Quote:Well, 70% more expensive than a 7970 would have been a launch price of ~$935. Of course, I know those numbers aren't serious. We'd probably be looking at something like ~30-40% more performance for a GK100 680, and ~15-20% more performance for a GK100 based 670. The thing is, x70 cards of the past have traditionally competed with AMD's high end offering (going back to Fermi that is). A GK100 680 would definitely cost a lot more than the 7970, but a hypothetical GK100 670 would probably aim to compete with the 7970 on price, at least initially.
Regardless of what Nvidia would have priced things had GK100 made its way to the market, the performance per dollar curve would still be a in a better state.
The second is what you're saying about the current cards being somehow 'not worth it'. A quick look at this chart confirms that GTX670/680 and 7950/7970 sit nuzzled in perfectly with where we would expect them to be on this chart (e.g they're right around comparable cards from previous generations). When you consider how much less power the 670/680 use vs. Fermi and how much quieter they run per FPS, there's really nothing out of line about Kepler's value (for Gaming at least) at all:
The issue with this is that these chips are significantly smaller than their last generation counterparts. Yes, these next generation cards are better value, but only barely so, without looking at power consumption and increased overclocking headroom. If I had to choose between getting a current generation and last generation card, I'd easily pick this generation, but the prices are still not interesting.
Let's look at GK104. I can't find the exact die dimensions without looking harder than I'd like to put the effort into right now, but according to this calculator here, we're looking at roughly 200 dies per wafer if we assume the die is a perfect square (17.1mm*17.1mm). GF110 is 520mm², or 22.8*22.8 (once again assuming it's a perfect square out of laziness), and around 100 fit on a wafer. Obviously, 28nm yields are going to be worse than 40nm, and we're also looking at higher wafer costs, but that is simply a lot more dies per wafer -- roughly double. Smaller dies yield better than larger dies as well.
Right now, the GTX 680 is retailing for $500 MSRP. The 580 was launched at $500 as well. At release, the 680 was only 20% faster than the 580. So despite being able to cram twice as many dies on a wafer, we're paying the same amount for 20% more performance and ½GB more of vRAM. It's a step forward, but knowing that there's so much more room for significant price drops in the future makes this generation a pass so far.
Not to pick on only Nvidia... AMD is no less guilty.