Alleged NVIDIA GK104 Specifications

[tweakboy]

Alleged NVIDIA GK104 Specifications

FYI, you need google chrome to translate this page from german to english. If you dont have google chrome wel shame on you...

 

[tweakboy]

According to the information R1000/Tahiti-Chip come from our forum now to the first piece of information GK104 nVidia chip performance of the 28nm generation. The first information we use is that nVidia wants to introduce this chip before the high-end chip GK100 - this increases the likelihood that the GK104 still on the Fermi-based architecture or take only certain bonds issued by the Kepler architecture. Nvidia is aiming for the GK104 but also a performance about half of the GeForce GTX 580 in what could very well go in the power range of R1000/Tahiti-Chips - depending on how much this turns out in practice. Performance assessment of GK104, there is also a specific Information: This is a performance of "well over 2 TFLOPs". As a GeForce GTX 560 Ti as the current performance solution from nVidia comes to 1.26 teraflops, a GeForce GTX 580 to 1.58 teraflops, however, this would be a significant increase in computing power at the chip-GK104.
Unless the system of nVidia "Hotclocks" (the shader units clocked much higher than the rest of the chip) and maintains the frequencies below the 28nm production not to be materially different than before(rather unlikely) , this is almost exclusively with a high standard nearly doubling the number of shader units available. At 768 (1D) shader units and a shader clock of 1600 MHz would thus achieved a computing power of 2.46 teraflops, this would be within the scope of this prediction. Conceivable, of course, are other models, such as 640 shader units with slightly higher clock rates, or 704 shader units at 1600MHz shader clock - in both cases you can get clear about two teraflops of computing power. Probably used for the nVidia chip GK104 then also the basic model of the shader clusters of GF100/GF110-Chips : This allows that one installed only 80-96 textures units in this high-performance chip - not the same as with the 112-128 shader clusters of GF104/GF114-Chips .
In the performance category of the GK104, you need not have such a high Texturierpower, moreover, this also saves transistors. The real from the 2 teraflops and more question arising is whether this is not a bigger memory interface is needed than was previously at nVidia in performance area set botene 256-bit DDR interface. In a performance well above half of the GeForce GTX 580, we would affirm this fact - on the other hand, nVidia also need a certain distance from the high-end chip, and also wants the GK100 GK104-They will not be too large. Consequently, it is vacant, whether the GK104 nVidia now has a 256 or 384-bit DDR memory interface built - both strategies are possible. The basic purpose of a performance just over half of the GeForce GTX 580 would, moreover, with "only" 640 unified shaders, 80 TMUs, and 256-bit DDR memory interface to be available, because in this model sufficiently more computing power and Texturierpower available to despite only slightly more memory bandwidth (only achieved by a higher memory clock) , the GeForce GTX beat 580th

 

[tweakboy]

Fermi-Refresh (GeForce 500 Series)Kepler (GeForce 600 Series)GF110 (GeForce GTX 570/580/590)
3 billion transistors on 530mm ² die area in 40nm
Fermi architecture with 4 Raster Engines
16 Engines polymorph tesselations with 16 units
512 (1D) shader units,
64 texture units (TMUs)
48 raster operation Units (ROPs)
384-bit DDR memory interface (up to GDDR5)
maximum clock chip 772 MHz, Shader 1544 MHz, 2000 MHz memory
maximum computing performance: 1.58 TFLOPS (SP)
Maximum Memory Bandwidth: 192 GB / secGK100
Kepler architecture estimated 1024 (1D) shader units estimated 128 texture units (TMUs)estimated 64 raster operation units (ROPs), estimated 512-bit DDR memory interface (up to GDDR5) Performance: still too uncertain for a forecast, but certainly about half the level of GeForce GTX 590 Launch: second quarter of 2012





GF114 (GeForce GTX 560/Ti)
1.95 billion transistors on a 358mm ² die area in 40nm
Fermi architecture with two raster engines
8 engines with 8 polymorph tesselations units
384 (1D) shader units,
64 texture units (TMUs)
32 raster operation Units (ROPs)
256-bit DDR memory interface (up to GDDR5)
maximum clock chip 822 MHz, Shader 1644 MHz, 2000 MHz memory
maximum computing performance: 1.26 TFLOPS (SP)
Maximum Memory Bandwidth: 128 GB / secGK104
supposedly slightly smaller die size than the GF114
unknown architecture, whether Kepler, Fermi or Kepler / Fermi mix adopted 640-768 (1D) shader units approved 80-96 texture units (TMUs) Memory Interface unclear, 256 to 384-bit DDR are possible maximum computing performance: rumors "well over 2 TFLOPs"Performance: allegedly a bit faster than the GeForce GTX 580 Launch: first initial / middle quarter of 2012

