Tuesday, December 13th 2011

GeForce Kepler 104 and 100 GPU Specifications Compiled

A quick stroll through our previous article about how the GeForce Kepler family of next-generation GPUs is laid out, would tell you that GeForce Kepler 104 (GK104), is going to be NVIDIA's answer to AMD's Tahiti. GK104 will be a high-performance (≠ high-end) GPU by NVIDIA that will have many of the features that were reserved for its previous high-end GPUs (such as a 384-bit wide GDDR5 memory interface), but will not be NVIDIA's most powerful GPU in the series. The throne will be kept empty for GK100, which will comply with NVIDIA's "go all in" design ideology for high-end GPUs.

3DCenter.org compiled a few specifications of the GK104 and GK100. They go like this:
GK104
  • 640 to 768 CUDA cores
  • 80 to 96 TMUs (depending on what the CUDA core count ends up being)
  • 384-bit GDDR5 memory interface, 48 ROPs
  • Built on the 28 nm TSMC process
  • Products based on this will launch in the first quarter of 2012


GK100
  • 1024 CUDA cores
  • 128 TMUs
  • 512-bit GDDR5 memory interface, 64 ROPs
If anything, this specifications disclosure is sure to have an impact on buyers saving up for AMD's HD 7900 series products that are heading toward launch on the 9th of January.Source: 3DCenter.org
Add your own comment

52 Comments on GeForce Kepler 104 and 100 GPU Specifications Compiled

#1
the54thvoid
by: btarunr
If anything, this specifications disclosure is sure to have an impact on buyers saving up for AMD's HD 7900 series products that are heading toward launch on the 9th of January.

Source: 3DCenter.org
I said this would happen in the 79xx thread. :cool:
Posted on Reply
#2
digibucc
by: the54thvoid
I said this would happen in the 79xx thread. :cool:
that specs would impact buyer's decisions? like, you're a psychic man ;)
Posted on Reply
#3
punani
Shader cores are same as CUDA cores, right? and GTX 580 has 512 of them, right ? :respect: GK100
Posted on Reply
#4
Live OR Die
Give me im going to get one of these :)
Posted on Reply
#5
halfwaythere
I call bullshit. If 7970 comes very close to 590-6990 performance wise nvidia is going to have a very hard time competing without breaking any consumption rules.
Posted on Reply
#6
treehouse
by: halfwaythere
I call bullshit. If 7970 comes very close to 590-6990 performance wise nvidia is going to have a very hard time competing without breaking any consumption rules.
i call bullshit on a 7970 coming close to a 6990
Posted on Reply
#7
Live OR Die
I call bullshit AMD will flop once again just like there cpus did.
Posted on Reply
#8
LAN_deRf_HA


285 had half the shaders of the 480, got 71% of the performance.

Safe to assume the 580 to 680 relationship will be similar?
Posted on Reply
#9
blibba
by: halfwaythere
I call bullshit. If 7970 comes very close to 590-6990 performance wise nvidia is going to have a very hard time competing without breaking any consumption rules.
Why? They're both going for a die shrink and about 40% more cores than their current products.

Remember, before Fermi, Nvidia was the king of power-efficiency - who's to say they've not gotten over that mistake now?

by: LAN_deRf_HA

285 had half the shaders of the 480, got 71% of the performance.

Safe to assume the 580 to 680 relationship will be similar?
Classic. The 480 had a little under double the raw computational performance of the 285. But double performance does not equal double FPS, because there are other factors. I'm sure you could find a graph showing the 480 outperforming the 285 by more than 100% if you looked in the right places (I'm thinking where the 1GB frame buffer is exhausted).
Posted on Reply
#10
halfwaythere
Comparing gpus across generations is a bit of a stretch because drivers dont stay the same. If they get the drivers right this time around I think its doable.
Posted on Reply
#11
treehouse
by: Live OR Die
I call bullshit AMD will flop once again just like there cpus did.
so you're saying AMD wont flop? or they will in which case you would not be calling bullshit?
Posted on Reply
#12
digibucc
by: halfwaythere
Comparing gpus across generations is a bit of a stretch because drivers dont stay the same. If they get the drivers right this time around I think its doable.
what? sure there are always driver hiccups but i think the fact that they are as optimized as they can be at the time of comparison is about as equal footing as you'll get, and that's enough imo.
Posted on Reply
#13
DarkOCean
Those specs meets my exact expectations from 28nm.
Posted on Reply
#14
nt300
It wonderful the speculate. But this time round the HD 7900 is a brand new design something that was in design for some time now. Looking at the specs of Nvidia's Kepler, it may be a close competition for them to keep up with AMDs HD 7970 IMO.
Posted on Reply
#15
Zubasa
by: DarkOCean
Those specs meets my exact expectations from 28nm.
The thing is will the yield meet nVidia's expectation?

by: nt300
It wonderful the speculate. But this time round the HD 7900 is a brand new design something that was in design for some time now. Looking at the specs of Nvidia's Kepler, it may be a close competition for them to keep up with AMDs HD 7970 IMO.
Well there is alot that can go wrong for both sides this round.
I rather reserves my judgment when the actual cards are released.
Posted on Reply
#16
v12dock
This is a TYPICAL Nvidia move lets release specifications near AMDs launch then set a launch date and never meet it
Posted on Reply
#17
HD64G
In my opinion the first batch of Kepler 104 and 100 will be reduced in cores just like 480. And it will launch very late to make someone wait to buy it since 7970 will prevail in every aspect whatever exists right now in GPU segment with normal power consumption...
Posted on Reply
#18
DarkOCean
by: Zubasa
The thing is will the yield meet nVidia's expectation?
That remain to be seen.
Posted on Reply
#19
pjl321
GK100 = Q4 2012 or Q1 2013

The article fails to mention that the GK100 is not due for at least another year and that is according to nVidia's times scales so large pinch of salt required.

