Fudzilla: Cypress yields 60-80% Fermi yields at 20% Fermi 20% faster than Cypress

[SlowSpyder]

Yet another rumor... grain of salt as always.

So if Fermi is being churned out, will it be here sooner then end of March? I don't know how much faith to put in to Fudzilla, but 20% faster is obviously good, but is it good enough given how much later it is and assuming it does indeed use more power and put out more heat? I guess it'll depend on what it sells for.



http://www.fudzilla.com/content/view/17205/1/

Claims that the current TSMC 40nm process that is being used for the AMD/ATI Cypress family is achieving yields of in the neighborhood of 40% don’t seem to have a shred of truth. Our sources close to TSMC claim that actually the 40nm process that is being used for the Cypress family is routinely achieving yields in the 60% to 80% range, which is actually confirmed by the fact that ATI is keeping product in the pipeline and you can actually buy boards at retail.

Chip yields seem to be the topic of discussion lately as the much larger Nvidia Fermi chip is struggling with yield issues as our sources suggest that the actual yields are as low as 20% and are not improving quickly according to sources close to the fabs. With wafer starts costing about $5K the numbers suggest that each chip would cost an astounding estimated $200 per chip which pegs the card with a sticker price of about $600.

Those in the known are claiming that Fermi despite the yield and thermal issues is only about 20% faster than Cypress, while Hemlock smokes it. The combination of low yields, high thermals, and marginally better performance than Cypress could be conspiring to place Nvida in the position of having to release the card, but have to sell it at a loss till they are able to address the issues in the next spin according to sources. Because of the situation a mole we know is suggesting that Nvidia may limit the sales of Fermi to consumers and instead use the chips for the Tesla and Quadro products where prices and margins are much better.

All of the talk of yields seems to be in some ways nothing more than a smoke screen to shine the light off of the current situation to stall sales by promising something that most consumers are likely going to be unable to buy. The moles claim you will not need a pair of 3D glasses to watch this shake out in the next few weeks.

 

[crisium]

A good improvement from 1.7%!

 

[Daedalus685]

Interesting read.. We will see I suppose. Not often we see such a scathing rumour from fud about nvidia.. I suppose it wasn't fuad that wrote it though.

20% faster seems about in line with the 280 vs 4870.. Could get dicey if this has any truth and a 5890 is out in line with the 380.

Regardless though.. fermi is going to cost a damn lot... No way they could make R&D costs back in consumer sales. Tesla will have to be popular.

 

[T2k]

Ohhhohohohooooo:

Those in the known are claiming that Fermi despite the yield and thermal issues is only about 20% faster than Cypress, while Hemlock smokes it. The combination of low yields, high thermals, and marginally better performance than Cypress could be conspiring to place Nvida in the position of having to release the card, but have to sell it at a loss till they are able to address the issues in the next spin according to sources. Because of the situation a mole we know is suggesting that Nvidia may limit the sales of Fermi to consumers and instead use the chips for the Tesla and Quadro products where prices and margins are much better.

SO Charlie was AGAIN right, after all, about the need for ANOTHER spin?

 

[PingviN]

"Written by David Stellmack "

20% better for being really, really late and giving poor yields? If that's true, Fermi is in for a rough ride.

 

[Daedalus685]

Ohhhohohohooooo:



SO Charlie was AGAIN right, after all, about the need for ANOTHER spin?
[/B][/I]

They are talking about a new spin to get yields up to make more money, which companies are always doing. I think it was expected they would retool something to improve yields. Charlie was on about how they could not get a working card without another spin. Still not great news if it is true, but I don' think a new spin would change anything but profit margins, not performance.. not until a revision anyway.

 

[Rezist]

I don't get how fud can even post this. They're pretty much dissing there own sponsors. Because of that I'm not sure if i will believe this but whats there motive for posting it.

 

[T2k]

They are talking about a new spin to get yields up to make more money, which companies are always doing. I think it was expected they would retool something to improve yields. Charlie was on about how they could not get a working card without another spin. Still not great news if it is true, but I don' think a new spin would change anything but profit margins, not performance.. not until a revision anyway.

