NVIDIA Kepler speculation thread

[CarstenS]

Either I am really behind things now (if so: SORRY), or Nvidia just officially moved Kepler to 2012 timeframe, accordingly Maxwell is now a 2014 arch.
http://www.pcgameshardware.de/aid,8...-2012-und-Maxwell-fuer-2014/Grafikkarte/News/

 

[Harison]

Either I am really behind things now (if so: SORRY), or Nvidia just officially moved Kepler to 2012 timeframe, accordingly Maxwell is now a 2014 arch.
http://www.pcgameshardware.de/aid,8...-2012-und-Maxwell-fuer-2014/Grafikkarte/News/

That was expected, there wont be any high-end 28nm GPUs this year, unless in marketing volume. However if NV officialy postponed, it means we wont see Kepler in 2012 January/February either, most likely Q2+ timeframe, thats when AMD high-end 28nm rumoured to be launched.

"Nvidia: 28nm Process Tech "Not Available" This Year."
http://www.xbitlabs.com/news/other/...8nm_Process_Tech_Not_Available_This_Year.html

 

[Alexko]

Either I am really behind things now (if so: SORRY), or Nvidia just officially moved Kepler to 2012 timeframe, accordingly Maxwell is now a 2014 arch.
http://www.pcgameshardware.de/aid,8...-2012-und-Maxwell-fuer-2014/Grafikkarte/News/

I don't think anyone really expected Kepler to be launched in 2011, but that's the first time NVIDIA has publicly admitted it, as far as I'm aware.

 

[CarstenS]

Frankly, when seeing the first version of the slide where Fermi was a 2009 product, I didn't think of a launch, but of an architectural presentation only.

 

[Man from Atlantis]

Frankly, when seeing the first version of the slide where Fermi was a 2009 product, I didn't think of a launch, but of an architectural presentation only.

hmm so we will see a paper launch before christmas then hard launch in Q1 '12

 

[DarthShader]

most likely Q2+ timeframe, thats when AMD high-end 28nm rumoured to be launched.

Where did you get that from? Those chips taped out in February and are set for a Q4 launch. That's the month the article you posted comes from btw. Just because 28nm isn't available for nVidia doesn't mean it's not available for everybody else.

Frankly, when seeing the first version of the slide where Fermi was a 2009 product, I didn't think of a launch, but of an architectural presentation only.

The GPGPU Technology conference is mid May 2012. That would be a bit late for an architectural presentation, that's probably the latest we can expect the launch to happen.

 

[iMacmatician]

http://fudzilla.com/graphics/item/23323-some-nvidia-notebook-28nm-chips-are-finished

Nvidia had a big partner thing in Lisbon Portugal and some key people even talked about 28nm chips and have confirmed that some of Nvidia’s 28nm notebook chips are almost ready for production even today. Far East sources are confirming this independently.

The yields are the factor that will decide when wills this actually happen but it all indicates that 28nm notebook chips should be on time to launch with Ivy Bridge 22nm notebooks. Some of these chips might show up as a part of Sandy Bridge, but this will depend on yields on 28nm and Nvidia will try to calculate if it makes sense to launch it before.

The design notebook GPUs usually starts from the ground up, or roughly from the middle and faster chips are introduced at a later date. For example, Nvidia launched GF108 based GT525M, GT540M and GT 550M at CES of January 5th 2011 together with GF106 based Geforce GT 555 M while the Geforce GTX 560 performance part launched in very late May (Computex 2011) and GTX 580M (high end) launched in late June.

We expect a similar scenario with Geforce 600 series mobile launch scenario and to see the mainstream and entry level first and the performance and high end to follow in a few months later.

At least one of these 28nm chips is finished and it seems to be in good shape.

Funny thing is they don't say whether these chips are Fermi, Kepler, or some Fermi and some Kepler.

The one chip being finished part reminds me of this post. The other parts of the report and posts don't really fit unless the other chips are Kepler.