 

[dangerman1337]

Translated version http://translate.google.com/transla...tp://www.3dcenter.org/news/2011-12-13&act=url

Sounds extremely fake, 768 shaders for the GK104 yet 'barely' better than 580? Also GK100 being 'half' wouldn't that put it in 580-ish levels?

Also 512 bit-bus and 1024 for the GK100? Looking at 280/285 > 480/580 in power consumption and that was reducing the 512 bit bus to a 384 bit bus for Fermi even considering that was a shrink and it ended up pretty hot and power consuming. If the specs are true then I'll be surprised if it consumes less than the 480 at release or even be smaller.

 

[Vesku]

Perhaps they will phase out the 580 and replace it with the GK104, assuming it performs at least a little better.

 

[maniac5999]

FYI, you need google chrome to translate this page from german to english. If you dont have google chrome wel shame on you...

Du kannst nicht Deutsch gesprachen?

Auch, Ich liebe Firefox.

 

[ocre]

I heard a good while ago about the fermi/kepler hybrid chip.


I have been waiting for some leaks on this chip. This is the first time i seen anything. Its not gonna be a full kepler, the full kepler is a big step (it will be very very powerful). This is a small step in the right direction. It should easily beat out the gtx 580. this is most likely why nvidia wanted to dump gf110 stock fast and came out wit 448sp 560 ti's. Nvidia could really do something great with this design. They could pile on the cores to make a very powerful version that uses 200+ watts. This would give AMD a good run for the money until they ready the full kepler.

 

[tviceman]

768 cores sounds really high for the performance segment part, if this is at all true. 768 cores for a performance chip would be an absolute monster - especially if they keep the die size similar to or smaller than GF114 (which the article says). If this is at all true, wow......

 

[Vesku]

If the full Kepler high end part is not coming until late Q2 early Q3 then this would be NVIDIA's 6970 type of part meant to bridge the gap.

 

[ocre]

i know there will be a fermi gpu with some kepler features. its more like fermi than not. I also know kepler is gonna be great. a lot more capable than a fermi dieshrink. Also there are rumors of the gk104 being high in power usage. Many have heard this. put this together and it seems Nvidia may have took the 28nm shrink in steps. A chip that is more fermi than kepler is less risky. It seems very logical to do this in steps. And i already heard they have silicon that is more fermi than kepler.

Why not push the limits. shoot for the extreme. worse case they shut off some of the cores and still have a decent card. Best case they have a killer GPU line to compete with AMD until the full kepler can be realized.

Kepler will take nvidia a step further in their future of gpu/compute supercards. its a pretty important step. This fermi/kepler hybrid likely will not be in their tesla brand. Nvidia sees AMDs direction with their core design. But nvidia is a good bit ahead. AMD is trying to play catch up. Nvidia has the advantage now but they must move fast for many many reasons. but being to ambitious could hurt them, again.


But why put all the eggs in one basket. Nvidia already has a powerful design that could be improved tremendously with a small investment. Fermi's little gem stone. Everyone wondered what would results if nvidia built a powerful GPU design based off the gf104/114. Its modular design makes it a perfect candidate for a very very interesting graphics powerhouse. Whats even more appealing is its a perfect route for nvidia to perfect the 28nm process and get kepler right the first time. Think about it. Its a win win win. No matter what, nvidia will better their 28nm skills, and have a very capable contender. a design that could be easily scaled as needed. Its a really good route.

I am speculating a lot, sure. But i have heard a few things. I am putting a lot of my own thinking into it. Its a very smart thing, and if nvidia doesnt end up doing it this way, they missed a great chance. Its got to be something, i have heard of this fermi chip with a few kepler attributes. I know its purpose was to perfect the 28nm node. I have waited till some leaks came out before i wanted to talk about it here. you know.

its a really neat plan of action! i mean, my speculation!

