NVIDIA Kepler speculation thread

[Psycho]

If and when the HD 7850 gets to $140 or below, I'm not sure how NVIDIA could respond.

Here the 7850/1G just went pretty cheap now, as in halfway between 6870/560 and 560ti, which all has stayed at the same prices for a looong time.
(6870/560 from 1200dkk, 7850/1G from 1350, 560ti from 1500, 7850/2G from 1600, 7870 from 1900, 660ti from 2300)

But one thing is having GK104 competing with 7850, I really doubt GK107 can reach any higher than 7750, so that's still quite a gap in a high volume segment.

 

[lanek]

Here the 7850/1G just went pretty cheap now, as in halfway between 6870/560 and 560ti, which all has stayed at the same prices for a looong time.
(6870/560 from 1200dkk, 7850/1G from 1350, 560ti from 1500, 7850/2G from 1600, 7870 from 1900, 660ti from 2300)

But one thing is having GK104 competing with 7850, I really doubt GK107 can reach any higher than 7750, so that's still quite a gap in a high volume segment.

I can imagine Nvidia continue to cripple the GK104 for do lower chips for take place in the 650 vs 7770... But wathever, there's a big gap between GK107 and GK104 ( and lets apart the cost of the chips itself for manufacture them). + they need to be able to supply all this line up.. 680-690 highly priced, not a problem.. 670-660TI-660 ... this could really start to be a problem, then if they add 650-650TI ...each time the proportion of shipment is increased. ( mid range sell more of high range in term of number of pieces, low range sell more of midrange ) + there's a last problem: thoses "discrete cards " for midrange notebook and OEM was still good sellers ( cause under, they hit APU and IGP ) ..

 

[jimbo75]

I suppose rebadging the 560 Ti and 560 to the 650 Ti and 650 is an option. Surely that's a better option than going down further than the 660 with GK104.

 

[Blazkowicz]

but, the 650 ti would eat more power and perform worse than the 660.
renamed 560 ti is "low" midrange performance, with the power use of a high end card.

 

[Alexko]

Apparently the GTX 660 (non-Ti) will be released next week. We'll find out for sure whether it's GK104- or GK106-based.

 

[xDxD]

http://videocardz.com/34621/nvidia-geforce-gtx-660s-kepler-gk106-gpu-pictured

 

[iMacmatician]

Assuming the two pictures are to scale (and going by the top grey pieces of each chips)…

The GK106 die is about 82.6% the size of the GK104 die (I got about 1.85" x 1.60" compared to 1.79" x 2.00" on Keynote measurements), to get about 243 mm^2.

I'm guessing 6 SMXes if true. I did a very simple chop job of the GK104 die and a 6 SMX GK104 ends up around 245 mm^2. (The aspect ratio on my chop is somewhat different though.)

My thoughts:

1. The larger number of SMXes of the GK106 compared to the geometric mean of the GK107's and GK104's SMX count might be a reason why the non-OEM GeForce GK104s didn't drop much in SMXes or clocks compared to the 680. (Although wouldn't keeping the memory bus at 256-bit for the 660 Ti give a larger distinction between it and a full GK106?)
2. What might this "SMX # anomaly" signal for GK114? Possibly 10 or 12 SMXes?

 

[Alexko]

Assuming the two pictures are to scale (and going by the top grey pieces of each chips)…

The GK106 die is about 82.6% the size of the GK104 die (I got about 1.85" x 1.60" compared to 1.79" x 2.00" on Keynote measurements), to get about 243 mm^2.

I'm guessing 6 SMXes if true. I did a very simple chop job of the GK104 die and a 6 SMX GK104 ends up around 245 mm^2. (The aspect ratio on my chop is somewhat different though.)

My thoughts:

1. The larger number of SMXes of the GK106 compared to the geometric mean of the GK107's and GK104's SMX count might be a reason why the non-OEM GeForce GK104s didn't drop much in SMXes or clocks compared to the 680. (Although wouldn't keeping the memory bus at 256-bit for the 660 Ti give a larger distinction between it and a full GK106?)
2. What might this "SMX # anomaly" signal for GK114? Possibly 10 or 12 SMXes?

