NVIDIA Kepler speculation thread

Also has anyone seen any discussion/data on Kepler DX11.1 support?

 

But how far? The 7970 seems clocked quite a bit more conservative than the 680 (especially if the 680 is mostly running 1050-1100 in those benches). Plus the fact that you won't get that much memory clocking, meaning that the bandwidth difference will start to kick in..
1250 vs 1335 (if it's really staying at that level?!) benches:
http://vr-zone.com/articles/asus-gtx-680-2gb-overclocking-review-win-some-lose-some/15322-5.html

 

DX11.1 is supported.

 

Do you guys know if the 680's adaptive vsync engages automatically on all games, or is it something that must be forced via control panel or only available on new games that support it? Didn't see that explained on the reviews I saw.

 

Nice to see Nvidia be cheaper, quieter, faster, more efficient and with WHQL drivers on launch day, something which AMD seemed to have forgotten about.

ATI made such a fuss about having 12 WHQL releases a year, how important WHQL was etc and now suddenly it's fine to wait over 2 months before they released a set for the 7900 series.

 

It's in the control panel, two settings actually.

 

Well, that makes more sense to me. If only 1/3 of the units are 1/2-rate DP capable, and only one of the GPCs are enabled for DP at any one time (or alternatively if 1/3 of units are full-rate DP capable, but only one SMX is active at a time) you wind up with this situation. I would think the 1/3 differentiator would be sufficient, and would wish for the GPC limitation to be removed. Remains to be seen if it can be, though, or even which one of these two explanations are correct. :sigh:

-Dave

 

http://forums.nvidia.com/index.php?showtopic=225312&view=findpost&p=1386352

http://forums.nvidia.com/index.php?showtopic=225312&view=findpost&p=1386379

 

Apparently, there are at least three review articles with different explanation for the Kepler architecture.

For now, I will stick with TR's version. Here's my personal interpretation: http://i.imgur.com/tnl2N.jpg

 

It's my thought that AMD should scale down the Double Precision rate(arguably the least important compute oriented feature) for the top card of Sea Island given the most slower pace for compute and professional markets.

 

Not really, they already have the same thing on their side and can just go back to a dual GPU card to beat the single GPU card. It seems they already have that ability with the chips they have out right now.

 

Interesting, was the GTX 680 supposed to be the 670Ti?

http://www.geforce.com/Active/en_US...ce-gtx-680/NVCPL-WindowsDesktopManagement.png

 

I'm actually underwhelmed with both the 7970 and 680. I really would love to see a single GPU pushing out 580SLI or 6970CFX numbers so I could switch to a single card solution without having to give up my current IQ running 3x 19x12 monitors. Based on the benchmarks it doesn't look like 2012 is my year for a single card solution

 

I wish someone would benchmark these cards in proper video editing program. I notice in Vegas Pro that using an NVidia card dramatically decreases encoding time, but I also notice that testing a 560ti and 580 yielded identical encode times on the same machine implying theres a cuda bottleneck some place with that generation of gpu. I'm wondering if this new gen of cards performs better.

 

I was wrong, Nvidia confirmed 1/24th

 

With the 8 special full-rate 64-bit units, I now fully expect that the GPGPU card from the Kepler generation will not actually be a graphics card.

Something 500mm²ish with no 32-bit units, triangle setup, interpolation, etc at all and just fully decked with 64-bit shader groups. It would be a monster.

 

I was thinking about the big Kepler -- its configuration may actually use a slimmer SMX architecture, like 128 ALUs (8x16-wide SIMDs) and half the texture units (incl., halving the SFU count). That way a 16-multiprocessor chip is much more likely and within reasonable die-size budget. Geometry and setup configuration should remain the same, for the Quadro market.

 

I have doubts that a 500mm2 chip on advanced process can subsist on a niche divorced from the cash brought in by the professional graphics market.

 

http://www.geforce.com/hardware/notebook-gpus/geforce-gtx-660m/specifications

http://www.geforce.com/hardware/notebook-gpus/geforce-gt-650m/specifications

http://www.geforce.com/hardware/notebook-gpus/geforce-gt-640m/specifications

Looks like GK107 GPU supports PCI Express 3.0?

 

I'm pretty sure it was and before Tahiti was known, hell it could have been a 660Ti...After Tahiti came out and nVidia knew where they stood, they probably were debating between 670 and 680 and ultimately ended up with 680 and possibly chose more aggressive clocks at some point. However that theory can probably only be true if that naming pushed back the release of the big boy as I can't imagine them being without a card named GTX 680 for long.