This is a rather long post, you dont have to read all the steps i did, just scroll to the bottom and read the conclusion and then scroll up to see how i got those numbers in case you wonder :D
and thx for anybody pointing out mistakes or irregularities! i dont expect those calculations to be 100% correct, if im somewhere in the ballpark or maybe on the parking lot thats still a good enough perspective to peek at where we are, whats going on, and what might happen ^^
cool, thanks for the link :toast:
so they are at 0.1-0.3 defects per square cm...?
fermi is 2.4x2.4cm so that makes it 5.76cm^2
so 0.576 defects per fermi chip, and there are roughly 100 of them on a wafer, so:
0.1=58 defects/wafer
0.2=115 defects/wafer
0.3=172 defects/wafer
assuming the defects spread evenly (they dont, they arent completely random but this is easier to calculate to get a rough idea)
fully functional yields (480+5870)
0.1 = 42 GF100s, 102 rv870s (40 vs 66% yields)
0.2 = 0 GF100s, 45 rv870s (0 vs 33% yields)
0.3 = 0 GF100s, 0 rv870s
thats fully functional chips, if they can disable the defect affected part of the chip and it works fine, thatll be a gtx470 or 5850. lets say 50% of the chips with 1 defect can still be used:
salvaged yields (470+5850)
0.1 = 29 GF100s, 29 rv870s (29% vs 18%)
0.2 = 43 GF100s, 58 rv870s (115 defects for 100 chips = 85 single defect chips) (43% vs 36%)
0.3 = 14 GF100s, 74 rv870s (172 defects for 160 chips = 148 single defect chips, 172 defects for 100 chips = 28 single defect chips) (14% vs 18%)
combined yields:
0.1 = 71 GF100s, 131 rv870s (71 vs 82% yields)
0.2 = 43 GF100s, 103 rv870s (43 vs 64% yields)
0.3 = 14 GF100s, 74 rv870s (14 vs 46% yields)
these numbers are very rough guesses!
i know they arent accurate and might be way off, i still thought its better than to have no numbers whatsoever... how evenyl the defects spread, and how well ati and nvidia can deal with single and multiple defects per chip and how many parts they can still salvage and turn into a partly or maybe even fully functional chip if they used some backup logic is hard to tell, if not impossible since those are some of their biggest secrets...
i didnt expect the difference in combined yields between rv870 and gf100 to change that much depending on defect rates... i thought the chip size difference would result in a more or less fixed delta in yields between the two. for example double the chip size is 1/4 the yields, more or less regardless of defect rate... but that doesnt seem to be the case at all?
since there are only ~100 fermis on a wafer, an average defect rate of 0.2/cm^2 would mean close to no fully functional 512core fermi... that perfectly fits the single digit yield rumors... at 0.2 there are almost no fully functional gf100s and not that many salvaged 470s either. at 0.3 even salvaged parts are scarce because most parts are hit by not only 1 but 2 defects...
so the worse the defect rate, the better rv870 looks yield wise, and the better the defect rate, the closer combined yields of gf100 and rv870 get...
he said that the iphone is a lot more powerful than most people think, does that sound like he had to tweak a lot for efficiency to get it working?
my main point wasnt that he sucks at efficiency, but that his talent is wasted on tweaking efficiency instead of pushing realism in graphics to the next level. lets say hes good at efficiency, then so what... a very efficiently coded game will run at good fps on cr4ppy hardware and uneccesarily high fps on highend hardware. a game that is coded unefficiently but looks awesome and is somewhat playable on highend hardware will always look much better than a game that focusses on efficiency.
now what do you think will create more hype and be a more successful game? a game that looks ok-good and runs on old and slow hardware, or a game that looks really nice but needs expensive highend hardware? if its only about graphics, then the latter will clearly be more successful and hyped and played than the prior.
it could be... look at 5830 and 5850...
and if heat is the main limiting factor, then 480s might need their vcore dropped to fit into the tdp envelope, the 470 has less blocks and can hence run slightly higher vcore and clocks within the same tdp envelope...
yeah but there arent too many people around that doom3 knocked out of their socks... :/
oh yeah, 3.5 was 55nm, right?
40nm are 5k each... i remember reading that somewhere...thx :toast:
95 or 100, no big difference, 5%, but 100 is easier to calculate :D
5k wafer cost means 42% higher wafer cost, and lets take the rough numbers i calculated.