That's the point, a "working card" - they had to lower their clock target, their cores target etc and it's still hot and still has super-low yields.

 

[SlowSpyder]

I don't get how fud can even post this. They're pretty much dissing there own sponsors. Because of that I'm not sure if i will believe this but whats there motive for posting it.

I'd like to think that some of these sites post these stories because it's news or a rumor that they believe to be true, not to please sponsors...

For whatever it's worth, I see an XFX, OCZ, and CoolerMaster ad on the page.

 

[Daedalus685]

That's the point, a "working card" - they had to lower their clock target, their cores target etc and it's still hot and still has super-low yields.

I know, but the thing could fly to mars and back on a table spoon of love and they would send it for another spin if the new spin would cost less than the expected profits of improving yields.

I understood from what charlie wrote that he thought that it was possible (not certain) that another spin was required before they could even release the thing. All we have here is another spin before it is making money. I'm not sure we will see it address anything like heat or clocks.. those ships have sailed and those targets missed. They have to hope those aspects come for the 385 (or whatever).

 

[Lonyo]

Well it's not that far fetched in some ways. It's (I would think) impossible for Fermi to have yields as good as Cypress (due to being bigger), but then it's not always about yields, but also what they do with the chips which don't make it.
What is this 20% yield number figuring? Is it based on an (apparently undetermined) core clock? Is it based on 512 fully functional pipes?

20% yield for 512-core fast enough clock speed Fermi isn't terrible if the failed chips are fine for the 448 pipe Tesla product, because they will take up a number of the failures, and sell for a nice price.

20% faster sounds entirely plausible though, given the 36% rumour that some people mentioned from CES, but obviously it depends on what sort of clocks NV go with.

All in all, it's in no way surprising news, apart from the size of the disparity, but even that is difficult to gauge.

 

[T2k]

, given the 36% rumour that some people mentioned from CES,

Had nothing to do with CES - it was from a single forum post on a gaming board, made by some very childish-behaving guy with some Indian name, claiming to be an ex-NV employee...

...most likely utter BS.

 

[ScorcherDarkly]

I understood from what charlie wrote that he thought that it was possible (not certain) that another spin was required before they could even release the thing. All we have here is another spin before it is making money.

I read it as the yields were too low to release a consumer card at a profit, so they're going to sell these chips in Tesla and Quadro cards and have to do a re-spin before consumer cards would be available. So they'd be able to field a working card, yes, but not one they want to release. The spirit of Charlie's prediction seems correct, if not the letter. Granted, this assumes the fud report is true, which is still a big assumption.

 

[Daedalus685]

I read it as the yields were too low to release a consumer card at a profit, so they're going to sell these chips in Tesla and Quadro cards and have to do a re-spin before consumer cards would be available. So they'd be able to field a working card, yes, but not one they want to release. The spirit of Charlie's prediction seems correct, if not the letter. Granted, this assumes the fud report is true, which is still a big assumption.

They mention limited consumer, not none. I thinjk the idea in this rumjor is that most GPUs will become tesla to make as much money as possible, but not all as they still plan to go ahead with a release on this spin, correcting the profit issues as soon as they can.

They are right of course, they have to get something out in Q1 or things will look kind of sad.

I think charlie was pretty spot on for fermi.. but it was likely another spin was goign to be done at some point anyway. I read it as another spin was goign to be needed before it is even possible to release fermi, not that it will just have worse profit. AT any rate, all semantics.. if fud is right (big if) this is bad news for nvidia.

 

[TemjinGold]

Chip yields seem to be the topic of discussion lately as the much larger Nvidia Fermi chip is struggling with yield issues as our sources suggest that the actual yields are as low as 20% and are not improving quickly according to sources close to the fabs. With wafer starts costing about $5K the numbers suggest that each chip would cost an astounding estimated $200 per chip which pegs the card with a sticker price of about $600.

Lol...

 

[akugami]

Well, we've seen rumors of 36% in other places. Both numbers could still be right. Perhaps 20% average from GF100 vs 5870 but 36% max in certain cases. It's still all rumors at this point.

