The GT300/Fermi Thread - Part 2!

[saaya]

Oh man, that was a good post! I read the whole thing, and I totally agree. $600 for a card that's 30% faster than the 5870 would be a good price, and with any kind of decent yield, it will be there. nVidia might have to settle for small margins with the first series of fermi based cards, but they will sell ok -as long as they perform 20% or more faster than the 5870. I'm sure there will be benchmarks and games where the difference is 50% or more, and there will be games where ATI actually outperforms it.

thx, i wouldnt be surprised if somebody mentions that its completely off
still, its interesting to look at those numbers and juggle them around, even if my yield numbers are off, the price and performance numbers arent, and those actually say a lot... those numbers are a fact, and they are rules of the real world nvidia needs to follow as well... so when they think about where to position gf100, or probably thought about it months ago, they must have been coming to at least similar conclusions...

The only thing i would argue is that ATI is better at getting good yields than nVidia, so even with the same defect rate, just based on the GPU design, ATI will have a bigger gap than you figured - just a small detail If nVidia can get to 50% yield, that would help a lot though!

i think nvidia doesnt even need 50% combined yields, i think if they get 40% combined yields thats good enough for them to make money even if the performance isnt that great... the funny thing about rv870 is that it really shines at bad yields and i heard of 2 tricks ati did to improve yields and even turn rv870 chips that have a defect into a full blown 5870.
ati doesnt need good yields to make money with rv870, and actually they dont benefit that much from improved yields as nvidia...
like i wrote, as soon as 40nm has OK yields they already get more full blown rv870 chips than they probably actually need to fill in demand... maybe that was their plan, to refresh rv870 and instead of binning parts based on defects they will then bin them on properties ie higher clocks. makes sense...

--Don't forget though, ATI is allegedly prepping an optimized 5870 with higher speeds, much like the 4890 from the 4870. This will be ~15-20% faster than the 5870 and fermi will need to compete with that card, not just the current model.

i heard they actually dont...
the yields are good enough for them to go for a rv890 now with 2000sps, hopefully bigger caches and most importantly more l2 cache bandwidth... but then again... they would be close to gf100... and does ati really want that? wouldnt it make more sense to focus on 28nm already? then again, its still 6-12 months away and atis strategy is to start with an entry level part and not highend... so a highend 28nm part is 12+ months away, possibly 18 months... and they need something to refresh the performance segment...

thats the one thing nvidia has worked out so well, the best business is offering the best, by being slightly faster than everybody else, but charging a nice premium for it... thats how you can make the most money...
so that would mean an rv890 would make a lot of sense, especially since a 28nm highend refresh is still a year or so away... and in a few months 40nm yields should be good enough to make a fermi sized chip with ok yields... and if its based on rv870 then they can even sell chips with several defects as 5870, 5850 or even 5830 and do a lot of recycling

NP
I mean, supply and demand is the major driving force for deciding the price of GPU, and everything else too.

Bad yield means more expensive manufacturing cost, but who care about nVidias cost? I have never asked for the cost before buying a GPU, do you?, nobody else does.

nVidia can spend $8000 on a chip, but nobody would give it to them, well that's if ATi could deliver a better performing GPU for less.

you should care about it, because if you know that a part costs xxx to make, you know that its price wont drop a lot, or will probably drop a lot when a new competitive product comes out etc...
if you know how much a part costs nvidia you know the minimum price it will cost you to buy it... its very very unlikely that nvidia sells parts below cost, and if they do, only in very very small volume.

and i agree, if supply of gf100 will be low, the prices will obviously go up and may be even more than 600$...

i have a feeling that nvidia is telling press and analysts that if prices of fermi are too high, that only means its so great and so awesome that everybody wants to buy one, which drives demand up and hence pushes prices up...
but i dont think anybody will fall for it... everybody knows that if demand outstrips supply thats actually bad for a manufacturers as it means they could sell even more and make even more money...