Kepler Yield Issues

[Oryon22]

Discuss.
http://www.xbitlabs.com/news/graphics/display/20120216125300_Nvidia_Kepler_Yields_Lower_Than_Exp ected_Chief_Executive_Officer.html

[Mutant]

Chalies take on this smells of truth
Nvidia comes somewhat clean on their quarterly conference call (http://semiaccurate.com/2012/02/15/nvidia-comes-somewhat-clean-on-their-quarterly-conference-call/)

"As usual, SemiAccurate immediately checked out the 28nm yield claims, and found absolutely no change. If anything, yields were better than the last time we checked on them, and no one queried could explain the reports. TSMC has publicly stated that yields are better than expected, and AMD has publicly said similar things about their 28nm experiences."

"During today’s call, Nvidia once again blamed TSMC for their yield problems. As we have said, Nvidia seems to be alone in having these problems. Again. The yield problems, combined with the gem that the 40nm per-chip pricing had expired or did not carry over to 28nm, dropped Nvidia’s numbers quite a bit."

[Vancha]

In other words, the price of Kepler will be sky high.

Joy.

[Rbstr]

Meh, yields are always too low. They always get better as they keep making them. Thus is the nature of building chips.

[Goku]

I'm still waiting for this to be released just for my competition. If for whatever reason my 560 Ti cannot handle PS2 on max setting I will likely be looking at a GTX 660 or 670.

[Fenrys]

It's the architecture that's giving them headaches. They use a massive die for each GPU compared to AMD's petite design.

For every millimeter of die, there is a % chance of something going wrong. A bigger die means fewer chips can be printed on each silicon wafer and a larger percentage of them will have "something going wrong". AMD's wafers probably have the same number of flaws, but since their chips are smaller, those flaws effect a smaller percentage of the total yield from each wafer.

http://img717.imageshack.us/img717/3654/greenvred.png

^ the dots are flaws and the squares without dots in them are usable GPUs

[Mutant]

Meh, yields are always too low. They always get better as they keep making them. Thus is the nature of building chips.

Indeed but TSMC and Some of their customers (inc AMD but I cant find a quote atm) have stated that TSCM 28nm yields are better than expected.
And if that is true it seems that Nvidia yield issues are more to do with chip design than the process.

http://www.xbitlabs.com/news/other/display/20120119204855_TSMC_Ships_More_28nm_Wafers_than_Ex pected_Projects_Rapid_Ramp.html
"Maria Marced, president of TSMC Europe, repeated what has been said before by herself and other TSMC executives; that defect density reduction is on track for the 28-nm node and ahead of where TSMC was with 40/45-nm process technology at an equivalent stage in its roll out."

http://www.electronicsweekly.com/Articles/15/02/2012/52982/tsmcs-28nm-is-great-says-xilinx.htm


It's the architecture that's giving them headaches. They use a massive die for each GPU compared to AMD's petite design.

For every millimeter of die, there is a % chance of something going wrong. A bigger die means fewer chips can be printed on each silicon wafer and a larger percentage of them will have "something going wrong". AMD's wafers probably have the same number of flaws, but since their chips are smaller, those flaws effect a smaller percentage of the total yield from each wafer.

http://img717.imageshack.us/img717/3654/greenvred.png

^ the dots are flaws and the squares without dots in them are usable GPUs


What you say is indeed totally right but note that;

GK104 is said to be ~ 342mm^2
Tahiti is 378 mm^2

While for the last 4 gen (inc 1/2 gen) Nvidia have also produced a 500++ mm^2 chip GK100 went MIA some time ago likely due to bad design coupled with the effect you describe. GK110 seems to have been rushed forward with some fixes to fill the gap, it is said to have only tapped out a month or so ago so it would be too early to get yield numbers back on it.

So it would seem the recent blow out by Nvidia on yields would be talking about GK104 which is smaller than Tahiti. (There is a possibility they were taking about GK100 but i got the impression that Nvidia had scrubbed it from existence)

edit:
For anyone interested in the choices in GPU design these articles are great;
The RV770 Story: Documenting ATI's Road to Success (http://www.anandtech.com/show/2679)
The RV870 Story: AMD Showing up to the Fight (http://www.anandtech.com/show/2937)

[Fenrys]

The GTX 660 Ti (GK 110, 550mm^2) supposedly is the performance equal of the HD 7950 (Tahiti, 365mm^2), and GK 110 is scheduled to be released before GK 104 or 106. If nvidia is having yield issues, it's probably with the 550mm^2 GK 110, since that is what is expected to be in stores first.

GK 104 will likely compete with Pitcairn which is rumored to be ~250mm^2.

[Mutant]

The GTX 660 Ti (GK 110, 550mm^2) supposedly is the performance equal of the HD 7950 (365mm^2), and GK 110 is scheduled to be released before GK 104 or 106. If nvidia is having yield issues, it's probably with the 550mm^2 GK 110, since that is what is expected to be in stores first.

GK 104 will likely compete with Pitcairn.

Got any sources?
Everywhere I have seen shows GK104 due in shops in a month or 2 and chips are with AIBs already.
and GK110 due in 8-10 months and only tapped out a month ago.

[Oryon22]

The wait for Kepler is killing me.

[Fenrys]

release dates (http://lenzfire.com/2012/02/entire-nvidia-kepler-series-specifications-price-release-date-43823/)

Now that I'm looking at it again, it's not only the 670 and 680 (gk 110) that are expected in April, but the 660 too (gk 104).

Unfortunately, neither the above article nor anything on SemiAccurate are very reliable.

[Mutant]

release dates (http://lenzfire.com/2012/02/entire-nvidia-kepler-series-specifications-price-release-date-43823/)

Now that I'm looking at it again, it's not only the 670 and 680 (gk 110) that are expected in April, but the 660 too (gk 104).

Unfortunately, neither the above article nor anything on SemiAccurate are very reliable.

I think that lenzfire article is bs and has had no independant confirmation on any of those specs.

Most sources seem to be converging on a few things naimly:

no hot clocks for Kepler
1536 shaders on GK104
256 bit memory bus on GK104
GK110 Q3 2012 at best.


I guess we will find out in a few weeks.


Also;
While Charlie like everyone is never 100% and will almost always put some sort of slant on his articles I have found Semmiaccurate to be more accurate than most especially when it comes to specs /dates / code names, but much less so on performance.

some sources:

3D Center: The current specification for GK104 Kepler performance chip (http://translate.google.com/translate?sl=auto&tl=en&js=n&prev=_t&hl=en&ie=UTF-8&layout=2&eotf=1&u=http%3A%2F%2Fwww.3dcenter.org%2Fnews%2Fdie-aktuellen-spezifikationen-zum-gk104-kepler-performance-chip&act=url)

sweclockers: NVIDIA GeForce GTX 670 Ti in March (http://translate.google.co.uk/translate?sl=sv&tl=en&js=n&prev=_t&hl=en&ie=UTF-8&layout=2&eotf=1&u=http%3A%2F%2Fwww.sweclockers.com%2Fnyhet%2F15083-nvidia-lanserar-geforce-gtx-670-ti-i-mars&act=url)

Nordic Hardware: NVIDIA Kepler GK104 gets 1536 CUDA cores (http://www.nordichardware.com/news/71-graphics/45257-nvidia-kepler-gk104-gets-1536-cuda-cores.html)