Less than 10 thousand fermi cards delivered,Nvidia stock falls.

[Will Robinson]

Looks like Charlie was right again...

http://www.xbitlabs.com/news/video/...t_20_30_Further_Delays_Possible_Analysts.html

http://www.google.com/hostednews/ap/article/ALeqM5gFq5LCjssCxEWoeXUM0_uqaXvsUAD9F4BQNO0

 

[DanielNenni]

I just did a blog on this, TSMC 40nm, where I talked with Dr. Shang-Yi Chiang, Vice President of TSMC R&D on 20nm, 28nm, and 40nm.

Redefining the Foundry Model: TSMC versus GlobalFoundries

Nvidia designed in 4B+ single vias (double via's are highly recommended at 40nm) and is paying the yield price. 40nm is yielding fine for dozens of other companies such as Altera, Qualcomm, etc....

TSMC 28nm has restricted design rules for better yield. Lets hope Nvidia follows them.

 

[BFG10K]

It looks like that 10,000 figure was right after all.

 

[BenSkywalker]

This is all binning though. Yields were still far below 20 percent for both the full 512 shader version and the 448 shader part combined.

That would put yields of the new lower spec GTX480 at well under the two percent Nvidia saw last fall.

http://www.semiaccurate.com/2010/03/29/why-nvidia-hacked-gtx480/

Real yields 20%-30% according to actual checks. Also, the analyst said that nVidia has 10K units as of 4/17- while not a large amount by any means that also indicates that a total of 0 units sold from 3/28 when they started hitting retail and then and also for Charlie to be right no more chips can ever be sold(10K units maximum for the lifetime of the product). I'd say this article puts into sharp focus how wrong Charlie was, off by an order of magnitude on the yields and clearly off by a decent amount on the lifetime run of the GTX470/480 parts already under a month into availability.

 

[TemjinGold]

But as you said, these numbers are as of 4/17. The yields COULD have been under 2% "before" and still be at 20-30% "right now." The analyst also said "less than 10,000," which would not bust Charlie just yet. Not looking to defend him but your argument has quite a few holes in it...

 

[Lonyo]

http://www.semiaccurate.com/2010/03/29/why-nvidia-hacked-gtx480/

Real yields 20%-30% according to actual checks. Also, the analyst said that nVidia has 10K units as of 4/17- while not a large amount by any means that also indicates that a total of 0 units sold from 3/28 when they started hitting retail and then and also for Charlie to be right no more chips can ever be sold(10K units maximum for the lifetime of the product). I'd say this article puts into sharp focus how wrong Charlie was, off by an order of magnitude on the yields and clearly off by a decent amount on the lifetime run of the GTX470/480 parts already under a month into availability.

Both stories talk about 10,000 units being available. However one talks about GTX480 (S|A) and the other Fermi (the AP article)
S|A talks about 9,000 wafers of 104 GPUs per wafer, while the other article talks about 20 to 30% yields.
If we assume that 20% is the yield, and it's 100 GPUs per wafer (rounded for simplification), and 10,000 cards, then that means 20 per wafer, meaning only 500 wafers. If it's 30% then that means 334 wafers. to produce all the cards.

Clearly something doesn't jive here.
So is it that Charlie is off on the yield, or that he is off on the number of wafers, or something else?

Source A: 20~30% yield producing 10,000 chips. = 334~500 wafers
Source B: 2% yield producing 10,000 chips + 20,000 harvested chips (giving en effective 6% yield). From 9,000 wafers.

Also this article implies only 1/3rd the total number of cards that Charlie claimed would be available. (10k vs 30k)

 

[BenSkywalker]

The yields COULD have been under 2% "before" and still be at 20-30% "right now."

In 2-3 weeks? Charlie stated 2% on March 29th.

Clearly something doesn't jive here.
So is it that Charlie is off on the yield, or that he is off on the number of wafers, or something else?

