NVIDIA Kepler speculation thread

[UniversalTruth]

Because waiting means less time alone on market for a product they could price with high margins. You can't really sit on high tech, there's a window to make money on these things.

Generally speaking, it's true but.. The 28 nm manufacturing process will be here (unfortunately) with us till early 2014 (I assume this reading the other threads here). So, they have plenty of time to milk customers first with GK104 as GTX680, and then with the big one. I mean the window is very widely open.

 

[lanek]

ZOTAC GeForce GTX going to release 680 overclocked to 2 GHz

yeah,.... well most big overclockers dont believe one second they will succeed... Kingpin have never been able to maintain 2000mhz, with an addon card and external VRM at -165°C + extremely high voltage.

Let see the cooling solution they have:

- Watercooling alone will never be enough
- Chilled water, will take too much places ( a case )
- Oil box.. ( well i dont really know what is the impact with it, i know it can replace air cooling, but subzero i doubt it )
- Tec + watercooling.. ( huum possible, but need a 450W PSU for the TEC alone )
- Phase change unit.. ( the only one i can see is something similar of my old XEII Vapochill who is inside a case ( well its the case of the pc for be honest ).
- Ln2 and Dry ICe are completely exclude ofc.

- Some cooling system used by the NASA ?

But i like dream, and if they really succeed to it.. I will be the first to applaude the exploit .

 

[AlphaWolf]

Generally speaking, it's true but.. The 28 nm manufacturing process will be here (unfortunately) with us till early 2014 (I assume this reading the other threads here). So, they have plenty of time to milk customers first with GK104 as GTX680, and then with the big one. I mean the window is very widely open.

If you believe the competition hasn't learned anything from their first effort. Or do you believe that Kepler is the best possible implementation of 28nm? I find that highly unlikely and even disputable right now.

 

[AlexV]

especially opencl.

NVIDIA isn't overly fond of OpenCL (and for reasonable reasons, IMHO), so optimising their CL stack or fixing whatever bugs might've been introduced along the way may not be top priority. I think one update at some point borked something wrt CL that's yet to be unborked in NV's stack (hope my failing memory does not fail now too!).

 

[MfA]

- Phase change unit.. ( the only one i can see is something similar of my old XEII Vapochill who is inside a case ( well its the case of the pc for be honest ).

Standard ones would only get you slightly below zero of course, which isn't that big a difference. In theory a mixed refrigerant compressor system can relatively cheaply cool to LN2 temperatures (cryocoolers are more efficient but would cost you in the area of 10K$). It would have to be a bigger compressor than those vapochills (probably around 1KW) and you'd still have to deal with condensation.

 

[lanek]

Standard ones would only get you slightly below zero of course, which isn't that big a difference. In theory a mixed refrigerant compressor system can relatively cheaply cool to LN2 temperatures (cryocoolers are more efficient but would cost you in the area of 10K$). It would have to be a bigger compressor than those vapochills (probably around 1KW) and you'd still have to deal with condensation.

This is the problem and the reason i was speak about the XEII ( who is a Vapochill munted in 1/4 of a big tower case ( a bit of joke of how much of place a unit capable of -40-60°C on full load can take )

If you refer to the little cpu cooler have made Asetek, they have never been really good. not even better of the best air cooler you can find. They are just not made for be able to work with more of a certain amount of watts, and some aircooler are capable of more on this sense.

Lets be honest, i really dont see how they will push a card who have not been able to hit 2ghz under LN2 with less extreme cooling solution...

 

[Alexko]

NVIDIA isn't overly fond of OpenCL (and for reasonable reasons, IMHO)

Are you referring to strategic or technical reasons?

 

[Megadrive1988]

Slightly off-topic:

"Nvidia deeply unhappy with TSMC, claims 22nm essentially worthless".


"One of the unspoken rules of customer-foundry relations is that you virtually never see the former speak poorly of the latter. Only when things have seriously hit the fan do partners like AMD or Nvidia admit to manufacturing problems, and typically only after postponed launches and poor availability have made protestations that everything is fine unsustainable.

That’s why we were surprised — and our source testified to being stunned — that Nvidia gave the following presentation at the International Trade Partner Conference (ITPC) forum last November.

According to Nvidia, the current model is unsustainable. What this slide states — we can’t even call it a suggestion — is that smaller processes no longer improve yields by leading to a greater number of chips per wafer. Instead, the complexities and difficulties of manufacturing at the new process create a cost structure that provides precious little incentive to manufacture at the new process. "

source (with slides): http://www.extremetech.com/computin...y-with-tsmc-claims-22nm-essentially-worthless

 

[Mize]

Slightly off-topic:





source (with slides): http://www.extremetech.com/computin...y-with-tsmc-claims-22nm-essentially-worthless

Well, they're talking about useless as far as $/transistor, not $/perf. or absolute perf., but it is an curious thing to indirectly call out one's partner like that.

