GK110 = Kepler-based Tesla K20 GPU, with up to 2,880 SPs, 7.1 Billion transistors!

[RussianSensation]

Nvidia is announcing two products today, two Kepler based HPC / GPGPU cards - K10 and K20.

NVIDIA Tesla K10

Tesla K10 dedicated to seismic analysis where most valuable resource is bandwidth. It has 3x single precision of today’s fermi tesla and 1.8x the memory bandwidth.

NVIDIA is preparing two Tesla GPUs, one is Tesla K10, based on dual-GK104 chips, which is to be devoted to high efficient computing in oil and gas exploration and defense industry. Card is based on dual-GK104 GPUs and has a memory bandwidth of 320 GB/s.

The Tesla K10 packs two Kepler-based GK104 GPUs and 8 GBs of VRAM.

Source

This is basically a GTX690 with 4GB of VRAM per GPU.

NVIDIA Tesla K20

Second Tesla, K20 is focused on double precision: 3x double precision of fermi, includes hyper q, dynamics parallelism, for stuff like physics, quantum chemistry, computational finance.

GK110: Actual chip is to feature 15 SMXs (streaming multiprocessors) with 192 CUDA cores each. This gives 2,880 CUDA cores in total.

GK110 GPU which is coming in Q4 of this year.

Source

Update: GK110 2-page PDF
GTC Live feeds - Click Here.

Update #2: NVIDIA disclosed that the Tesla K10 offers a memory bandwidth as high as 320 GB/s. - TechPowerup

Thoughts: It is now official -- GK104 is not the high-end Kepler chip. It just happens to be the high-end gaming chip at the moment. Looks like NV is pushing for higher profit margins by selling GK110 in the professional markets. I think GK110 will become GTX780 and launch in 2013 at this point.

Update #3: Rumored Specs for GK110

Full physical GK110

15 SMX @ 192 CUDA Cores = 2,880 SPs
K20 Tesla parts will likely have 13-14 of SMX clusters enabled = 2,496 - 2,688 SPs
384-bit memory bus
TDP > 225W
K20 Tesla Launch Date: Q4 2012
Consumer GK110 Launch Date: 2013

Source 1
Source 2

Thanks to Sontin and Cookie Monster for finding that info and breaking it down.

 

[Quantos]

No GeForce out of this? Wow, that's very interesting! That thing sounds incredibly powerful. Perhaps too powerful to be kept in check by conventional cooling methods?

 

[Don Karnage]

Green PCB = Fail

 

[railven]

Thoughts: It is now official -- GK104 is not the high-end Kepler chip. It just happens to be the high-end gaming chip at the moment. Looks like NV is pushing for higher profit margins by selling GK110 in the professional markets.

I remember presenting this scenario to you in one of our debates and you called me all kinds of mean things for even pondering it as a possibility

Vindication!

 

[JTsyo]

No GeForce out of this? Wow, that's very interesting! That thing sounds incredibly powerful. Perhaps too powerful to be kept in check by conventional cooling methods?

Maybe so powerful that it's weapons grade and would violate arms trade embargoes. They should have saved there crowbar promo for this card

 

[blackened23]

So whens the MSI lightning Tesla K20 coming out! I want to OC that sucker for raytracing.

 

[RussianSensation]

I remember presenting this scenario to you in one of our debates and you called me all kinds of mean things for even pondering it as a possibility

Vindication!

No, the reason I disagreed with you because you insinuated that GK104 is Kepler generation's high-end chip. Also, many people continued to claim that GK110 is just a made up chip that might never be released.

I don't remember ever saying that GK110 won't launch in professional markets. Also as far back as I remember I always stayed firm that GK104 was a mid-range chip and I didn't expect GK110 until August/Q3 2012 at the earliest. I never claimed to know when exactly GK110 would launch in the gaming market but I said from the beginning that GK104 was a mid-range chip and could be sold for $400-500 because it's good enough to go against 7900 series. Also I am not sure how you are vindicated since I don't recall stating that I expected GK110 to launch alongside GK104 in March-May time frame.

I even said that there was a strong correlation to GTX7800 256mb --> 7900GTX scenario where GK110 would likely follow up closer to the fall. I think we got into an argument because you claimed that GK104 is the high-end Kepler chip and I disagreed. This finally proves that notion false. NV couldn't get it out on time which is why they had to use GK104 and which is why they are having a hard time getting GTX660 / 660Ti parts out since GK104 was supposed to be that part all along.

 

[blackened23]

I remember presenting this scenario to you in one of our debates and you called me all kinds of mean things for even pondering it as a possibility

Vindication!