Perhaps nV hasn't gone through their initial run of wafer orders? It's not like a 9K run is a trivial matter.

Also this article implies only 1/3rd the total number of cards that Charlie claimed would be available. (10k vs 30k)

Which Charlie claim are we using? I liked the 10K total ever claim, I thought that was the most represenative of Charlie and it fit in with his 2% yield theory the best.

 

[NoQuarter]

In 2-3 weeks? Charlie stated 2% on March 29th.

Hmm closest thing I could find to a 2% claim in the March 29th article was this..

As things stand, the 9,000 risk wafers seem to have produced less than 10,000 GTX480s and about twice that many GTX470s if the rumored release numbers are to be believed. That would put yields of the new lower spec GTX480 at well under the two percent Nvidia saw last fall.

Is that what you're talking about? Cuz he makes a very caveated statement here.

 

[soccerballtux]

I just did a blog on this, TSMC 40nm, where I talked with Dr. Shang-Yi Chiang, Vice President of TSMC R&D on 20nm, 28nm, and 40nm.

Redefining the Foundry Model: TSMC versus GlobalFoundries

Nvidia designed in 4B+ single vias (double via's are highly recommended at 40nm) and is paying the yield price. 40nm is yielding fine for dozens of other companies such as Altera, Qualcomm, etc....

TSMC 28nm has restricted design rules for better yield. Lets hope Nvidia follows them.

They didn't know this until too late.
The other companies are not producing humassive chips like Nvidia.

 

[Lonyo]

Hmm closest thing I could find to a 2% claim in the March 29th article was this..


Is that what you're talking about? Cuz he makes a very caveated statement here.

And the rumoured release numbers ended up being 1/3rd of what he claimed to be hearing.
Which means that his estimates of yields based on 9,000 wafers must be way off.

So that begs the question of how he managed to get an understanding that 9,000 wafers had been run at TSMC and already harvested for chips.
Clearly the problem is he managed to come up with a 9,000 wafer number that doesn't fit with yields or production numbers. The flawed yield calculation then came from this 9,000 wafer figure, which was either wrongly interpreted or just flat out wrong.

 

[Lonyo]

http://www.theinquirer.net/inquirer/news/1594021/tsmc-40nm-yield-issues-explained

Here's a link talking about 80k wafers per quarter, so for 9k wafers to have been produced before launch would have required over 10% of TSMC's 40nm production to be devoted over the course of 3 months to just making Fermi wafers.

 

[NoQuarter]

And the rumoured release numbers ended up being 1/3rd of what he claimed to be hearing.
Which means that his estimates of yields based on 9,000 wafers must be way off.

So that begs the question of how he managed to get an understanding that 9,000 wafers had been run at TSMC and already harvested for chips.
Clearly the problem is he managed to come up with a 9,000 wafer number that doesn't fit with yields or production numbers. The flawed yield calculation then came from this 9,000 wafer figure, which was either wrongly interpreted or just flat out wrong.

Yea I wasn't really addressing that, just the 2% claim. It's hard to put it all together. Does his 2% claim only apply to the first revision (my impression)? Are all the statements properly using yield to describe all usable chips or is it including only fully functional GTX 480 parts? If there was only 1 9,000 wafer run have we seen all the GPUs (10,000? 30,000?) from it?

 

[Lonyo]

Yea I wasn't really addressing that, just the 2% claim. It's hard to put it all together. Does his 2% claim only apply to the first revision (my impression)? Are all the statements properly using yield to apply to all usable chips or is it including only fully functional GTX 480 parts? If there was only 1 9,000 wafer run have we seen all the GPUs (30,000?) from it?

The 2% claim was talking about just the revised GTX480 (the 480SP part) and there was also a claim about sub-2% yields on the original (512SP) GTX480 part.
The 2% didn't include harvested dies (GTX470), although at the bottom it talks about their being twice as many GTX470's as GTX480 (480SP)'s, he talks about 10k GTX480's and 20k GTX470s, meaning that yields overall for both cards would be something under 6% after harvesting, assuming 30k cards and 9k wafers.

