Nvidia Kepler Yields Lower Than Expected –CEO. Fermi 2.0?

[nyker96]

Link: http://www.xbitlabs.com/news/graphi...er_Than_Expected_Chief_Executive_Officer.html

 

[blackened23]

This doesn't bode well for prices. Although I think April-May for gk104 is mostly set in stone

 

[notty22]

Here is more on Kepler from that article which some deny is even being made.

Nvidia has very high expectations for its Kepler generation of graphics processing units (GPUs). The company claims that it had signed contracts to supply mobile versions of GeForce “Kepler” chips with every single PC OEM in the world. In fact, Nvidia says Kepler is the best graphics processor ever designed by the company. [With Kepler, we] won design wins at virtually every single PC OEM in the world. So, this is probably the best GPU we have ever built and the performance and power efficiency is surely the best that we have ever created,” said Mr. Huang.
Unfortunately for Nvidia, yields of Kepler are lower than the company originally anticipated and therefore their costs are high. Chief exec of Nvidia remains optimistic and claims that the situation with Fermi ramp up was ever worse than that.
“We use wafer-based pricing now, when the yield is lower, our cost is higher. We have transitioned to a wafer-based pricing for some time and our expectation, of course, is that the yields will improve as they have in the previous generation nodes, and as the yields improve, our output would increase and our costs will decline,” stated the head of Nvidia.

 

[AnandThenMan]

Here is more on Kepler from that article which some deny is even being made.

Like who?

 

[Phynaz]

signed contracts to supply mobile versions of GeForce “Kepler” chips with every single PC OEM in the world.

I find that very difficult to beleive. Maybe every tier one OEM, but every OEM?

 

[Crap Daddy]

AMD might not be bragging that the yields are poor but the price on the 7900 series might be a direct consequence.

 

[Silverforce11]

28nm, TSMC is charging everyone per wafer, unlike at 40nm where NV had the benefit of being priced only for "good" dies. Price-war? No way. Expect high prices if their yields are bad.

SA just had a long article all about this, and specifically, only NV is having yield troubles at 28nm.

 

[Olikan]

In fact, Nvidia says Kepler is the best graphics processor ever designed by the company

yeah, like Nvidia would just go and say..."LOL... Kepler sucks, lolololol"

I find that very difficult to beleive. Maybe every tier one OEM, but every OEM

why not?
there is only 2 companies that oems have to chose, and they can choose both without a problem...

 

[MaxPayne63]

I'm curious about what is generally considered low yield. Assume 100 chips per wafer. What would the approximate breakdown be between chips working as designed and total losses for 'low yield?'

 

[Stuka87]

SA just had a long article all about this, and specifically, only NV is having yield troubles at 28nm.

I think this may come down to die size. AMD has again gone with a small die design, whereas NV has gone with a large die. Larger dies have shown to be more difficult to produce initially.

 

[Olikan]

I'm curious about what is generally considered low yield. Assume 100 chips per wafer. What would the approximate breakdown be between chips working as designed and total losses for 'low yield?'

it's hard to answer that...it depends on many things (most important is profit)

you can use amd see the yield problem, amd could had the tahiti at 1,1Ghz (it's the spot that it get heavy bandwidth limited)

yet, all 7xxx cards, so far, are overpriced, even when they are downcloked and smaller than the old cards

 

[Lonyo]

I'm curious about what is generally considered low yield. Assume 100 chips per wafer. What would the approximate breakdown be between chips working as designed and total losses for 'low yield?'

Wait for IDC to get some graphs and drop some knowledge bombs. I'm fairly sure there have been yield over time graphs floating around.

The main issue with saying what "yields" are, is defining yield.
You could say that the GTX580 die has X yield, but the yield increases if you include the GTX570.
If you then go forwards in time, retrospectively the yield increases when you add the GTX560 448-core in, since that uses the same chips.

TSMC's 40nm yields were below 60%, which is pretty bad, but not dire, and as mentioned, it depends on how you define yield. When TSMC talks about yield, their definition is going to be different to NV/AMD, because of their wide customer base who may not be able to harvest defective dies so easily.

http://www.xbitlabs.com/news/other/...SMC_s_40nm_Yields_Improved_to_60_Company.html

 

[Arzachel]

I think this may come down to die size. AMD has again gone with a small die design, whereas NV has gone with a large die. Larger dies have shown to be more difficult to produce initially.

That's the problem GK104 is rumored to be about the same size as Tahiti, with GK100/110 delayed until August or so. It's understandable that they have issues with the later, but having issues with the former isn't exactly good news.

 

[maddie]

I think this may come down to die size. AMD has again gone with a small die design, whereas NV has gone with a large die. Larger dies have shown to be more difficult to produce initially.

This time Nvidia is supposed to be starting with dies even smaller than Tahiti. They should have even better yields than AMD assuming everything else equal, Unrealistic assumption however.

 

[Idontcare]

Wait for IDC to get some graphs and drop some knowledge bombs. I'm fairly sure there have been yield over time graphs floating around.

The main issue with saying what "yields" are, is defining yield.
You could say that the GTX580 die has X yield, but the yield increases if you include the GTX570.
If you then go forwards in time, retrospectively the yield increases when you add the GTX560 448-core in, since that uses the same chips.

