Nvidia has serious yield problems with the GT200.

[highoctane]

It seems that the tables have turned this time around. Nvidia does not produce enough GT200s to meet demand; yields are terrible. Not all second tier AIC (add in card) companies can actually get their hands on the GTX 260 and GTX 280 cards, forcing them to switch to the ATI brand or skip this round and waith for the GT200 successor (die shrink and/or simplification?) that should appear within this quarter (Q3).

With 1.4b transistors on 65nm, the GT200 looks to me like a rushed chip in an effort to keep up with ATI.

Source: Me

Send the 2nd teir AIC companies to Neweggs source since they seem to be loaded with gtx280's. Personally I would say demand or lack of demand is more an issue than supply at the moment.

We all know cards like the 8800gt's sold like hotcakes when they where released, it just hit the right performance/price ratio to be successful considering what the competition was at the time.

Now the tables have turned, ATI has finally delivered something very solid after two product cycles.

As for 1.4b transistor chips, I don't think that is something you can "rush" by any means and poor yields would seem to be directly related to TSMC not yet being able to reliably produce the fully operational dies.

Nvidia is still ahead of ATI so I wouldn't say they're keeping up as far as performance is concerned but their performance per $$$ needs a little work in light of what ATI is delivering in that regard.

EDIT: Is there any solid facts about yields and availability that are verifiable or is this thread another journey into the Xtremetheory/speculation zone.

[[XC] gomeler]

Always funny how if AMD/ATI starts doing well again the AMD/ATI fanboys come out of the closet en mass and everything turns into a nVidia bashing fest.



I wouldn't at all mind having a ATI card for my next purchase, but I just think it's absolutely hilarious how this always happens.

And no don't start with the it's nVidia's fault, you and me would have done everything they did as well, ie releasing all those G80 variants, just milking the cash cow, that's why they still excist, they know how to play the game and are an excellent company.

It has been an Nvidia bash-fest since G92 and the huge ripoff it was compared to G80 That being said, it's been an ATI bash-fest with R600 but RV670 kind of turned that around and RV770 is the current poster-child GPU

[alexio]

How has it scared away Nvidia's partners? I saw Gainward is now moving over to the red camp, but it's not as if they left green, they're just putting their feet in both doors. Besides, I think that says alot more about the 4000 series than any failure from the g92. g92 was and still is a great chip.

It is still a very good chip, I'd say it is about as impressive as the RV770. The problem here is not the performance, but rather the marketing.

Nvidia pushed the 9800 series to sell as many GPUs as possible (this is just normal business). The problem however is that their board partners still needed to sell all those G92 8800 cards. You and I know the difference between the two series is quite small, but 50% of the potential buyers don't. As a result the price of the 8800 cards drops to a very low point. But Nvidia wants to sell even more G92 GPUs, so they launch the 9800GX2. The result is that many people buy a 9800GX2 instead of 2 seperate cards, now the board partners have a buttload of 8800 and 9800 cards. Now Nvidia pushes the G92+ forwards making it harder for their partners to sell their normal 9800GTXs. The low-margin 9800GX2 is just becoming absolete throught the launch of the new GT200 based cards, so these cards have to be sold cheaply.

My point here is that this is indeed business. Nvidia is indeed doing great business. However Nvidia finally understands that they have screwed their partners, so the weird pricing agreement comes in effect to save their smaller/weaker partners.

Makes sense?

[Hornet331]

Send the 2nd teir AIC companies to Neweggs source since they seem to be loaded with gtx280's. Personally I would say demand or lack of demand is more an issue than supply at the moment.

We all know cards like the 8800gt's sold like hotcakes when they where released, it just hit the right performance/price ratio to be successful considering what the competition was at the time.

Now the tables have turned, ATI has finally delivered something very solid after two product cycles.

As for 1.4b transistor chips, I don't think that is something you can "rush" by any means and poor yields would seem to be directly related to TSMC not yet being able to reliably produce the fully operational dies.

Nvidia is still ahead of ATI so I wouldn't say they're keeping up as far as performance is concerned but their performance per $$$ needs a little work in light of what ATI is delivering in that regard.

very good post.

[ownage]

??

That is a very strange way of calculating die cost.

