NVIDIA Kepler speculation thread

[iMacmatician]

More from EXPreview (translated): "Sometimes coincidence: the GK104 graphics inadvertently specifications."

…
According to the "insiders" said that the GK104 graphics core is indeed 1536 CUDA the Core, corresponding to the product configuration, 256-bit 2GB GDDR5 memory and core frequency 700MHz or so, the specific performance to be tested. From the number of CUDA Core and memory specifications, the news spread with the current online basically the same, only the core frequency has been more definite figures.
…

700 MHz? That's in contrast to earlier rumors of ~950 MHz…

 

[Man from Atlantis]

maybe it is GK104-335 (1344CCs).. GTX560Ti base clock was lower too, but many models are clocked ~880-900MHz.. maybe nvidia wants to keep TDP rather low ~170W.. earlier rumors similar die size with Tahiti and 950MHz@180W tdp was unrealistic IMO..

_____

Fudzilla - Kepler GK104 cards in production

We got word that Nvidia finally talked to its partners about its highly confidential Kepler products.

They got briefed, they got the reference design and the production of the cards is staring as we speak. It looks like most of the cards available at launch are going to look alike, with different cooling from some partners, but the non-reference partner cards will come a bit later.

 

[mczak]

Base clock of 560Ti was 810Mhz though, and no GF114 products are afaik shipping with less than 776Mhz. Even the second grade GF110 chips are clocked at 732Mhz (GTX570). Not quite so high clock than rumors suggested for power concerns wouldn't be that surprising, but only 700Mhz imho would be. Plus I thought that removing hotclock should potentially make it slightly easier to achieve higher clocks.

 

[Man from Atlantis]

I dont speak a single word of Chinese but maybe author tries to say 700ish MHz, could be 790MHz

790MHz GK104-335
840MHz Oc models..
840MHz GK104-400
900MHz Oc models..

 

[Rangers]

I dont speak a single word of Chinese but maybe author tries to say 700ish MHz, could be 790MHz

790MHz GK104-335
840MHz Oc models..
840MHz GK104-400
900MHz Oc models..

So out of text stating 700 mhz-ish, you somehow get =900 mhz. Great

 

[Man from Atlantis]

So out of text stating 700 mhz-ish, you somehow get =900 mhz. Great

i could stop at 890 MHz seriously though imagine MSI Lighting @900MHz-ish @225W TDP
155W, 780MHz, 2GB VRAM GK104-335 ~GTX580 -5% performance would make everyone happy.. great oc headroom too..

 

[neliz]

i could stop at 890 MHz seriously though imagine MSI Lighting @900MHz-ish @225W TDP
155W, 780MHz, 2GB VRAM GK104-335 ~GTX580 -5% performance would make everyone happy.. great oc headroom too..

We're making a GK104 Lightning?

 

[Man from Atlantis]

We're making a GK104 Lightning?

uh you haven't heard yet then
_____

http://vr-zone.com/articles/nvidia-kepler-power-circuitry-revealed-300w-tdp-/15011.html

A1 silicon, 8+6pin 5 phase pwm etc..

Spoiler

 

[neliz]

I always love this kind of confidence:

"but it is clear that this might be another winner for NVIDIA. Or… not."
Clear ... might be... or not, from 100 to 0 in 10 words.

Anyway, we have rules for what kind of product can actually "become" a Lightning.

 

[Man from Atlantis]

Anyway, we have rules for what kind of product can actually "become" a Lightning.

I guess you make Lightnings only for top GPU and top SKU (e.g. GF110 GTX580).. so GK104 is not top GPU in 600 series?? when do you expect GK110 in august? september? or you'll do it for GK104x2 ? or it is really power efficient gpu that it doesnt need that fancy coolers and pwm hence you'll enjoy with TFIII instead

 

[Ailuros]

You don't seriously expect him to answer all those questions do you?

 

[Man from Atlantis]

You don't seriously expect him to answer all those questions do you?

well i know some of them, asking for the sake of double confirm
but he has his own way to express himself RG: :smile:

 

[mczak]

It can't really be top sku only since there were lightning's for both gtx 260 and 275 neither of which was the top sku (both were the top chip though). Newer ones seem to use top sku however.

 

[Psycho]

Lightnings are only for top SKUs (well, except for the 275) - otherwise it'll just be twin frozr etc.
So that kinda rules out GK104-400 as GTX680 (so it's probably either too slow for the name (ie below 7970), or GK110 isn't far enough away to be called 780)


...
300W TDP for some kepler board: http://vr-zone.com/articles/nvidia-kepler-power-circuitry-revealed-300w-tdp-/15011.html
(but is it for GK104 - and in that cas just some GK104 engineering board with plenty of power?)

 

[Ailuros]

Probably just a GK104 engineering sample. I severely doubt that GK110 taped out in Q3 11' already.

 

[Man from Atlantis]

Three phases are dedicated to feed the GPU (NVVDD-PH2, PH4, PH5), and dual phase is for driving the rest of the board - a couple of gigabytes of GDDR5 memory and 3rd party chips (NVVDD-PH1 and PH3).

i thought NVVDD is only for GPU and there should be FBVDD for memory and off chip, it doesnt make sense 2 phases are dedicated for memory esp for such low amount memory and speed, and we know there is no nvio chips etc to increase offchip power req..

 

[iMacmatician]

300W TDP for some kepler board: http://vr-zone.com/articles/nvidia-kepler-power-circuitry-revealed-300w-tdp-/15011.html
(but is it for GK104 - and in that cas just some GK104 engineering board with plenty of power?)

If that's the case it may fit with the 4Gamer report last year that contained info of ~250 W GK104—more recent report(s) suggest closer to ~170 W.

Maybe NVIDIA planned ~950 MHz for the (top) GK104 but had to drop clocks because of power/yield or if they wanted to hit a lower TDP.

Or there's two cards using GK104-400, one at ~700 MHz and ~170 W to go between Pitcairn and Tahiti, and another at ~950 MHz and ≥225 W (so 8 + 6 pin) to go up against Tahiti, at least until GK110 shows up. Most likely not though.

 

[ninelven]

I don't quite agree though with CarstenS that it wouldn't be impressive for a successor of gf104. I think more than doubling alu capacity is indeed quite impressive (if they can achieve the clocks do actually exceed the factor 2).

Indeed, Tahiti achieves ~1.5x FLOPs / mm^2 compared to Cayman (but it is also more efficient so more like 1.7x).

GK104 with 1536 CCs and 360 mm^2 die size would achieve vs GF114:

~2x FLOPs / mm^2 @ 775 MHz
~2.3x FLOPs / mm^2 @ 925 MHz
~2.5x FLOPs / mm^2 @ 1 GHz

 

[Psycho]

GK104 with 1536 CCs and 360 mm^2 die size would achieve vs GF114:

..But probably less "efficient" due to rumored double SIMD length at 96 (3 times GF100..) and no SFU.

 

[whitetiger]

Indeed, Tahiti achieves ~1.5x FLOPs / mm^2 compared to Cayman (but it is also more efficient so more like 1.7x).

GK104 with 1536 CCs and 360 mm^2 die size would achieve vs GF114:

~2x FLOPs / mm^2 @ 775 MHz
~2.3x FLOPs / mm^2 @ 925 MHz
~2.5x FLOPs / mm^2 @ 1 GHz

Whoooaa!
Think you might be in danger of exceeding the official speculation limit there, buddy!


The forum thought-police will be on you like a ton of bricks if you're not careful!