Well, that and its not close to ready yet The presentation was pretty interesting, I enjoyed it. Except for the music....they need to get some better music for their presentations!

 

[Vesku]

Q4 for the real compute Kepler, seems like some people should grab some salt to go with that crow they are eating.

In the meantime they will be selling the GTX 690 at a nice markup to the specific business uses in which it can do better than the current Tesla product.

 

[blastingcap]

The more astute forumers have been saying this for a long time. That gk104 was not the flagship, as should be obvious not only from the various leaks and GTX670Ti renaming cover-up, but by the design and lack of DP and memory bandwidth compared to the previous flagships.

I guess NV took a page out of AMD's playbook when it comes to gaming GPUs, but executed better: go small die, high efficiency, lose the HPC stuff, and put two on one PCB to combat the other side's monolithic GPU. Just like AMD did with HD4870/HD4870x2 vs. GTX280, and HD5870/HD5970 vs GTX480, and HD6970/HD6990 vs. GTX580.

The difference is that NV's top gaming GPU actually edges out AMD's top gaming GPU, i.e., 680>7970, whereas 4870<280, 5870<480, 6970<580.

When 28nm process becomes more mature, NV can probably get much better yields on massive GK110 GPUs, so it makes sense to wait. AMD is not really a threat in HPC and pro graphics so NV can afford to wait.

And I guess there is a market for HPC that does NOT require DP, hence the K10 launch.

 

[blackened23]

I don't really think they're just holding it back just because. If others are suggesting that they're just not releasing it because they don't feel like it, thats not true. It only recently taped out a few months ago, and do keep in mind it generally takes 9-10 months from tape out to get a product. So -- again -- they're not holding it back, its just not ready.

GK104 was the high end at release because its the highest end part that nvidia had, and still is the highest part that they have that isn't going through iterations of tapeouts and validations. What will be interesting is how they gear it to the consumer market, hopefully AMD can really step it up with Sea Islands. AMD GPU's are always pretty good at being released on time on a yearly basis, one can hope they make a killer like they did with the 5870 around December of this year. 7xxx was pretty promising but nvidia definitely won this round, but ATI has overcome that in the past.....hopefully this will put the pressure on prices and/or performance of consumer level GK110s later this year.

 

[Joseph F]

Green PCB = Fail

They are professional cards, it doesn't matter what the PCB looks like when they spend their entire life in a rack, obscured from sight.
Actually, I prefer plain green PCBs to the ridiculous, gaudy colors sometimes used that are associated with gamers, nowadays.

Though, I'm sure you're just kidding.

 

[tviceman]

And just as I suspected, there will be a larger performance gap between GK104 and GK110 than there was between GF114 and GK110. But then again, with the amount of time that it will have taken to get GK110 out, all of this info is both appropriate and expected. I am guessing GK110, when it makes it's way into a geforce product, will be 60-80% faster than GK104 across the board in graphics performance. I think we will be "lucky" to see it's initial msrp at or below $650.

 

[greenhawk]

8GB ram on a video card?

A good enough indication that I need to upgrade. A video card with more RAM than I originally had in this current system (4GB).

I upgraded to 8GB recently as a stop gap, but looks like I need to set it up somewhat.

 

[blackened23]

8GB ram on a video card?

A good enough indication that I need to upgrade. A video card with more RAM than I originally had in this current system (4GB).

I upgraded to 8GB recently as a stop gap, but looks like I need to set it up somewhat.

8gb 690s will not be released if history is any indication, and this 8gb dual gk104 tesla k10 card is geared for professional use. It will definitely be priced accordingly....

 

[RussianSensation]

GK104 was the high end at release because its the highest end part that nvidia had, and still is the highest part that they have that isn't going through iterations of tapeouts and validations. What will be interesting is how they gear it to the consumer market, hopefully AMD can really step it up with Sea Island.

Fair enough. The way I look at it is NV had trouble delivering GK100 (True GTX480 successor) out on time for whatever reasons (yields, capacity, profitability, product stack). VR-Zone's article reinforced that they were impressed by GK104 and after seeing HD7900 series they simply launched a highly overclocked "GTX460/560" successor. Instead of getting out a hot and power hungry GTX685 as they did with GTX480, it looks like they might go directly to GTX580 (GK110 --> GTX780). If you want to call GK104 high-end based on current product stack, I won't disagree but internally it was designated mid-range within NV.