I get it to 5000/(0.6*94) = $88.65 per die

Depends on the way you look at it. Cores should be 53 dollar, but because of the failure rate it costs almost 89 dollar. These are raw prices of course. There are more costs you should add to a single die.

Is it really, if it scares away your partners?

I'm curious about how dependent AIB's are on nVidia. I know from some AIB's like EVGA and XFX that they work very closely with nVidia. Both brands have step-up programs which are only available because nVidia supports EVGA and XFX with this. ASK Technology Limited, where Inno3d is a part from, would be bankrupt if nVidia would collect his debts. From what i heard nVidia wants more small AIB's instead of being dependent on a handful of big AIB's. I don't see why small AIB's would choose for nVidia for the high-end products.

[nn_step]

well it is better yields than when IBM started making POWER processors

[Warboy]

Weee! rumor mill thread, Best thread evar. NOT!

Every company now days have yield issues in the news atleast.
Intel, nvidia, AMD, ATi.

AMD and it's Phenom.
Intel and their Penryn.
Nvidia and their G80 / GT200.
ATi and their past 2 product cycles.

God damn people. It's called a Revision. There is a meaning behind that.

[savantu]

afaik no yield rate ever was publicated by any big semi conductor company (at least not when the product was still selling).

So forgive me my skeptic. There where a lot of BS rumors about intel, how utter horrible there 45nm yield were, same goes for 65nm when they where first, same for amd etc.

http://download.intel.com/technology...Technology.pdf
Some nitpicks :

Well , Intel recently posted info about its 45nm process.Systematic cross wafer ring oscillator frequency variation slightly lower in 45 nm while random cross wafer variation reduced by about 50%. Mean cross die Vt variation reduced from 20 mV for 65 nm to 11 mV for 45 nm for NFETs and from 9 mV to 7 mV for PFETs.

Put it simply : the process rocks compared to their own 65nm.High-k and metal aren't the only "goodies" put in the mix.

The "horrible yield" card is played far to offen in the recent days, but we'll see when NV brings it 55nm shrink of the G200.

IBM had less than 20% yeilds on Cell , 200mm^2 logic only.G200 is 450mm^2 and logic only.I'd be surprised if TSMC yields on it are better than 20%.

Fortunately , I think they can save most parts and sell then as inferior chips , with disabled paths.

[MAS]

Always funny how if AMD/ATI starts doing well again the AMD/ATI fanboys come out of the closet en mass and everything turns into a nVidia bashing fest.

Always funny how if Nvidia starts making monstrous videochips again the NVidia fanboys keep on crying Geforces have a great performance advantage over competition. 10 per cent difference (between 4870 and GTX280) is undoubtely a GREAT Nvidia success

[GoldiClocks]

The low-margin 9800GX2 is just becoming absolete throught the launch of the new GT200 based cards, so these cards have to be sold cheaply.

Sorry, just can't resist it, if we are talking about graphics card that have gone obsolete overnight and thus need to be now sold cheaply, someones gotta mention ATI's 38xx cards too.
Seriously, I really wanna know who wants to buy them now, and for what price?

[Aberration]

Yeild has more to do with process than design.

[Wiggy McShades]

http://download.intel.com/technology...Technology.pdf
Some nitpicks :



Put it simply : the process rocks compared to their own 65nm.High-k and metal aren't the only "goodies" put in the mix.


IBM had less than 20% yeilds on Cell , 200mm^2 logic only.G200 is 450mm^2 and logic only.I'd be surprised if TSMC yields on it are better than 20%.

Fortunately , I think they can save most parts and sell then as inferior chips , with disabled paths.

so just because ibm has low yield on Cell that means TSCM has low yield on the gt200? they are two totally different chips being produced by two totally different fabs. Im sure there are bad yield rates but your just throwing guesses around. You have no idea what the yield rates are.

[s1nykuL]

Sorry, just can't resist it, if we are talking about graphics card that have gone obsolete overnight and thus need to be now sold cheaply, someones gotta mention ATI's 38xx cards too.
Seriously, I really wanna know who wants to buy them now, and for what price?

38xx Obsolete? By what standards? inviable at current price point yes, but obsolete no. All previous cards by Nvidia have been superseded by Nvidia products so by your definition anything but a GTX260 or GTX280 from the Nvidia camp is also obsolete. Though perhaps not quite so obsolete with GTX2xx being so ridiculously expensive.



