NVIDIA Kepler speculation thread

[UniversalTruth]

Fate, something similar to fruiti.co's. People are ready to spend enormous insane amounts of money for products which have none or negative performance advantages over competition's... Maybe some unnatural forces could influence it too.

 

[CarstenS]

AMD winning a [strike]benchmark[/strike] marketing tool created by a company that sells the same underlying video processor technology created by Tensilica is hardly surprising. It's cheating, d'oh.

Silicon Optix uses Xtensa in their HQV video processor product. ATI/AMD also license the same tech in their UVD.

Of course it's not cheating! AMD simply uses some post-processing in their Catalyst Control Center as the default like Mosquito Noise, Dynamic Contrast or Deblocking. This positively influences the HQV score but does not deliver movies the way they were intended - some like "300" add film grain as a DI aftereffect and the CCC default settings do a good job at removing those.

On the low side, this is responsible for the rather high power consumption, even the latest GCN cards show during Blu-ray playback. Turning those enhancement off, reduces power from 45-46 to 32-33 watts (graphics card only) on a HD 7800 as shown here (in the text):
http://www.pcgameshardware.de/aid,8...nell-und-sparsam-dank-28-nm/Grafikkarte/Test/

 

[xDxD]

http://translate.google.it/translat...SE-auf-Kepler-Basis/Grafikkarte/News/&act=url

 

[RecessionCone]

Remember when people were saying Kepler was impossible to find because it was unmanufacturable? I think it's interesting that Qualcomm is turning to
other foundries because they too are dissatisfied with TSMC supply.

http://www.theverge.com/2012/7/4/3136890/qualcomm-partners-samsung-umc-produce-snapdragon-s4

 

[Kaotik]

Yet AMD hasn't had similar problems AFAIK, so is it something in Qualcomm & nVidia designs causing too poor yields for enough supply (despite smaller dies), or perhaps they haven't just ordered enough wafers and AMD got there first with big orders?

 

[trinibwoy]

Or there's lower demand for AMD products.

 

[Homeles]

Or there's lower demand for AMD products.

You don't say?

 

[Lightman]

Or there's lower demand for AMD products.

Or AMD planned for lower demand taking projected yields into account and ordered enough wafers to sustain adequate supply.

We can spin this however we want. Fact is that AMD got it mostly right.
Don't forget AMD has 3 different chips covering $100-$500 GPU market and is able to keep up with demand. I doubt GK104 sold same amount of silicon area as combined AMD products till today. Granted, nVidia has GK107 in mobile market and just launched it for desktop as GT640 as well.

I've got impression nVidia got more design wins in mobile market they were expecting and not to loose them promised to deliver in a timely fashion hoping TSMC will magically improve yields which they didn't. Just my speculation without any insider knowledge.

 

[UniversalTruth]

The interesting thing here is the NV planned demand for GK104.

Imagine the circumstances (Tahiti was more up to the initial expectations) were a little bit different and they had to position GK104 with respect to mainstream type of product demand, and not like a little bit earlier with something very similar to satisfying ultra high-end demand.
How would they satisfy the demand then? Uber fail it could have been. (headbang) Poor AMD. ;(

BTW, what we can see right now is dramatically improved availability at newegg.

 

[CarstenS]

Remember when people were saying Kepler was impossible to find because it was unmanufacturable? I think it's interesting that Qualcomm is turning to
other foundries because they too are dissatisfied with TSMC supply.

http://www.theverge.com/2012/7/4/3136890/qualcomm-partners-samsung-umc-produce-snapdragon-s4

Yes, and for the gazillionth time: There's ONE (!) fab in TSMC that produces 28nm chips with another one scheduled to go 28 in autumn (IIRC). This ONE fab has to satisfy not only AMD with it's whole line-up of 28nm GPUs and Nvidia with their half-line-up (including the high-volume Apple stuff that utilizigin kepler) and Qualcomm with probably ub0r high volume for their SOCs.

That demand in this scenario far far far outstrips supply (even with a 100 percent yield) is simply a matter of more than three functioning brain cells.

 

[UniversalTruth]

Yes, and for the gazillionth time: There's ONE (!) fab in TSMC that produces 28nm chips with another one scheduled to go 28 in autumn (IIRC). This ONE fab has to satisfy not only AMD with it's whole line-up of 28nm GPUs and Nvidia with their half-line-up (including the high-volume Apple stuff that utilizigin kepler) and Qualcomm with probably ub0r high volume for their SOCs.