Assuming the rumors are true, if the GF100 is released at roughly a $600 MSRP and it's only going to be about 20% faster then nVidia is going to have a problem. They won't have issues moving the initial few lots of cards but nVidia will have their work cut out. And that's not counting a likely refresh of the Radeon 5 series, perhaps a 5890 released soon after Fermi.

 

[Jaskalas]

I'd love to find a nice long and detailed article on how it came to this. Maybe some months down the road post-launch.

 

[sandorski]

If it's all true, Nvidia is in for a tough year(even if you ignore the $200/chip). There simply won't be enough Supply of Nvidia cards to dent the AMD/ATI domination of Sales, 20% faster or not. If AMD/ATI respond by upping their performance, that 20% performance advantage will disappear.

Time will tell.

 

[v8envy]

NV still has the goold old 8800/9800/250/(and dare I predict it, 340?) to carry them through 2010. Crank up the marketing machine, get the viral guys to yammer incessantly about PhysX & 3d glasses and everything should turn out well even without enthusiast parts.

To listen to Wreckage, nv is dominating the market even without any competitive products. Who knows, that may continue for yet another year.

 

[sandorski]

NV still has the goold old 8800/9800/250/(and dare I predict it, 340?) to carry them through 2010. Crank up the marketing machine, get the viral guys to yammer incessantly about PhysX & 3d glasses and everything should turn out well even without enthusiast parts.

To listen to Wreckage, nv is dominating the market even without any competitive products. Who knows, that may continue for yet another year.

Wreckage will be happy no matter what happens, and I'm sure those cards will Sell well, but AMD/ATI will only gain Marketshare against that level of competition.

 

[uclaLabrat]

I'd love to find a nice long and detailed article on how it came to this. Maybe some months down the road post-launch.

It came to this because Nvidia has decided that they find an emerging market in CUDA, or using the GPU as a general processor. The problem is, they want gamers to subsidize the effort, by making chips with dedicated CUDA logic that has no real application to 95% of Nvidia's customers. That dedicated logic (of what, 50% of the transistor budget? Someone can quote the right numbers) results in increasingly higher costs for the consumer, by increasing chip complexity, for negligible performance benefit.

 

[MarcVenice]

"TSMC claim that actually the 40nm process that is being used for the Cypress family is routinely achieving yields in the 60% to 80% range, which is actually confirmed by the fact that ATI is keeping product in the pipeline and you can actually buy boards at retail."

How does being for sale in retail confirm such high yields? Yields could be shit but some boards might still pop up in retail, if ATI only orders enough wafers

I went from not believing Fudzilla to simply disliking them. They pull so much shit from their ass and post it as truth, it's mindboggling.

 

[nitromullet]

I think it will be interesting to see if NV is successful in harvesting enough not fully functional flagship Fermi chips to use in their #2 card ("GTX 360"?). If the 20-36% performance rumor has any truth to it their #2 card might slot itself in between the 5870 and 5850 in both cost and performance, which would make it a contender.

 

[Lonyo]

I'd love to find a nice long and detailed article on how it came to this. Maybe some months down the road post-launch.

How it came to this, if this is true:
TSMC had 40nm problems.
NV decided to make a very large GPU chip.
(The problem is mainly #2)

Fin.

 

[nitromullet]

"TSMC claim that actually the 40nm process that is being used for the Cypress family is routinely achieving yields in the 60% to 80% range, which is actually confirmed by the fact that ATI is keeping product in the pipeline and you can actually buy boards at retail."

How does being for sale in retail confirm such high yields? Yields could be shit but some boards might still pop up in retail, if ATI only orders enough wafers

I went from not believing Fudzilla to simply disliking them. They pull so much shit from their ass and post it as truth, it's mindboggling.

"Confirmed" is probably too strong a term, but I do think there is some sense to their conclusion. Consider that we do know that availability of Cypress has improved and that generally yields do increase as a process matures, and I'd say it's a pretty safe assumption to make that yields have gotten better.

That being said, I agree that 60-80% is far from confirmed by anyone on the outside. Then again, the yield itself is not very relevant when you have availability but no competition to drive the price down.