Nvidia BigK GK110 Kepler Speculation Thread

[A1xLLcqAgt0qc2RyMz0y]

More on Tesla K40 and IBM hookup

In August this year, IBM announced their OpenPOWER consortium which marked a development alliance between IBM, Google, NVIDIA, Mellanox and Tyan. The group aims to build server, networking, storage and GPU-acceleration technologies for developers of next-generation, highly scalable cloud data centers. Today, IBM and NVIDIA are jointly announcing co-development of GPU-accelerated IBM enterprise software and applications on IBM POWER Systems. The companies plan to work together integrating IBM’s POWER8 Processor architecture with NVIDIA Tesla GPUs.

The two should make for a potent HPC tag-team. IBM’s POWER8 architecture is a monster multithreaded 12-core chip with 96MB of on-chip eDRAM L3 cache that is capable of processing up to 96 threads simultaneously. It’s a ton of CPU horsepower for feeding hungry GPUs data for crunching. The agreement will also open up sizable opportunities for NVIDIA as well, where IBM is a market leader in the enterprise data center. Patrick Moorhead, principal analyst of Moor Insights and Strategy notes: “This is the most interesting piece of NVIDIA’s announcements at SC13. IBM is very competitive in HPC and they’ve announced that they will standardize on Tesla and CUDA going forward. Even bigger is that IBM will pull NVIDIA Tesla and CUDA into WebSphere and Analytics where IBM is an 800 pound gorilla.”

http://www.forbes.com/sites/davealt...ategic-partnership-with-ibm/?partner=yahootix

...

More color added by Ryan Smith of AnandTech

http://www.anandtech.com/show/7522/ibm-and-nvidia-announce-data-analytics-supercomputer-partnership

 

[DSC]

http://www.anandtech.com/show/7521/nvidia-launches-tesla-k40

 

[trinibwoy]

nVidia picks up another spot in the top 10 supercomputer list at number 6. It's also #1 on the green list with highest perf/w. Rmax/Rpeak is pretty impressive at 80%.

http://www.top500.org/lists/2013/11/

 

[ams]

nVidia picks up another spot in the top 10 supercomputer list at number 6. It's also #1 on the green list with highest perf/w. Rmax/Rpeak is pretty impressive at 80%.

http://www.top500.org/lists/2013/11/

Systems with NVIDIA Tesla GPU's + Intel Xeon CPU's have swept the top 10 spots in the November 2013 Green 500 list: http://green500.org/news/green500-list-november-2013 . The full Green 500 list is here: http://green500.org/news/green500-list-november-2013?q=lists/green201311 . What is remarkable is that the number one Green 500 system from Japan is nearly 25% (!) more power efficient than any other supercomputing system on the Green 500 list.

 

[RecessionCone]

What happens when you give a GK110 the same power budget as an R9 290X:
http://www.hardwarecanucks.com/foru...1-gigabyte-gtx-780-ti-ghz-edition-review.html

 

[kalelovil]

What happens when you give a GK110 the same power budget as an R9 290X:
http://www.hardwarecanucks.com/foru...1-gigabyte-gtx-780-ti-ghz-edition-review.html

Why do review sites continue with rather useless power consumption measurements?

In this case using a canned benchmark which is not even part of the article's performance comparison.

With the increasingly sophisticated and variable dynamic clocks used by both major GPU vendors, to give an accurate power consumption the reviewers should ideally be measuring the power consumed during each game run.

 

[lanek]

Why do review sites continue with rather useless power consumption measurements?

In this case using a canned benchmark which is not even part of the article's performance comparison.

With the increasingly sophisticated and variable dynamic clocks used by both major GPU vendors, to give an accurate power consumption the reviewers should ideally be measuring the power consumed during each game run.

Thanks you... finally, someone who look to understand it.