That demand in this scenario far far far outstrips supply (even with a 100 percent yield) is simply a matter of more than three functioning brain cells.

The only missing thing to support this line of thoughts is the link to prove that somehow TSMC managed to improve dramatically wafer starts, so at the moment you have supply far outstripping demand at least for all AMD cards plus 680, 670, etc.

TSMC has always been the only company but I don't remember they had severe manufacturing capacity problems.

 

[xDxD]

Yes, and for the gazillionth time: There's ONE (!) fab in TSMC that produces 28nm chips with another one scheduled to go 28 in autumn (IIRC). This ONE fab has to satisfy not only AMD with it's whole line-up of 28nm GPUs and Nvidia with their half-line-up (including the high-volume Apple stuff that utilizigin kepler) and Qualcomm with probably ub0r high volume for their SOCs.

That demand in this scenario far far far outstrips supply (even with a 100 percent yield) is simply a matter of more than three functioning brain cells.

http://www.eetimes.com/electronics-news/4376633/Qualcomm-signs-UMC-Samsung-for-28-nm-says-report

 

[silent_guy]

TSMC has always been the only company but I don't remember they had severe manufacturing capacity problems.

Maybe you should reconsider changing your insider sources?

 

[xDxD]

Now with the mention of the GTX 670 SE, that means one SKU will be based on the GK 104 chip i.e. the 670 with the other two GTX 660’s based on the GK 106 chip. The shorter PCB therefore perfectly matches up with the 670 but then again the 660 will be itself an inch smaller than the 670 and being based on the GK106 will relatively also pack a smaller cooling solution for having an overall TDP of 100W.

http://wccftech.com/nvidia-working-gtx-670-se-earlier-660-se-rumors-squished/

 

[lanek]

so IF we trust this rumor.. no 660TI based on GK104 but a 670SE... I should stop read rumors

 

[A1xLLcqAgt0qc2RyMz0y]

so IF we trust this rumor.. no 660TI based on GK104 but a 670SE

This makes more sense to me as the GK104 will then only be used for 670-690s and the GK106 for the 660s.

What are the guesses on pricing for the 670SE, 660Ti and the 660?

Mine are:

$324.99 for the 670SE
$249.99 for the 660Ti
$199.99 for the 660

 

[CarstenS]

The only missing thing to support this line of thoughts is the link to prove that somehow TSMC managed to improve dramatically wafer starts, so at the moment you have supply far outstripping demand at least for all AMD cards plus 680, 670, etc.

TSMC has always been the only company but I don't remember they had severe manufacturing capacity problems.

Really - I don't get what your trying to tell me here.

 

[trinibwoy]

Or AMD planned for lower demand taking projected yields into account and ordered enough wafers to sustain adequate supply.

There's a finite number of wafers available. Ordering "enough wafers" can't change that if that number is higher than TSMC's available capacity.

 

[jimbo75]

Nvidia won't waste the chance to milk the 660 name for all it's worth.

If there are two different chips, the faster 660 Ti will be GK104 and the slower 660 will be GK106, with 660 SE for the salvage part of the GK106. This let's them ride on the success of the 660 Ti being faster than the 7870, while the actual 660 (GK106) is slower but almost certainly more expensive. I'd put money on this.

They did this with the 560, they'll do it here. If Nvidia excels at anything it's confusing their buyers and they'll lap up the slower 660 based on the prowess of the 660 Ti.

 

[McHuj]

This makes more sense to me as the GK104 will then only be used for 670-690s and the GK106 for the 660s.

What are the guesses on pricing for the 670SE, 660Ti and the 660?

Mine are:

$324.99 for the 670SE
$249.99 for the 660Ti
$199.99 for the 660

I'm thinking 299 for the 670SE, and 349 for the 670 proper.

There seems to be slight price movement on the 670. You can get a Zotac one for $390 so a price cut maybe coming. It's probably closest competitor 7950 can be had for sub $350 now. My guess is when Nvidia finally releases the 660 series, we'll start seeing some real price competition and the price scheme be closer to what the NV 5xx/ AMD 6xxx was.