 

[Harison]

Where did you get that from? Those chips taped out in February and are set for a Q4 launch. That's the month the article you posted comes from btw. Just because 28nm isn't available for nVidia doesn't mean it's not available for everybody else.

That article doesnt mention AMD high-end 28nm launch date. Earlier this year it was expected Q3-4 '11, but lately I hear Q2 '12. Even if they would launch in Q4, it would be for testers and few lucky end-users only

"According to TSMC, wafers made using 28nm fabrication process will account for 2% - 3% of its revenues in Q4 2011"

If AMD could get hands on all those wafers, they could do a hard-launch, but they most likely will only get a fraction of it. Not just because many customers wants, but because Apple made a big 28nm order, and they will come first.

 

[LordEC911]

Even if they would launch in Q4, it would be for testers and few lucky end-users only

Come November...

 

[Alexko]

http://fudzilla.com/graphics/item/23323-some-nvidia-notebook-28nm-chips-are-finished

Funny thing is they don't say whether these chips are Fermi, Kepler, or some Fermi and some Kepler.

The one chip being finished part reminds me of this post. The other parts of the report and posts don't really fit unless the other chips are Kepler.

Given how it happened for 40nm, I'd say those chips are low-end Fermi-derivatives, assuming the report is correct, of course.

 

[Arun]

GF117 with the same specs as a GF116 but a 128-bit GDDR5 memory bus? It seems like the sweetspot for discrete notebook graphics to me and they're very unlikely to be pad limited there (GF116 is 238mm²!)
Personally I'd say it wouldn't be worth the trouble to come up with a full line-up since TSMC won't have the capacity for it anyway. They'd be better off accelerating Kepler derivatives.

 

[Harison]

Come November...

You expecting high-end SI 28nm launch in November? Lets say AMD would do it:

a) AMD most definitely wont have enough cards to satisfy demand, but would kill 69xx sales.
b) AMD already have the fastest card in the market.
c) NV wont launch Kepler till sometime 2012, no real competition for AMD.

From any point you look, it makes no sense of doing it. I personally expect mid-range pipe-cleaner with limited availability in 2H '11.

 

[Psycho]

Even if production constrained by TSMC it could make sense to sell high end parts at a high price, like 6970 +50% in both price and performance (which wouldn't kill 69xx sales either) .
And then wait with the rest - the high volume parts (maybe except for the early pipe cleaner) until production picks up.
But even with a (VLIW4?) pipe cleaner in September, I guess november is a bit optimistic for such a real high end.

 

[Squilliam]

Is there any chance GF 28nm will be ready for AMD by the end of the year? Maybe thats what they based their estimates on?

 

[Alexko]

You expecting high-end SI 28nm launch in November? Lets say AMD would do it:

a) AMD most definitely wont have enough cards to satisfy demand, but would kill 69xx sales.

How do you know that?

 

[Harison]

How do you know that?

Very low supply of 28nm, hence "AMD most definitely wont have enough cards to satisfy demand", plus since AMD would launch high-end SI to replace 69xx, it "would kill 69xx sales", and those are the only premium GPU's AMD could actually meet demand of usually the most profitable Q4.

Bottom line, launching premium SI cards this year would be like killing goose who lays golden eggs (69xx) Makes no sense whatsoever. If NV would launch Kepler in Q4, then its understandable soft (or paper) launch in a speed-race, but AMD wont face Kepler till sometime 2012, therefore no rush.

Psycho thinks AMD could launch SI with +50% price, I would doubt that as well, unless AMD pulls NV and starts selling cards for 1k.

 

[Alexko]

Very low supply of 28nm, hence "AMD most definitely wont have enough cards to satisfy demand", plus since AMD would launch high-end SI to replace 69xx, it "would kill 69xx sales", and those are the only premium GPU's AMD could actually meet demand of usually the most profitable Q4.

Bottom line, launching premium SI cards this year would be like killing goose who lays golden eggs (69xx) Makes no sense whatsoever. If NV would launch Kepler in Q4, then its understandable soft (or paper) launch in a speed-race, but AMD wont face Kepler till sometime 2012, therefore no rush.