0.1 = 71 GF100s, 131 rv870s (71 vs 82% yields)
0.2 = 43 GF100s, 103 rv870s (43 vs 64% yields)
0.3 = 14 GF100s, 74 rv870s (14 vs 46% yields)
if you put it in a graph you get something like this:
GF100 vs rv870
10 - 45
20 - 50
30 - 55
40 - 60
50 - 70
60 - 75
70 - 90
while gf100 gets more and more functional chips overall, rv870 combined yields dont improve much in quantity, but in quality, they get less damaged 5850 parts and hence have more fully functional part they can use for 5870s. since the demand is much higher for 5850s compared to 5870s, good yields at tsmc actually dont really help ati, as they get more fully functional parts than they need, and their overall yields dont improve much, meaning the price barely drops... rv870 is clearly engineered for bad yields, and if defect rates actually reach a good level, rv870 looses its price advantage.
still, even at the same yields, gf100 will cost notably more than rv870, so the die size advantage results in a constant price advantage.
looking at prices...
assuming 100%=100 fGF100 per wafer
assuming 100%=160 rv870s per wafer
GF100 vs rv870
10 - 45 = 500$ vs 70$
20 - 50 = 250$ vs 63$
30 - 55 = 166$ vs 57$
40 - 60 = 125$ vs 52$
50 - 70 = 100$ vs 45$
60 - 75 = 83$ vs 42$
70 - 90 = 71$ vs 35$
nvidia clearly never expected to get lots of fully functional chips back, just like with GT200 where most of their business is with cut down chips, 260 and 275 cards.
looking at the price difference between gf100 and rv870:
GF100 vs rv870
10 - 45 = 500$ vs 70$ = 430$
20 - 50 = 250$ vs 63$ = 187$
30 - 55 = 166$ vs 57$ = 109$
40 - 60 = 125$ vs 52$ = 73$
50 - 70 = 100$ vs 45$ = 55$
60 - 75 = 83$ vs 42$ = 41$
70 - 90 = 71$ vs 35$ = 36$
even with great yields gf100 costs double the price of an rv870... but thats not that important, the actual price is what matters, or the price difference...
GF100 will need a more complex pcb, bigger heatsinks and more memory chips... but i think those price differences are not that big... lets say 25$ for all of that...
10 - 45 = 500$ vs 70$ = 430$= 455$
20 - 50 = 250$ vs 63$ = 187$= 212$
30 - 55 = 166$ vs 57$ = 109$= 134$
40 - 60 = 125$ vs 52$ = 73$= 98$
50 - 70 = 100$ vs 45$ = 55$= 80$
60 - 75 = 83$ vs 42$ = 41$= 66$
70 - 90 = 71$ vs 35$ = 36$= 71$
how much faster does a vga have to be to cost how much more?
its hard to tell because everybody has a different idea of value and brand and everybody has a certain limitation in their head, both a delta limitation of how much more a card can cost than another comparable one, as well as a fixed limitation of the max amount of money they are willing to spend overall...
lets just look at the current pricing structure and see how its working out right now:
lowest newegg prices + performance from tpu, relative to 5970 in 1920x1200:
http://www.techpowerup.com/reviews/ASUS/EAH5970/31.html
HD5970 2gb 700$ 100%
GTX295 2gb 525$ 81%
HD5870 1gb 400$ 71%
GTX285 1gb 375$ 58%
HD5850 1gb 325$ 62%
HD5830 1gb 250$ 45%*1
GTX275 1gb no stock 54%
GTX260 1gb 225$ 47%
HD5770 1gb 175$ 43%
HD4870 1gb 175$ 44%
HD5750 1gb 150$ 37%
GTS250 1gb 125$ 36%
HD4850 1gb 125$ 35%
*1 3% faster than the 4870 according to:
http://www.techpowerup.com/reviews/A...rectCu/28.html
3% of 44%=1.2%=~45% of a 5970
*2 5% faster than a GTS250 according to:
http://www.techpowerup.com/reviews/P...50_PCS/30.html
5% of 36%=0.9%=37% of a 5970
now lets sort it according to performance and then look at the prices relative to the 5970:
HD5970 2gb 700$ 100% 100%
GTX295 2gb 525$ 81% 75%
HD5870 1gb 400$ 71% 57%
HD5850 1gb 325$ 62% 46%
GTX285 1gb 375$ 58% 53%
GTX260 1gb 225$ 47% 32%
HD5830 1gb 250$ 45%*1 35%
HD5770 1gb 175$ 43% 25%
HD4870 1gb 175$ 44% 25%
HD5750 1gb 150$ 37% 21%
GTS250 1gb 125$ 36% 18%
HD4850 1gb 125$ 35% 18%
there are some interesting conclusions we can make right away...
despite complaints, the 5870 and 5850 are not overpriced and offer a good price perf ratio for highend cards. and despite complaints about the 5970 itself, its price performance ratio compared to the 295 and 285 is pretty good.
the 5770 seems to be the best deal, offering almost half the performance of a 5970 and weighing in at only 1/4 the price, with the 260 and 5830 looking pretty good as well.
lets look at difference between perf and price relative to the 5970 to see how far the price can be pushed away from where it should be according to relative performance:
+ = cost is higher than perf relation = bad deal
- = perf is above price relation = good deal
name - cost - %perf - %price - %/%
HD5970 2gb 700$ 100% 100% 0
GTX295 2gb 525$ 81% 75% -6
HD5870 1gb 400$ 71% 57% -14
HD5850 1gb 325$ 62% 46% -16
GTX285 1gb 375$ 58% 53% -5
GTX260 1gb 225$ 47% 32% -15
HD5830 1gb 250$ 45%*1 35% -10
HD5770 1gb 175$ 43% 25% -18
HD4870 1gb 175$ 44% 25% -17
HD5750 1gb 150$ 37% 21% -16
GTS250 1gb 125$ 36% 18% -18
HD4850 1gb 125$ 35% 18% -17
the lower we go perf wise, the better the deal gets, but keep in mind there is a minimum performance people need or want which makes them less attractive at some point. in general cards seem to be priced to offer around 10-15 points more perf than they cost, both in relation to the fastest card.