The text says

I’m not sure if the size is correct, so take this comparison with a grain of salt.

Where I'm guessing "size" means "scale".

 

[iMacmatician]

From TechPowerup: "MSI GeForce GTX 660 Pictured, Spills The Beans on Specifications," although the link goes to a 404 page for me, the images are up.

960 CCs @ 980/1023 MHz, 192-bit GDDR5 @ 6.008 GHz, 2 GB, 170 W, note that this card is factory OCed.

 

[lanek]

Look legit, but make lie the 660OEM specification.

 

[A1xLLcqAgt0qc2RyMz0y]

Look legit, but make lie the 660OEM specification.

Not if the GTX660 (OEM) is GK104 based.

The GTX660 (OEM) specs are still available on www.geforce.com

http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-660-oem/specifications

 

[jimbo75]

170W TDP? That doesn't make sense unless it's GK104 with a large voltage bump.

 

[A1xLLcqAgt0qc2RyMz0y]

170W TDP? That doesn't make sense unless it's GK104 with a large voltage bump.

I agree the 170 watts do not make sense.

And it is not a GK104 as the MSI Specs shown in the link above clearly states GK106.

 

[Kaotik]

Well it could be a reason for all the delays, for whatever reason they just couldn't get it lower even if it doesn't make sense considering GK104

 

[tviceman]

Well it could be a reason for all the delays, for whatever reason they just couldn't get it lower even if it doesn't make sense considering GK104

That makes even less sense. It's 5/8 a GK104 shaders with 3/4 the memory bus and lower clock speeds. There is no way at the same clock speeds as a gtx680 this chip will consume as much power - it's the exact same technology and transistors, just significantly less of them. The slide probably just has the TDP wrong. I doubt TDP issues had anything to do with it's delayed release. GK107 obviously doesn't have TDP issues, GK104 doesn't either. More than likely GK107 and GK104 were heavily priotized, and GK106 only makes sense now to fabricate as demand is catching up to supply.

And also, 5 SMX units lines up with the early rumors of a ~205mm^2 chip that were floating around at around GK104's release. I don't think the videocardz link is accurate with it's 6 SMX specs prediction.

 

[madyasiwi]

170W TDP? That doesn't make sense unless it's GK104 with a large voltage bump.

It's a factory OCed card

MSI R7870 Hawk TDP rated at 200watts for comparison:
http://media.msi.com/main.php?g2_view=core.DownloadItem&g2_itemId=76172

 

[Psycho]

Usually the factory oc cards are just showing the same tdp as the reference clocked - even though they obviously use more power (unless it's gk104 based and power capped to the same - then it's just not boosting as much)

More than likely GK107 and GK104 were heavily priotized, and GK106 only makes sense now to fabricate as demand is catching up to supply.

So you don't think there has been demand for $150-250 cards since march?

 

[AnarchX]

Could be ~240mm^2 GK106 with probably 1152SPs/192-bit an indication for a stronger gamer/performance Kepler (>1536SPs/>256-bit) in early 2013?

 

[iMacmatician]

gdm.or.jp (translated) says that the upcoming GTX 650 (GK107) will have 1058 MHz base clock and 1 GB 5000 MHz GDDR5, and it and the GTX 660 (GK106) will be released on September 6.

IT168 (translated) has a spec table and a few benchmarks of the GTX 660 (GK106) and the GTX 650. Interestingly the 660's specs in the table are about the same as those of the apparently OCed MSI card above: 980 core, 1033 boost, and 6000 memory. The TDP is 140 W though. The table also shows that the GTX 650 doesn't have boost.

IT168 also mentions this interesting tidbit of speculation in their report:

And we believe that the GK106 will no longer use the GK104 the SMX unit in combination, but to allow the use of more efficient SMX unit combination (reference GTX460 (GF104 core) before and GTX480 (GF100 core)).

 

[AnarchX]

Even GK110 has 192SPs SMX. A change for GK106 seems highly unlikely.