 

[ocre]

768 cores sounds really high for the performance segment part, if this is at all true. 768 cores for a performance chip would be an absolute monster - especially if they keep the die size similar to or smaller than GF114 (which the article says). If this is at all true, wow......

There was some rumors a while back on nvidia making a gf104 based monster chip. I believe there was early 580 rumors leaked saying it would be a gf104 based gpu with 576sp. other leaks had even more cores. Well i know the 580 rumor was based on a real engineering GPU. The mishap was assuming it was the 580. Bits an pieces get caught in the wind, then a lot of speculation fills in the gaps. a lot of times rumors are based on something real. A lot goes on behind the scenes

 

[tviceman]

There was some rumors a while back on nvidia making a gf104 based monster chip. I believe there was early 580 rumors leaked saying it would be a gf104 based gpu with 576sp. other leaks had even more cores. Well i know the 580 rumor was based on a real engineering GPU. The mishap was assuming it was the 580. Bits an pieces get caught in the wind, then a lot of speculation fills in the gaps. a lot of times rumors are based on something real. A lot goes on behind the scenes

Well my question is, on the assumption that there is some truth to this rumor - will GK104 have more in common with GF114 or will it be more like GF110 core? Being a "performance part" and not a GPU that will likely be aimed at Tesla, my guess would be it's more like GF114 with superscalar cores. Either way, if this is all true it will be a performance monster for not being a high end solution and should easily beat out AMD's Pitcarn (hd7870, hd7850) while probably also beating Tahiti's 2nd tier option (hd7950) soundly.

 

[RussianSensation]

This rumor is almost too good to be true. Based on this rumor, the GF104 chip will have 768 CUDA cores, which is 50% more than GTX580, and yet that's not even the flagship GK100 Kepler part....so where would that put Kepler's performance at 2x GTX580? 2012 is shaping up to be an interesting year for GPUs, after one of the most boring years for GPUs. Can't wait.

 

[tviceman]

This rumor is almost too good to be true.

That's almost what I'm thinking. Especially since rumor also says the die size is equal to or smaller the GF114. Then again, it has been 2 years since the fermi core was basically finalized, so who knows. I expect Kepler to be a monster in the performance department though.

 

[Will Robinson]

i know there will be a fermi gpu with some kepler features. its more like fermi than not. I also know kepler is gonna be great. a lot more capable than a fermi dieshrink. Also there are rumors of the gk104 being high in power usage. Many have heard this. put this together and it seems Nvidia may have took the 28nm shrink in steps. A chip that is more fermi than kepler is less risky. It seems very logical to do this in steps. And i already heard they have silicon that is more fermi than kepler.

Why not push the limits. shoot for the extreme. worse case they shut off some of the cores and still have a decent card. Best case they have a killer GPU line to compete with AMD until the full kepler can be realized.

Kepler will take nvidia a step further in their future of gpu/compute supercards. its a pretty important step. This fermi/kepler hybrid likely will not be in their tesla brand. Nvidia sees AMDs direction with their core design. But nvidia is a good bit ahead. AMD is trying to play catch up. Nvidia has the advantage now but they must move fast for many many reasons. but being to ambitious could hurt them, again.

Umm...no.
AMD are the current performance per watt leaders.
NVDA graphics chips are bigger,hotter and use more power.:whiste:

 

[thilanliyan]

Du kannst nicht Deutsch gesprachen?

Auch, Ich liebe Firefox.

Ja, das ist richtig.
Ich liebe Chrome nicht. (never know where to put the nicht lol...I can put it before or after "Chrome" right?)

 

[Arkadrel]

Launch: second quarter of 2012

Does that mean ~June or later?
Then AMD might just have ~6 months lead time with its series.

That said... man those are some insane specs.

estimated 1024 (1D) shader units (double up)
estimated 128 textures units (TMUs) (double up)
estimated 64 raster operation units (ROPs) (+33&#37

estimated 512-bit DDR memory interface (up to GDDR5)

Sweet mother of unholy gods! 512bit bus
That chip is gonna be HUGGEEEE... but a crazy beast of a thing.

Power of this card? probably around 590 levels.


Im with others on this, rumor sounds abit toooo crazy, most likely just rumor to be proved untrue.