With 3072 CUDA cores, NV is ready to increase performance by at least 50% for the next round. Let's hope we get something like this out on the desktop within 12 months. :thumbsup:

And just as I suspected, there will be a larger performance gap between GK104 and GK110 than there was between GF114 and GK110. But then again, with the amount of time that it will have taken to get GK110 out, all of this info is both appropriate and expected. I am guessing GK110, when it makes it's way into a geforce product, will be 60-80% faster than GK104 across the board in graphics performance. I think we will be "lucky" to see it's initial msrp at or below $650.

With those specs it might have 80% of GTX690's performance. $650 for that is a bargain by today's standards .

 

[Olikan]

3x double precision of fermi

WOAH!

they might use 1:1 DP rate to reach that

 

[KompuKare]

So how much will this K20 cost? Seems that NV were smart this time: do the monster chip later and charge so much for it so that yields are not so important. I still have this notion that all those gamers shouting 'I want big Kepler' will be disappointed. Not only will it launch first in Tesla, there's nothing to indicate that GK110 will even perform well in games: I suspect in the same way that GK104 had it's compute neutered, GK110 will not waste die space on gaming features. And that going forward, Nvidia want to put a larger distance between their gaming card and Tesla: the gamer's will get a bit of CUDA so that Nvidia can claim a certain market share penetration and say 'x million can run CUDA' but they will make sure it's only a bit faster than running that CUDA code on a CPU.

Surprising is the K10 since GK104 is so poor at compute in general. Have they got a bit market for single-precision or are the gaming Keplers crippled really well (would the laser-cut parts?)

 

[Grooveriding]

double post

 

[3DVagabond]

This just in. Someday nVidia will release a bigger faster chip than GK104. When they do, the GK104 will no longer be top of the line. Then everyone calling it a midrange chip will be right and they will proclaim it to the rooftops. In the meantime, GK104 = top of the line nVidia GPU. It's the biggest baddest thing they have the engineering skills to make.

I was going to say skills to make in commercial quantities, but that remains to be seen, actually.

 

[tviceman]

Fair enough. The way I look at it is NV had trouble getting GTX480 out on time, so they overclocked GTX460 and will go directly to GTX580 (GK110 launched as GTX780). If you want to call GK104 high-end based on current product stack, I won't disagree. However, with 3072 CUDA cores, NV is ready to increase performance by at least 50% for the next round. Let's hope we get something like this out on the desktop within 12 months. :thumbsup:

50%? That is super conservative!! GF110 was 40% faster than GF114, and that was with only one third more cuda cores running a little slower. GK1110 will be TWICE the cuda cores GK104 is. And it won't be bandwidth starved! This sucker is going to be a fracking monster!!!

 

[blackened23]

So the K10 is a just a GTX690.

The K20 is a new chip. Have they mentioned die size ?

I see 2 vapor chambers on 1 chip so......

Big. No holds barred? Possibly.

 

[Grooveriding]

So the K10 is a just a GTX690.

The K20 is a new chip. Have they mentioned die size ?

With 3072 CUDA cores, NV is ready to increase performance by at least 50% for the next round. Let's hope we get something like this out on the desktop within 12 months. :thumbsup:

That is being way too optimistic. 50% more shaders is not going to mean 50% more performance with the overhead of a compute GPU. GK104 has nothing in the way of nvidia's usual compute arsenal and is aimed squarely at gaming. This card most likely will give the true performance leap over the GTX580 that we are used to from nvidia with a node shift, 60-70%, and make up the slack that the GTX680 left with it's small 30% gain on the GTX580.

There will be some nice gains in memory bandwidth though. 50% more sounds like a best case scenario. For gaming, my best guess is 30-35% at most over a 680. It's unrealistic to expect the perf/mm2 characteristics of GK104 to play out in the compute heavy GK110.

Hopefully they remove GPU boost altogether from this chip as well, along with allowing for voltage control.

 

[thilanliyan]

This just in. Someday nVidia will release a bigger faster chip than GK104. When they do, the GK104 will no longer be top of the line. Then everyone calling it a midrange chip will be right and they will proclaim it to the rooftops. In the meantime, GK104 = top of the line nVidia GPU. It's the biggest baddest thing they have the engineering skills to make.

Lol. Speculation is fun I guess but too bad we don't have any sort of release date for a desktop GK110. If it's coming out at the END of this year, it may not look so impressive. If it's coming out within like a month then it would be impressive.

 

[railven]

No, the reason I disagreed with you because you insinuated that GK104 is Kepler generation's high-end chip. Also, many people continued to claim that GK110 is just a made up chip that might never be released.

I never insinuated that, you continue to misrepresent my posts, but regardless.

Vindicated!