When will Kepler Launch? What's the "hold up"?

tviceman · **Joined:** Sat Jan 16, 2010 5:02 pm **Posts:** 1168

Kyle Bennett of [h] is now seeing gaming benchmarks. http://hardforum.com/showpost.php?p=103 ... stcount=51

It's all happeneing. Nvidia's "mid range" chip is going to beat AMD's hd7970. All we need now is decent pricing on the part of Nvidia....

grstanford · **Joined:** Sat Apr 24, 2010 11:19 am **Posts:** 5166

Ok, here is a guess. Nvidia originally said of Kepler that it would double Fermi's performance in DP tasks (the famous slide that I've linked to many times before).

So what if all CUDA cores in Kepler are fully 64 bit and the schedulers can treat them as though they were two 32 bit CUDA cores for most graphics operations (but not all like branching and things where all the precision is needed and things like two multiplies at once etc which would explain the variable performance)?

SirPauly · **Joined:** Wed Jan 12, 2011 1:56 pm **Posts:** 818

That's an interesting theory.

BoFox · Posted: Tue Mar 06, 2012 11:33 pm

Ahh, I get it.. so it's really 768 SP units, but with up to 1536 "CUDA" threads.. completely unified with texture filtering/addressing capabilities (thus no need for TMUs or TAUs).

grstanford · **Joined:** Sat Apr 24, 2010 11:19 am **Posts:** 5166

Yes, except not on the unified texturing. That was just me speculating on how nvidia managed to get 1536 physical cores into the chip and still have it do useful graphics work given the die size (it is coming in the future though). There is only 768 real physical cores that can pretend to be 1536 cores thanks to clever schedulers, registers files and cache, though sometimes you won't be able to use the virtual cores and that is where the variable performance comes in. Sometimes you'll get 768 core performance, sometimes 1536 core performance.

It's an old trick borrowed from nv35. It had FP32 registers, but you could pack two FP16 values into 1 FP32 register and get more work done. Of course the "scheduler" (such as it was back then) was nowhere near as sophisticated as what Kepler should have.

BoFox · Posted: Wed Mar 07, 2012 3:25 am

tviceman wrote:

Kyle Bennett of [h] is now seeing gaming benchmarks. http://hardforum.com/showpost.php?p=103 ... stcount=51

It's all happeneing. Nvidia's "mid range" chip is going to beat AMD's hd7970. All we need now is decent pricing on the part of Nvidia....

Is Kyle under NDA?

He was "wise" to not do "3 strikes" in a row against Nvidia. But it's still a foul this time anyways.

See, in his HD 7770 review, he said that GTX 560 Ti does not offer any gameplay benefits over HD 7770.

And then in this HD 78xx review, he's saying that HD 7850 does not offer any gameplay benefits over GTX 560 Ti.

Whoa!! The [H]ardcore [L]oyalists must be confused now, not understanding that HD 7850 is in a wholly different league from HD 7770.

Alrite, Kyle.. you're just trying to be a good boy now.. Nvidia was about to give you a spanking, but you escaped it just now by saying that HD 7850 is too expensive (saying this about HD 7770 instead would have made a lot more sense by looking at the bang4buck chart where HD 7770 is by FAR the worst of all 28nm and 40nm cards: http://hexus.net/tech/reviews/graphics/ ... s/?page=12

apoppin · Posted: Wed Mar 07, 2012 10:26 am

http://www.tomshardware.com/news/nvidia ... 14921.html

Quote:

Chiphell and Expreview published what apparently is a GK104-based card. What we gather from these pictures is that Kepler cards could be huge with potentially high power consumption. . . .

According to reports, the 290 mm2 GK104 chip shown in the pictures will be used in GTX650Ti and GTX 660 graphics cards, which will arrive with 448 and 512 activated stream processors, respectively, as well as a core clock of about 850 and 900 MHz. The first Kepler cards are expected to be announced next month.

Where's the beef??
---The GTX 680?
:scratch:

i'd love to stay and discuss ... but i gotta fly ...
. . . steak dinner tonight!

UPDATE: And now Charlie says TSMC halted production of 28nm chips
http://semiaccurate.com/2012/03/07/tsmc ... roduction/

Quote:

word has reached SemiAccurate that TSMC halted 100% of 28nm production in mid-February to make unnamed changes to the process. Since only Nvidia was having issues, and those are not really process related, we have no explanation as to why they would take this drastic step, but yields don’t seem to be it. That said, we are quite certain that it happened, and is still in effect.

Sources tell SemiAccurate that the stoppage was quite sudden, and production should resume in short order. One source said that by the end of March, volumes would be back to where they were before if not a lot better. Checks with the channel have confirmed an abrupt stop in 28nm chip shipments

C-ya!

grstanford · **Joined:** Sat Apr 24, 2010 11:19 am **Posts:** 5166

Frankly I don't believe Tom's on the shader counts. I still think GK104 is a doubled GF114 (with hot clocks and "hyperthreaded" cores), which means there will be a minimum of 768 cores assuming all SM's are active. Disable one SM and your should be 96 cores down, disable two and you are 192 cores down giving total core counts of 672 and 576 respectively.

I imagine nvidia will serve up a healthy portion of "food for thought" to accompany your steak

People at B3D seem to think that the TSMC stoppage info is correct, so Charlie is probably correct there.

Ocre · **Joined:** Mon Nov 22, 2010 5:22 am **Posts:** 2234

apparently Nvidia's issues is not the only issues with tsmc's 28nm node. Holy Smokes (joint)

imagine that.

CharLIE's again

Nsavop · **Joined:** Fri Jan 06, 2012 12:05 am **Posts:** 362

Gtx 680 to feature turbo boost

Quote:

First, and most interesting: with some models of the GeForce 600, NVIDIA will introduce a load-based clock speed-boost feature (think: Intel Turbo Boost), which steps up clock speeds of the graphics card when subjected to heavy loads. If there's a particularly stressing 3D scene for the GPU to render, it overclocks itself, and sees the scene through. This ensures higher minimum and average frame-rates.
Second, you probably already know this, but GK104 does indeed feature 1,536 CUDA cores, which lend it a strong number-crunching muscle that helps with shading, post-processing, and GPGPU.

Third, the many-fold increase in CUDA cores doesn't necessarily amount to a linear increase in performance, when compared to the previous generation. The GeForce GTX 680 is about 10% faster than Radeon HD 7970, in Battlefield 3. In the same comparison, the GTX 680 is slower than HD 7970 at 3DMark 11.

Fourth, the NVIDIA GeForce GTX 680 will very much launch in this month. It won't exactly be a paper-launch, small quantities will be available for purchase, and only through select AIC partners. Quantities will pick up in later months.

Fifth, there's talk of GK107, a mid-range GPU based on the Kepler architecture, being launched in April.

Next up, NVIDIA is preparing a dual-GPU graphics card based on the GK104, it is slated for May, NVIDIA will use Graphics Technology Conference (GTC), as its launch-pad.

http://www.techpowerup.com/161972/GeForce-GTX-680-Features-Speed-Boost-Arrives-This-Month-etc.-etc..html

Heise.de (source)

When will Kepler Launch? What's the "hold up"?

Who is online