 

[Genx87]

This rumor is almost too good to be true. Based on this rumor, the GF104 chip will have 768 CUDA cores, which is 50% more than GTX580, and yet that's not even the flagship GK100 Kepler part....so where would that put Kepler's performance at 2x GTX580? 2012 is shaping up to be an interesting year for GPUs, after one of the most boring years for GPUs. Can't wait.

I would expect the high end kepler part to perform like the GTX590.

 

[Arkadrel]

If true its gonna be a huge chip... probably bigger than 520mm^2 that 580 was.

even if 40nm->28nm is a huge step in shrinkage, the fact that so many things get doubled up, along with bus width = its bound to be bigger than the 580 was.

huge chip -> pricey to buy.

 

[tviceman]

NVDA graphics chips are bigger,hotter and use more power.:whiste:

And they're faster and preferred by 60% of PC gamers. :whiste: :biggrin:

 

[tviceman]

If true its gonna be a huge chip... probably bigger than 520mm^2 that 580 was.

even if 40nm->28nm is a huge step in shrinkage, the fact that so many things get doubled up, along with bus width = its bound to be bigger than the 580 was.

huge chip -> pricey to buy.

For reference. a 55nm GT200b chip is ~480mm^2. Fermi is ~520mm^2, more than doubled the core count, but reduced the memory interface. Going from 55nm to 40nm is technically a little bit smaller reduction than going from 40nm to 28nm (27% smaller node process vs. 30% smaller).

I don't think Nvidia is going to make a chip so large that they have to price it beyond affordability. Since there is sooo little information on Kepler right now, everything - even these rumors - are pure speculation. But I can say that I highly doubt Nvidia is going to make another G80 sized die (576 mm^2). I expect the high end Kepler chip to be around the same size of Fermi's high end chip. If they did anything to optimize their cores a little bit in size (even if only by 5%), then there is room to double the core count while staying very close to Fermi's die size. But again, pure speculation. I imagine that around the hd7000 release date, we are going to get some more substantive Kepler rumors.

 

[ocre]

Well my question is, on the assumption that there is some truth to this rumor - will GK104 have more in common with GF114 or will it be more like GF110 core? Being a "performance part" and not a GPU that will likely be aimed at Tesla, my guess would be it's more like GF114 with superscalar cores. Either way, if this is all true it will be a performance monster for not being a high end solution and should easily beat out AMD's Pitcarn (hd7870, hd7850) while probably also beating Tahiti's 2nd tier option (hd7950) soundly.

it would be the gf114. And i am hearing something performance like +30% of the 580. This would put it up there with AMDs 79XX series. But this is from a toss up of information. I cannot verify it at this point. Both nvidia and AMD are being very very quit. I can tell we are getting pretty close......

i dont expect AMD availability until later in January at the earliest. If we are lucky we can buy them by February, I cannot wait to see the 28nm stuff come out. Its been so long, i am very excited.

Umm...no.
AMD are the current performance per watt leaders.
NVDA graphics chips are bigger,hotter and use more power.:whiste:

i meant in compute. AMD's core next is playing catch up in the fermi direction. I actually cannot wait to see what they accomplish with it. I owned a 5850 for almost a yr. AMDs 6000series just didnt seem to much of an improvement over their last gen cards. I cant wait to see AMDs next big evolution. I just dont want to set expectations too high. My interest is very high!

 

[Zed03]

And they're faster and preferred by 60% of PC gamers. :whiste: :biggrin:

Can you show me benchmark where nVidia beats an AMD card in the same price range?

I don't think I've seen one since 2008.

 

[Silverforce11]

If its Q2 2012 launch, why even bother with leak specs, its purely speculation at this time.

Even with Tahiti's multiple-confirmed "rumor" of an early Jan launch, leaked specs we're getting is inconsistent.

 

[tviceman]

it would be the gf114. And i am hearing something performance like +30% of the 580. This would put it up there with AMDs 79XX series. But this is from a toss up of information. I cannot verify it at this point. Both nvidia and AMD are being very very quit. I can tell we are getting pretty close......

i dont expect AMD availability until later in January at the earliest. If we are lucky we can buy them by February, I cannot wait to see the 28nm stuff come out. Its been so long, i am very excited.

If that is true, and with what AMD's rumored pricing is going to be, AMD is going to have a very, very short ride. Nvidia will stomp all over them in the high end and performance segment. The only real traction AMD will have is with whatever they do with the barts shrink.