Comparing the spec of an AMD chip launching in Jan and possible nVidia chip lauching 12 months later is not exactly apples for apples is it now.
Posted on Reply
#20
entropy13
by: v12dock
This is a TYPICAL Nvidia move lets release specifications near AMDs launch then set a launch date and never meet it
Typical? It only happened during the GTX 400 series. The GTX 500 series "met" their dates because they were already set at a later date in the first place. The GTX 200 series didn't miss their launch dates too.
Posted on Reply
#21
the54thvoid
by: HD64G
In my opinion the first batch of Kepler 104 and 100 will be reduced in cores just like 480. And it will launch very late to make someone wait to buy it since 7970 will prevail in every aspect whatever exists right now in GPU segment with normal power consumption...
Your opinion is based on what? It's all very well saying 'in my opinion' but at least base it on some current tidbit of contemporary knowledge, NOT what happened with Fermi. Current noise suggest NV will release low end 28nm process designs first to make sure they get it right on the high end. I don't see another Fermi coming - it would be too harmful to NV to let it happen again. Going on the rumours they learned some lessons from Fermi.

by: pjl321
The article fails to mention that the GK100 is not due for at least another year and that is according to nVidia's times scales so large pinch of salt required.

Comparing the spec of an AMD chip launching in Jan and possible nVidia chip lauching 12 months later is not exactly apples for apples is it now.
You got a source for that?
(please don't say it's this site http://www.4gamer.net/games/120/G012093/20111124085/) My jap aint too good.

And this table looks very, well.. very.

Posted on Reply
#22
mamisano
by: pjl321
The article fails to mention that the GK100 is not due for at least another year and that is according to nVidia's times scales so large pinch of salt required.

Comparing the spec of an AMD chip launching in Jan and possible nVidia chip lauching 12 months later is not exactly apples for apples is it now.
I agree, was about to mention the same thing.
Posted on Reply
#23
Benetanegia
by: LAN_deRf_HA
http://tpucdn.com/reviews/NVIDIA/GeForce_GTX_480_Fermi/images/perfrel_1920.gif

285 had half the shaders of the 480, got 71% of the performance.

Safe to assume the 580 to 680 relationship will be similar?
Tesla (GTX200) and Fermi are very different architectures. Kepler is thought to be similar to Fermi. It will have imprvements on the shaders, TMUs, etc, etc, but it will probably look very close to what Fermi looks like and perform on the same ballpark SP vs SP.

With a different architecture comes a different way of doing things. Fermi can pack more SPs in the same die area/transistor budget than Tesla (or ir can do 64 bit much faster, depending of we are talking about GF100/110 or the lower end ones), but they are less efficient in general because they share more resources. For example the GTX460 with 336 SPs is about as fast as the GTX285.

Once we compare any Fermi card against anoother Fermi card, scaling is near perfect though and this is most likely going to be inherited to Kepler.

by: halfwaythere
I call bullshit. If 7970 comes very close to 590-6990 performance wise nvidia is going to have a very hard time competing without breaking any consumption rules.
Fermi was the foundation to a new architecture. It was like a 5 story building with foundations of a 20+ story building. It's overkill, innefficient in a way, yet necessary and unnavoidable. Remember R600? ~750 million transistors 320 SP, RV670 680 million transistors, same 320 SP, RV770 800 SPs, 2.5x the amount, more than double the performance, only 995 million transistors. Look how far the architecture reached after that in terms of SP count using the same "foundations" (efficiency did suffer a bit), same thing is posible with Kepler. Fermi was the fist step, the beachhead so to speak (i.e. R600), Kepler will be the balanced chip (i.e. RV770), and Maxwell if rumors are true and is still based on the same overall arch it will be less efficient than Kepler (bottlenecked in some ways) (i.e. Evergreen), but arguably still more efficient (perf/die area) than a new architecture would be.
Posted on Reply
#24
the54thvoid
Just found this familiar table. This helps kybosh the table i posted earlier (which is the source of the timings for GK releases)



GF110 did/does not have 768 cores. Bad info is bad info.
Posted on Reply
#25
Bjorn_Of_Iceland
by: Benetanegia

Once we compare any Fermi card against anoother Fermi card, scaling is near perfect though and this is most likely going to be inherited to Kepler.
Yes.. precisely what they were aiming for in the first place. It was built with a mindset on what would be more conventient to them in terms of releasing a card in a rapid manner for the target demograph
Posted on Reply
Add your own comment