I wish I could find the links, I tried. Sorry I couldn't. I did read that the GTX280 was costing around $100-$110 per GPU to produce and that around 97 of these would fit on a 300mm wafer. Wafer costs are dependent on process size and the number of mask layers.

[Clint]

Depends on the way you look at it. Cores should be 53 dollar, but because of the failure rate it costs almost 89 dollar. These are raw prices of course. There are more costs you should add to a single die.

Well, that was what I said..

FYI, a 40% failure rate out of 94 gives 56 dies left for sale..not 37

Ah well, splitting hairs here...

[Scubar]

Would explain why Gainward have dropped the nVidia only cards and are now going to be producing ATi aswell. Id definitely be interested in seeing what Gainward come up with for their Golden Sample cards. They have their own cooling methods which are far more effective than the standard ones.

[FischOderAal]

Overconfidence?

Or arrogance

[Starscream]

So partners cant get enough GPUs?

Isnt it usual with Nvidia that the first batch of cards is made by 1 OEM appointed by Nvidia. This OEM ships the cards tot he partners with cooler on it.

could be that this OEM is having problems or ofc that this si all BS made up by someone cause the stores i visited today dont have a problem with 360 and 280 supply.

And that Nvidia only partners are moving to AMD is normal cause it increases their potential sales.
Asus and MSI did the same after the Geforce FX fiasco and in return Nvidia punished them by not supplying them as many Geforce 6 GPus as they wanted.

[ownage]

Well, that was what I said..

FYI, a 40% failure rate out of 94 gives 56 dies left for sale..not 37

Ah well, splitting hairs here...

Sorry. I was talking about the costs of the failed dies not the ones for sale.

Yeild has more to do with process than design.

Of course but design is also important. Alexio explained this to me last week.
G70 was identical to G71 but G71 had less transistors. G70 had lots of back-up transistors so if important transistors would fail, the process would be taken over by the backup transistor. G71 had less backup transistors because they where familiar with the technique and expected good yields.
If GT200 is already big, and they don't want more they use less backup transistors.
Produce large dies is not easy but if a core doesn't have much backup transistors you experience more failed dies. If a core has failed transistors, but the tasks can be taken over by the backup transistor then the core still works.

[zerazax]

See this wafer of RV770 chips?


Assuming that RV770 is indeed 16 x 16 = 256mm^2, then GT200 at 24 x 24 = 576mm^2 will mean for every 3 by 3 block of 9 RV770 dice, GT200 can only have 4. So in the best case where both have the same yields, Nvidia will still be at a losing 2:1 ratio

[YukonTrooper]

With G200 being so overpriced I'm surprised there is any 'demand' at all. :p

Because it's "The way it's meant to be played."

[Devious]

With G200 being so overpriced I'm surprised there is any 'demand' at all. :p

[Clint]

Sorry. I was talking about the costs of the failed dies not the ones for sale.

Ahh, now when you say it...sorry for that.

In any way It feels good in some perverted way that Nvidia are going to get it a bit harder in near future. Go ATI!

[Indes]

I'm confused, how come every company who has the performance drops almost all R&D and decides they don't need to make better products. Seriously nvidia had enough time to make a killer card, we've seen it from intel/nv/amd and ati. I must have missed the memo.

[Shintai]

IBM had less than 20% yeilds on Cell , 200mm^2 logic only.G200 is 450mm^2 and logic only.I'd be surprised if TSMC yields on it are better than 20%.

Fortunately , I think they can save most parts and sell then as inferior chips , with disabled paths.

Cell yield sucked because they wanted app SPEs etc to work And its very logic heavy. With a GPU you can make sure there is 260-280 shaders if you only need 240 etc to increase yield multifold. And then others can be sold as 260GTX (Or even something lower). Cell was either or nothing.

[Shintai]

Well you can only fit around 90-100 of those dies on a wafer....
Has anyone got any idea what a single wafer costs for the likes of NVIDIA? $5000? $10000?

Alot lot less.

http://www.digitimes.com/NewsShow/Ar...ages=A5&seq=19

I think a CPU grade 12inch wafer cost about 3000$ when "ready". TSMC is much cheaper. Perhaps 1500-2000$.