Anyway, this remind me some R290x review launch, where they was increase the temp limit to the same level of the 290 on the 780TI...( 85 to 94°C ), ofc the 780TI was running faster as it clocks was staying in higher boost clock as the limit of temp and TDP was increased.. Completely stupid, do we have test the 7970 with similar setting overclocked when the 680 have been launched ? No.. maybe we could have increase the clockspeed for match the turbo clock speed of the 680 ? for use " similar setting " ? ( if follow the tester and review community since 1998 and this is the first time i see a review done like that. ) .. this is ok if you test a 780 vs 780 or an AMD vs AMD cards.. but why tests gpu this way ? whats next ? if a Nvidia card is clocked at 1100mhz turbo, you will overclock the 290x to 1100mhz for the review ?

 

[CarstenS]

Why do review sites continue with rather useless power consumption measurements?

In this case using a canned benchmark which is not even part of the article's performance comparison.

With the increasingly sophisticated and variable dynamic clocks used by both major GPU vendors, to give an accurate power consumption the reviewers should ideally be measuring the power consumed during each game run.

Hm, I understood RecessionCone's slightly different, since the 780 Ti and it's OC-variants are not bound by the 250 watt limit, the Titan always had to make do with. I really don't think that he was referring to the single measurement of power at the end of the article.

 

[kalelovil]

Hm, I understood RecessionCone's slightly different, since the 780 Ti and it's OC-variants are not bound by the 250 watt limit, the Titan always had to make do with. I really don't think that he was referring to the single measurement of power at the end of the article.

Fair enough. I wasn't disputing RecessionCone, rather making an observation of the link he posted.

 

[RecessionCone]

Fair enough. I wasn't disputing RecessionCone, rather making an observation of the link he posted.

FWIW, I agree with you that power numbers from important games are more useful than Unigine.

 

[Frontino]

Why doesn't NVIDIA sell GK110 with more than 3 SMX disabled (not withstanding the rumored 790)?
And why there are no GeForce GK110 with only 2 SMX disabled?

 

[silent_guy]

Why doesn't NVIDIA sell GK110 with more than 3 SMX disabled (not withstanding the rumored 790)? And why there are no GeForce GK110 with only 2 SMX disabled?

Diminishing returns.

Numbers pulled from thin air, but if you have a perfect yield of 60%, that means that the defect density is already pretty low, and most failing perfects will have only 1 defect. So after recovery rate by disabling one will be high too. Say 20%. At that point, there is only 20% left to recover. With 2 SMX disabled, you will recover yet another big chunk, and 3 disabled should get you close to 100%.

(At least if the defects are located on an SMX. Defects on other places don't enter the equation.)

3 SMX disabled was pretty high, but they must have stockpiled them when they were making Tesla cards. That was relatively early in the life of 28nm with yields lower. Process quality should be much better now. It wouldn't surprise me one bit if most today's GTX780 dies have more than 12 SMX working but they disable 3 to keep the market segmented correctly.

 

[DSC]

http://videocardz.com/49559/nvidia-geforce-gtx-titan-black-black

 

[Skinner]

http://videocardz.com/49559/nvidia-geforce-gtx-titan-black-black

Should it run on 1+ ghz base clock?

 

[Skinner]

nm.

 

[iMacmatician]

From Videocardz: "NVIDIA GeForce GTX TITAN BLACK ‘officially’ confirmed."

 

[DSC]

http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-titan-black

 

[pjbliverpool]

Base/Boost clock of 889/980 vs 875/928 of the 780Ti so we seem to have a new king of the hill if only by a marginal amount.

 

[DavidGraham]

Stupidly stubborn NVIDIA still can't seem to bother releasing a 1GHz high-end chip, what a bummer.

 

[mczak]

Base/Boost clock of 889/980 vs 875/928 of the 780Ti so we seem to have a new king of the hill if only by a marginal amount.

One thing is for sure, if you're getting it because of the higher clocks, you're crazy indeed considering the price difference. But you still get the 6GB vs. 3GB, and of course the non-crippled DP. It is nice though that the ridiculously priced 1000$ card now at least is the fastest nvidia card (even if it might be essentially a tie, at least it's not slower any longer!).