a -5 value seems to be acceptable, but already limits demand...
lets say a 0-20 point relation is the playing field. less and you have a show stopper, more and your just stupid, giving away free performance, or your manufacturing costs are so low that you can afford flooding the market with great price perf parts. this happened several times in the past, and its always lots of fun to surf on those flood waves :D
so, realistically, gf100 the highest price fermi COULD charge for is around 0 points, which means just as many percentage points of the fastest cards price as it reaches percentage points of the fastest cards performance.
looking at the yields you can tell that it costs around 200$ more than an rv870 card at bad yields and roughly 100$ more at ok-good yields. if the yields are truly as terrible as some people claim, ie ~10%, its impossible for the card to reach a performance level that would justify that price. lets assume yields are bad-ok or ok-good, ie ~100$ additional costs compared to rv870 or ~200$. lets see how fast the cards need to be make as much money in revenue as rv870 does.
200$ premium per card over rv870
name - cost - %perf - %price - %/%
HD5970 2gb 700$ 100% 100% 0
GF100-U 3gb 600$ 85% 85% 0
GF100 1.5gb 525$ 75% 75% 0
GTX295 2gb 525$ 81% 75% -6
HD5870 1gb 400$ 71% 57% -14
HD5850 1gb 325$ 62% 46% -16
GTX285 1gb 375$ 58% 53% -5
GTX260 1gb 225$ 47% 32% -15
85% vs 71% = 14% 5970 perf points faster = 14/0.71=20%
75% vs 62% = 13% 5970 perf points faster = 13/0.62=21%
so GF100 has to be AT LEAST 20-21% faster than rv870 to cost 200$ more per card...
that means if the yields are bad (~15-30%) then fermi could still sell at least some cards for an "ok" price performance level IF either the 470 or the 480 is ~20% faster than the 5870 or 5850. if GF100 performs even better then it might even turn into a good deal. if a GF100 part can beat an rv870 part by 40% and costs 200$ more, then it would have a %/% relation of 15 points which would make it an excellent deal.
Bad yields (15-30) - good performance (40)
name - cost - %perf - %price - %/%
HD5970 2gb 700$ 100% 100% 0
GF100-U 3gb 600$ 100% 85% 15
GF100 1.5gb 525$ 90% 75% 15
GTX295 2gb 525$ 81% 75% -6
HD5870 1gb 400$ 71% 57% -14
HD5850 1gb 325$ 62% 46% -16
so according to these calculations and speculations(!) even bad yields of 15-30% for GF100 would only hurt its success if its performance would be only 20% faster than rv870 or even less... if its performance is 30-40% faster than rv870 cards it would make it a great price perf offer (same price/perf but higher perf) even if gf100 cards cost 500-600$ (200$ more than rv870 cards)
if performance is only 20% faster, they will be bad price perf cards if they cost 200$ more than rv870...
BUT even if the performance isnt that great, if nvidia can reach yields of 40% or more, the price difference in card cost between rv870 and gf100 cards drops to ~100$ or less. at that point the price performance would be equal to that of rv870 cards, assuming identical margins on the cards.
OK yields (40+) - bad performance (20)
name - cost - %perf - %price - %/%
HD5970 2gb 700$ 100% 100% 0
GTX295 2gb 525$ 81% 75% -6
GF100-U 3gb 500$ 85% 71% 14
GF100 1.5gb 425$ 75% 60% 15
HD5870 1gb 400$ 71% 57% -14
HD5850 1gb 325$ 62% 46% -16
conclusion:
nvidia needs either 40% perf advantage over rv870 at bad yields (15-30%) or at least a 20% perf advantage over rv870 at ok yields (40%+) to be competitive.
thats the GF100 attack window, worse yields than 15% and the card becomes too expensive to possibly offer reasonable performance for the price it costs. less than a 20% perf advantage over rv870 and nvidia would need great yields to make GF100 competitive. oh and before somebody points out that ati could lower their cards costs... yes, they could, and most likely will, at least to some degree, BUT all these calculations are based on the yield relation between gf100 and rv870... its unlikely that there are too big differences in the relation of yields of rv870 and gf100... neither of them will suddenly improve their yields while the other one has the same yields, both SHOULD improve their yields more or less at the same pace. so, IF ati lowers their card prices, they will do so by reducing their margins. i assumed identical margins for both rv870 and gf100 cards, so both the rv870 and gf100 cards prices contain a healthy margin which both of them can and HOPEFULLY will, cut down in a price war this summer as soon as fermi starts shipping in real volume :D
but again, if ati custs their costs 50$, nvidia can do the same and they will still both have identical margins and make the same money per sold card.
pheew, what a post ^^
thx :D