Kepler GK104: Some new rumors on Nvidia's next generation cards.

It’s me again, keeping track of graphic hardware for you folks as usual :smiley:

For those not aware of Kepler, it´ll be the successor to the current Nvidia Generation of GeForce cards. The GPU is the direct successor to the GTX560 TI, which has a GF114 chip.

It seems the first GK104 beasts (the new chip) are in the hands of various insiders.
No one knows yet what actual name the card will have, most likely it’ll be the GF700 series but who cares about the name at this point anyways.

What’s important thus far is, that the cards have 8*GDDR5 chips which should result in 2GB VRAM on a 256 bit bus.
Rumors have it that to achieve a high performance with 256 bit (current GTX have 384bit) the memory clock will be quite high.
This already keeps the hopes up we’ll get 4GB cards from board partners and some space where to put CUDA scene data :slight_smile:

The cards seem to have 26pin PCIe power plugs, which would mean 225W, while the AMD Thaiti got 28pin. So the nvidia might be more power efficient.

Very good news is that the chips are labled with A2 stepping, meaning those are not the first generation of chips and we can expect to get a mature chip already, not a betachip like the GF100 Fermi chips :wink:

It also seems that nvidia is already massing the chips to be able to provide the card on launch en masse and not just paperlaunch it.

In chinese forums the rumor is the card will be launched early march, in the western world the rumors are towards early april.

In any case it’s quite exciting.
Some of you might know Charlie Demerjian the founder of SemiAccurate, and he’s very, very critic to Nvidia, but he said that he had several of those cards in his hands already and that Kepler GK104 is superior to the AMD Thaiti in all aspects.

Remains to see the pricing, the benchmarks and most important how it’ll perform in OpenGL and CUDA.
I am the crying clown on this, on one site I am sitting on ants, excited about the new card which supposedly is a quantum leap again in performance, on the other side I am already afraid of nvidias “product strategies” shrugs

Which reminds me, anyone here has a HD7970 yet and can tell something about OpenGL performance in Blender?

Exciting. And here I was just about to get something from the 500 series to replace my old GTX 295. I think I’ll wait for the benchmarks. :slight_smile:

curious, they stepped down from 384 to 256 bit width bus…

Been holding out for the 6xx/7xx series. I also read a rumor yesterday that they were moving up to a February release date in response to the success of ATIs offerings. I hope it’s true, I know what my birthday present will be for myself if it is!

it said kepler will enter as gfx 680 with 1024 cores, 590 has two cards in one, each 512cores = 1024 cores.

if that’s true and the price is not insane that will be so awesome.

the next chip after kepler will be maxwell, seems they gonna take a huge leap again, or atleast planned to.

chinese forum said it might even be out late february to compete with amd/ati 7970 cards.

1024 cores sounds nice, but you probably won’t find that many in what you find on the shelf. You can expect some of them to be turned off to compensate for the initial production yields, just like with 2xx, 4xx and 5xx series.

I think it’s gonna be 1024 in the 680, but… looking at the 5xx series, the 560 looks like the best —bang per buck—
so, the 660 might turn out like 700ish cores but also good ratio between performance and money spent.

It will absolutely be 1024, or even more, depends on what changed in the architecture, and they have to release the flagship first, as they got to go into direct competition with the HD7970 which is the fastest single GPU card for the next 2 months at least.

I guess what MM meant that they’ll run out of stock in the first day and you’ll be stuck with a preorder or a weaker card.
I for one doubt it this time, the C2 stepping suggests they got a good yield already.

We’ll see.

Lets hope so! Either that or they had a lot of bugs. :slight_smile:

will Kepler utilize PCIE 3.0?
I’m thinking of getting the ASUS P8z68 pro v GEN 3 if Kepler does use PCIE 3.0…

I thought the only thing that matters to blender when it comes to GPU is it’s VRAM? does it matter if…let’s say a person has HD5870 2gb and a newest ATI 7XXX series… and just for this story sake… let’s just say the 7XXX series card posses only 1gb worth of Vram, wouldn’t it be a better choice to get the lower end 2gb HD 5870 instead?

I keep seeing people comment on threads saying that GPU power does not mean much to blender (if any) is this true? enlighten me.

Sorta kinda. You don’t need much GPU power for modelling or simple animation. Last generation cards will do fine. You do need power for Cycles, but if you’re considering a Radeon you must be planning on using something else.

one more question:
how does the GPU VRAM effects blender viewport? I thought that it’s the RAM that decides as to how many polygons a scene can handle?

I don’t actually know. If I ever managed to run out of vram, I never noticed, so I don’t know what happens if you do. I’m guessing rapid performance degradation as data is swapped between system memory and vram? A developer would have to answer this one.

I heard that the GK110 chip will be much faster in GPGPU than the previous graphics cards. So maybe it’s a good idea to wait until nvidia releases graphics cards based on that chip? Does anyone know something about it?

Well you’d hardly expect the next generation of cards to be slower. Of course it will be faster. It might be a good idea to wait in any case because the new chips will probably bring a new compute model with them.

CPU= render time.

RAM= more ram more polygons

gpu power= even if you have more rams therefore are able to produce more polygon doesn’t mean you are able to work on the 3d scene somoothly and fast. this is where gpu power comes in right?

GPU VRAM= decide how big the scene can be. I always thought that blender 3d world has unlimited space with no end… turns out that it depends on how much VRAM your gpu have.

correct me if I’m wrong people~

Not necessarily. A GeForce doesn’t need fast CUDA. It’s very much possible that they cap the SP/DP performance again.
The current GeForce and Quadros have the DP performance capped at 1/8th of what is possible to sell Tesla cards.

If the GK110 is over 8 times faster than the current one, they’ll cap it for sure, else a cheap GeForce would outperform the current Teslas and “professional customers” would most likely be angry buying a 5000 euro card just to see a 500 euro card outperform it a month later.

Nvidias marketing strategy brought them in a downward crippling spiral they can’t escape anymore unless the make a magic product 100 times faster than the previous generation and start from zero with their strategy.

Furthermore rumor has it that the first cards are not expected before june :confused:
Up to everyone to wait for nvidia to announce a date to announce a date where the NDA falls to announce a releasedate lol.

But seeing the HD7970 the GK110 can only be a success. While the new Radeon is the fastest Card in the world at the moment it’s only 15% faster in average than the GTX580 which I don’t find very impressive - unfortunately.

Well that would be too bad. Crippling GPGPU performance just when it’s taking off for practical, non-scientific applications is a really bad idea I think. Oh well. With Apple championing OpenCL and using it promiscuously throughout their OS, I’m sure we will eventually see either AMD win, or Nvidia cave and decripple their cards.

I really doubt that AMD will win anything anytime soon. The Bulldozer was a fail on many fronts and the 7900 is mediocre at best. And I really like AMD and have to say that from a practical point of view.
And they released their quarterly results yesterday, they’re deep in the red figures, lost a lot on the graphic market, only 20% of their income is based on the graphic sector anymore. Even the sales in console graphic units can’t make up for their loss of ground to the mobile sector where Nvidia and Intel are the topdogs.

And I am 99% certain that GK104 will be crippled for GPGPU in performance and with the least VRAM possible on GeForce cards, just because.
If only there was a healthier competition.