, making a giant-chipped GTX-680 both unnecessary and too expensive.
The effect has been for GK-104 to be prioritised and recast as both Nvidia's top part and the basis of their entire enthusiast range. The chip is currently found 'whole' in the GTX 690 and GTX 680 and, with one segment disabled after manufacture, powering the GTX 670. Further 'harvested' cores should also power the upcoming GTX 660. Rounding out the rest of the 600 series is the GK-107, designed to be basically a quarter of a GK-104. This has popped up as a mobile chip and is slated for entry-level desktop parts such as the Geforce GT 630 and 640. A few sneakily rebadged previous-gen Fermi parts will play supporting roles.
Meanwhile the giant GK-110 has been shifted further down the line, when yields are greater and the extra speed is needed, we suspect to become the basis for Nvidia's 700 series. Performance of a single-GPU GK-110 card should rival the GTX 690. On top of the murmurings doing their murmuring in tech-land it ties up logically. A 7.1bn transistor chip is about the most that can fitted onto the 28nm - a process that will be used for next generation or two - without drastic heat issues. Higher production yields in the future should also enable Nvidia will be able to produce it for only a little more than they produce their 3.5bn transistors today.
Timing wise, availability for the 'K20' supercomputer part announced is expected around the end of this year. Given such server-grade parts are the cream of the crop, this implies the 7.1 bn transistor based GeForce card will arrive before then. Given the server model is close to finished, it would make sense for Nvidia to aim for a GeForce part to be available pre-Christmas this year. Expect AMD to respond in kind. All in all, this reminds us of a scene (which we'll have to paraphrase a lot) in Third Rock from the Sun where doctors dramatically receive yet another critical trauma patient entering their Emergency Room:
"Dear Lord, when will... when will they stop coming? When will it all end?!?"
"Let's hope... let's hope never?"
"God you're right, onwards!"
Cet article est dédié à ma grand-mère. Bonne vovyage ma chère.