 

[BRiT]

Already being discussed in this thread: http://forum.beyond3d.com/showthread.php?t=61700

 

[A1xLLcqAgt0qc2RyMz0y]

BigK when 28nm yields improve and 28nm supply expands

If they could release 'BigK' they would. Simple as that.

The could have but with 28nm yields sucking right now and not enough 28nm wafer capacity it would not have been a good financial decision.

#1 The price for each good GK110 die would have a very high cost

#2 Producing the GK110 would eat into the supply of all other 28nm parts (GK104, GK106/7, 28nm Fermi).

Nvidia did the right thing by delaying BigK and will release it when 28nm yields improve and 28nm supply expands.

 

[cal_guy]

However launching GK110 now would allow nVidia to start the long process of qualifying GK110 for compute and professional applications.

 

[Jawed]

Why does TechReport say this:

The organization of the SMX's execution units isn't truly apparent in the diagram above. Although Nvidia likes to talk about them as individual "cores," the ALUs are actually grouped into execution units of varying widths. In the SMX, there are four 16-ALU-wide vector execution units and four 32-wide units. Each of the four schedulers in the diagram above is associated with one vec16 unit and one vec32 unit. There are eight special function units per scheduler to handle, well, special math functions like transcendentals and interpolation. (Incidentally, the partial use of vec32 units is apparently how the GF114 got to have 48 ALUs in its SM, a detail Alben let slip that we hadn't realized before.)

http://techreport.com/articles.x/22653

 

[fellix]

Bah, that's nothing. Look what Anand have figured out:

GK104 SMX Functional Units

• 32 CUDA cores (#1)
• 32 CUDA cores (#2)
• 32 CUDA cores (#3)
• 32 CUDA cores (#4)
• 32 CUDA cores (#5)
• 32 CUDA cores (#6)
• 16 Load/Store Units (#1)
• 16 Load/Store Units (#2)
• 16 Interpolation SFUs (#1)
• 16 Interpolation SFUs (#2)
• 16 Special Function SFUs (#1)
• 16 Special Function SFUs (#2)
• 8 Texture Units (#1)
• 8 Texture Units (#2)
• 8 CUDA FP64 cores

Everybody else pretty much insists on the 12*16-lane SIMDs. Silly season...

 

[Alexko]

The could have but with 28nm yields sucking right now and not enough 28nm wafer capacity it would not have been a good decision.

The Price for each good GK110 die would have a very high cost and second producing the GK110 would eat into the supply of all other 28nm parts (GK104, GK106/7, 28nm Fermi).

Nvidia did the right thing by delaying BigK and will release it when 28nm yields improve and 28nm supply expands.

The earliest report of tape-out I've seen for GK110 is December 2011, which would not make a March release possible. In fact even June would be unlikely.

 

[KimB]

Mmmmmm....just read through AnandTech's review. Low power consumption. High performance. This is exactly what I want. One or two of these cards will be perfect for my new PC in a month or so...plus I think I'll also go for three 23" IPS displays....yesssss...

 

[denev2004]

Bah, that's nothing. Look what Anand have figured out:





Everybody else pretty much insists on the 12*16-lane SIMDs. Silly season...

It not really silly....I guess it to be different opinion

Is there any more academical or official reference about the structure within SMX?

 

[Alexko]

Mmmmmm....just read through AnandTech's review. Low power consumption. High performance. This is exactly what I want. One or two of these cards will be perfect for my new PC in a month or so...plus I think I'll also go for three 23" IPS displays....yesssss...

At 5760×1080, it's a tie with the 7970: http://www.hardware.fr/articles/857-22/recapitulatif-performances.html

 

[Sinistar]

Mmmmmm....just read through AnandTech's review. Low power consumption. High performance. This is exactly what I want. One or two of these cards will be perfect for my new PC in a month or so...plus I think I'll also go for three 23" IPS displays....yesssss...

Why is that important for you now, high power consumption, low performance per watt never bothered you before.

Personally, I'll be getting 2 HD 7970. I waited to see what Nvidia had, and I am not impressed.

I'm still trying to figure out who pays $500 dollars for a card and plays at 1080P or less!

 

[Mize]

Mmmmmm....just read through AnandTech's review. Low power consumption. High performance. This is exactly what I want. One or two of these cards will be perfect for my new PC in a month or so...plus I think I'll also go for three 23" IPS displays....yesssss...

Wait.
If you run 3x 19x12 you should wait for a 3 or 4 GB version.