Psycho thinks AMD could launch SI with +50% price, I would doubt that as well, unless AMD pulls NV and starts selling cards for 1k.

Yes, but how do you know that: "Very low supply of 28nm"?

Psycho may be pushing it, but if SI ends up, say, 30% faster than GF110, I don't see why AMD couldn't sell it for 20% more than the latter, which would be a good bit more than then can ask for Cayman.

 

[Harison]

Yes, but how do you know that: "Very low supply of 28nm"?

"Nvidia: 28nm Process Tech "Not Available" This Year."

"According to TSMC, wafers made using 28nm fabrication process will account for 2% - 3% of its revenues in Q4 2011"

If AMD could get hands on all those wafers, they could do a hard-launch, but they most likely will only get a fraction of it. Not just because many customers wants, but because Apple made a big 28nm order, and they will come first. NV usually could get a bigger supply slice than AMD, but this time even they gave up. What are the chances of AMD gaining majority of 28nm from TSMC? None. Much bigger sharks wants those wafers, and they will get it.

Global Foundries:

"process isn't expected to become mature enough until late next year."

"For example, AMD plans to use it for the upcoming Radeon HD 7000 family of graphics processors."

http://news.softpedia.com/news/GlobalFoundries-Tests-28nm-Chip-Production-210089.shtml

Bottom line: we wont see sufficient 28nm supply from TSMC till mid-2012 (even new refresh of Tablet APUs in Q2 2012 will be on 40nm!), and from GF we wont see 7xxx generation till the END of 2012.

 

[LordEC911]

"According to TSMC, wafers made using 28nm fabrication process will account for 2% - 3% of its revenues in Q4 2011"

Glad to see you don't look at any history...
http://www.xbitlabs.com/news/other/..._Revenues_Quadruple_in_the_Third_Quarter.html

4% of TSMC's revenue came from 40nm in Q3 2009 with RV740 already in production and RV8x0 ramping up plus the other 40nm contracts they had to other customers...
And that was with AMD only getting thousands of wafers a week, not the tens of thousands they were promised from TSMC, due to their ramp up problems.

Hopefully things go much smoother for TSMC's 28nm and they don't have the same problems they had with 40nm but even if they do, you shouldn't be so sure that AMD won't be launching GCN this year.

Edit- Word is Cayman and maybe Barts are set for EOL sometime in Q3... I wonder what that means?

 

[Harison]

Glad to see you don't look at any history...
http://www.xbitlabs.com/news/other/..._Revenues_Quadruple_in_the_Third_Quarter.html

4% of TSMC's revenue came from 40nm in Q3 2009 with RV740 already in production and RV8x0 ramping up plus the other 40nm contracts they had to other customers...
And that was with AMD only getting thousands of wafers a week, not the tens of thousands they were promised from TSMC, due to their ramp up problems.

Hopefully things go much smoother for TSMC's 28nm and they don't have the same problems they had with 40nm but even if they do, you shouldn't be so sure that AMD won't be launching GCN this year.

I remember 40nm history, and how much AMD and NV were burned by it. I would expect them to learn from it, and NV is showing it by postponing Kepler launch as we speak, but you think AMD havent learned?

Now the supply numbers you like:
http://www.xbitlabs.com/news/other/...m_Chips_Up_Slightly_Quarter_over_Quarter.html

Thats 9% 40nm in Q4 2009, and we have only 2% - 3% of 28nm in Q4 2011. Do you remember how many problems AMD had? Now picture this with 3-4x less supply, and big sharks like Apple getting majority of wafers. Not a pretty picture.

Word is Cayman and maybe Barts are set for EOL sometime in Q3... I wonder what that means?

If thats true, it would mean AMD is about to shoot in its leg, with double-pump gun, no less

In '09 GTX280 were the king, and it was expected Fermi would be launched in Q4, and AMD rushed, so they wouldnt be left far behind. Situation is different now. They have the fastest card, no Kepler in '11, TSMC wont be able to supply enough 28nm, but 40nm would, etc. You decide what makes more sense, I did.