TSMC's 40nm yields were below 60%, which is pretty bad, but not dire, and as mentioned, it depends on how you define yield. When TSMC talks about yield, their definition is going to be different to NV/AMD, because of their wide customer base who may not be able to harvest defective dies so easily.

http://www.xbitlabs.com/news/other/...SMC_s_40nm_Yields_Improved_to_60_Company.html

Yeah, what folks have to keep in mind (or become aware of) is that there are two general types of yields - Parametric Yield and Functional Yield.

Functional yield is your basic "does the chip power up? does it do math correctly? (1+1=2? or does 1+1=3?)"

This type of yield is heavily dependent on the defect density in the fab (how many "killer" particles are falling onto the wafers while the wafers are in the fab). It is the type of yield that starts out really bad and then improves as the node matures.

Forgive me for recycling this graph from 2 yrs ago, but its already prepared and in my photobucket account so we'll use it here to speak to:

In terms of functional yield, if Kepler is much larger than AMD's GPU or qualcomm's other chips being made on 28nm then it would be of no surprise to hear that Kepler yields were problematic while everyone elses yields were OK.

The second yield type is Parametric yield. This is the yield that is based on clockspeed binning, power-consumption binning, and reliability binning.

The chip might function just fine but it can't clock over 500MHz, or it can clock just fine but it consumes 500W of power. All sorts of electrical issues can come into play which limit parametric yield.

If TSMC was having parametric yield issues you would expect it to impact all customers.

Final yield is the product of functional yield and parametric yield.

Based on the general statements made by NV, and the anecdotal observations of other 28nm customers not citing yield as a problem, I am inclined to believed the issue is simply one of functional yield (big die = lower yield).

Which, if true, means we should not expect a Fermi 2.0 in terms of power-consumption (a parametric yield issue), but maybe a Fermi 2.0 in terms of scheduling slips for functional yield enhancement purposes,

 

[blackened23]

Great post IDC, very informative. Kinda sheds light on why (possibly) AMD may have been overly conservative with tahiti, and the price.

 

[AnandThenMan]

Great post IDC, very informative.

Agreed, much appreciated. :thumbsup:

I'd like to know the relationship (if any) as to how overclocking headroom lines up with yields. I ask because Tahiti overclocks are generally excellent, in my view that would seem to correlate into good overall yields. Power consumption is also very healthy. Is it possible to have poor yields but the functional chips work very well?

 

[MaxPayne63]

Great post IDC, very informative.

Indeed. That was exactly what I was looking for, and just after I finished dinner. :thumbsup:

 

[Ajay]

I'm curious about what is generally considered low yield. Assume 100 chips per wafer. What would the approximate breakdown be between chips working as designed and total losses for 'low yield?'

Well, IIRC, Fermi was in the 20% range, and that was really bad. According to the Xbitlabs article, Kepler isn't as bad. What NV considers to be poor, visa vi yields, is somewhat subjective. 40% would be pretty low, with capacity being an issue ATM, that would put NV in a bad position. As TSMC brings more process equipment online, things will improve - plus yields will go up over time as well.

As blackened23 already said, this does not bode well for Kepler pricing.

 

[Silverforce11]

Agreed, much appreciated. :thumbsup:

I'd like to know the relationship (if any) as to how overclocking headroom lines up with yields. I ask because Tahiti overclocks are generally excellent, in my view that would seem to correlate into good overall yields. Power consumption is also very healthy. Is it possible to have poor yields but the functional chips work very well?

Totally separate issue.

AMD clock ref design that low because they were targeting a certain TDP. The chips are able to clock much higher but then power consumption also goes way up as we can see from reviews.

Functional yield affect bigger chips more, that's obvious. But the signs point towards poor yields for kepler as a whole whereas other companies using TSMC's 28nm has not expressed yield concerns. This suggest something inherently difficult with making kepler that TSMC is struggling with, not just die size. Lets face it, if NV didn't have problems, kepler would be here already.

 

[badb0y]

Good thing I picked up a 7970 early on I guess I will wait for the Kepler refresh to change my card.

 

[tviceman]

Wow spin city is working hard tonight.

Thread-crapping is not allowed.

Moderator jvroig

 

[notty22]

Nvidia turned profits and just missed estimates in earnings. So the CEO is expected to report why.
When you see losses, layoffs and top executives leaving then it's time to get worried.

 

[blastingcap]

Nvidia turned profits and just missed estimates in earnings. So the CEO is expected to report why.
When you see losses, layoffs and top executives leaving then it's time to get worried.

As a consumer, I want both companies to be healthy. We should work to get more people into PC gaming and not buy consoles.

 

[Silverforce11]

Well, IIRC, Fermi was in the 20% range, and that was really bad. According to the Xbitlabs article, Kepler isn't as bad. What NV considers to be poor, visa vi yields, is somewhat subjective. 40% would be pretty low, with capacity being an issue ATM, that would put NV in a bad position. As TSMC brings more process equipment online, things will improve - plus yields will go up over time as well.

As blackened23 already said, this does not bode well for Kepler pricing.

Fermi yields were bad but they had a pricing agreement with TSMC, they only paid for functional chips, and not per wafer, which only hurts TSMC's bottom line and not NVs. This is the critical difference now.

Typically a 40nm wafer costs ~$5,000, probably much more on 28nm. Actually, not probably but definitely given the news surrounding TSMC's 28nm production.

I guess they can squeeze ~160 gk104 onto a wafer given its die size estimates. Raw die cost are only a fraction of the final product costs though.

Edit: whatever the case, i'm delaying gpu upgrades until i see gk104 benchmarks. With something concrete to work with, one could speculate on the big kepler's potential performance as well.