Now since the actual number of cards was under 10k (according to S|A it was 8k and according to this story it was 10k, assuming for GTX470 and 480 since it just says "Fermi") that would put overall yields of all chips (based on 9k wafers) at the sub 2% mark, rather than just the GTX480 (480SP) being at sub 2%.

The obvious assumption would be that the 9k wafer claim is the main issue, and all other figures derived from that are by association completely off base.

 

[Janooo]

It does not appear that all 9k of wafers were used, just a portion of it.

 

[yh125d]

Charlie charlie yadda yadda


Can anyone attest to the reliability of Needham & Co?

 

[DanielNenni]

Charlie charlie yadda yadda


Can anyone attest to the reliability of Needham & Co?

Needham is reliable, serious Wall Street player.

The article I saw had it marked as satire:

ATI's Southern Islands tapes out
SemiAccurate (satire)
Southern Islands (SI) taped out

 

[Meghan54]

Can anyone attest to the reliability of Needham & Co?

Apparently the Street Insider considers them reliable enough to post about their warning.....

http://www.streetinsider.com/Downgr...Fermi+Ramp+and+Gross+Margin+Risk/5536612.html


And this warning from Needham was enough to drop Nvidia's stock price measurably.

 

[MrK6]

Why is anyone surprised? This has been coming for months. Charlie didn't have to do any major detective work to figure this one out, and neither should anyone else have.

 

[Madcatatlas]

Charlie Demrijan is one of the few people (related to hardware news) who goes out on a limb, making himself a pretty good target if he fails. But it seems, for your average reader/forumvisitor, that he is more right with his often way-too-early conclusions, than he is wrong.

The world of hardware news would be a boring place without the likes of Charlie in it to spice things up. nVidia fans are hardware fans aswell, they appreciate it in their own way

Personally I think his articles are a great read. Mostly because of the humor. Just look at his article..lol

 

[Compddd]

"With ati re-punching nvidia in the face this year with new 6000 series on 40nm so no delays there, and moving down to 28nm on GF next year for 7000 series, I suddenly feel sorry for nvidia."

That comment on the article made me lol for some reason

 

[scooterlibby]

I have to give the obligatory 'competition is good' admonition. If Nvidia becomes insolvent, as Charlie alluded to, and it's only ATI and Intel, we are fucked.

 

[ZimZum]

I have to give the obligatory 'competition is good' admonition. If Nvidia becomes insolvent, as Charlie alluded to, and it's only ATI and Intel, we are fucked.

Its going to take alot more than one bad product launch to sink nvidia.

 

[waffleironhead]

Its going to take alot more than one bad product launch to sink nvidia.

too true, but is you look at their current product offerings they are diminishing. With the loss of the chipset business, low volume of fermi, and low adoption of tegra where do you see the revenue coming from to cover the r&d for future products?

 

[RobertR1]

I read the article elsewhere this morning and though "I should head ever to anand and see how Ben is spinning the hell out of this one." Sure enough, 4th post...

 

[Lonyo]

too true, but is you look at their current product offerings they are diminishing. With the loss of the chipset business, low volume of fermi, and low adoption of tegra where do you see the revenue coming from to cover the r&d for future products?

Fermi was never going to be a high volume product, and doesn't need to be.
From one slide from Oak Ridge about a different project, they were talking about GPU based HPC installations, and they had a test case of 250 GPUs scheduled for 2010. The real project (of greater scale) was scheduled for 2012.

Do you really think that current state Fermi, even if it had been flawlessly executed, would have ended up being a primetime product? I don't. It's a first revision product designed to address the shortcomings of GPUs in the HPC market. That doesn't make it a product for the HPC market to widely adopt, it makes it a product to get the HPC market to test out using GPUs and then in the future Fermi II or whatever gets put into large scale projects, after the concept has been proven useful.