NVIDIA Kepler speculation thread

Tridam's just published an article about GPU Boost, revisited: http://www.hardware.fr/focus/65/gpu-boost-gtx-680-double-variabilite.html

Basically, his press sample was qualified up to 1110MHz while retail cards may be limited to 1097, 1084, 1071, or perhaps as low as 1058MHz. So he benched a random retail card against his press sample, measured a 1.5% difference on average, up to 5% in Anno 2070.
Does anyone know which cards he tested? Not that I don't believe him, but the article does not provide much in the way of scientific proof. Perhaps a more in-depth test is in the works?
 
Does anyone know which cards he tested? Not that I don't believe him, but the article does not provide much in the way of scientific proof. Perhaps a more in-depth test is in the works?

What do you mean? He tested a Gigabyte, "up to 1084MHz" cards against his "up to 1110MHz" press sample.
 
He claims he has observed "several" retail cards and some didn't make it higher than 1071 MHz, the number 1058 MHz also appears in the article.
Also, two cards are quite an insufficient statistical sample. The Gigabyte card could have worse cooling, worse contact between the die, paste and cooler, which would inhibit it from hitting higher boost clock. It does not necessarily mean that there are five bins of chips. Where would they set the limit, in the BIOS, or by blowing a fuse or something?
 
Last edited by a moderator:
http://www.newegg.ca/Product/Product.aspx?Item=N82E16814130781

Wonder if we'll ever see one for anything near $999. On the bright side, you can buy one.

Daaamn, its nearly the cost of an Asus Mars ...

He claims he has observed "several" retail cards and some didn't make it higher than 1071 MHz, the number 1058 MHz also appears in the article.
Also, two cards are quite an insufficient statistical sample. The Gigabyte card could have worse cooling, worse contact between the die, paste and cooler, which would inhibit it from hitting higher boost clock. It does not necessarily mean that there are five bins of chips. Where would they set the limit, in the BIOS, or by blowing a fuse or something?

The problem is not the cooling, it is the qualification of the chip... Nvidia have some qualification tests, following thoses tests, Nvidia set the Turbo Boost at a predeterminate clock. ( the max it can attain ).

Ofc the limit is set or in the bios, or in the hardware who control the turbo boost..

See it like AMD who determine the voltage on 7970 following the quality of the Asic ( not only ), the voltage can be 1050 -1120- 1175mv ..
 
Last edited by a moderator:
The problem is not the cooling, it is the qualification of the chip... Nvidia have some qualification tests, following thoses tests, Nvidia set the Turbo Boost at a predeterminate clock. ( the max it can attain ).

Ofc the limit is set or in the bios, or in the hardware who control the turbo boost..

See it like AMD who determine the voltage on 7970 following the quality of the Asic ( not only ), the voltage can be 1050 -1120- 1175mv ..
And the cards with different GPU voltage have different BIOSes? That seems needlessly complicated for manufacturing. I wonder how they do that.
But about those qualification tests and binning to different maximum boost clocks - is this like a well known thing, or just speculation?
 
And the cards with different GPU voltage have different BIOSes? That seems needlessly complicated for manufacturing. I wonder how they do that.
You certainly wouldn't do it via different BIOSes. Likely the chips have a series of fuses that indicate the ASIC "quality" and the BIOS programs the voltage appropriately.
 
GTX 670 Review

http://www.tweaktown.com/articles/4...2gb_video_card_performance_preview/index.html

looking to power plugs and length of the card it's probably a Gigabyte card

IMG_3329-2.jpg


IMG_3336.jpg
 
Last edited by a moderator:
For a site that's upset with nVidia they sure have a very nVidia friendly game selection. Not to mention outdated.

The 670 will probably put downward pressure on 7970 pricing though.

I will be honest, i dont see the interest of make 8 pages of test without AA, and then put 3x games with AA / AF 1900X1200 on a single pages..

Are they joking ? They are the only tester who are able to get the 680 in front of the 7970 in Metro... ( 87fps at 1920x1080? ( what settings are used ? )

And what a nice list of games .. lol
 
Last edited by a moderator:
Mafia 2 is DX9 only


The list of games is a thing... the way they have test them with an high end card is even more, with all the respect i have for Tweaktown guys,... This article look to be made for a reconciliation with Nvidia ..
 
http://www.tweaktown.com/articles/4...2gb_video_card_performance_preview/index.html

looking to power plugs and length of the card it's probably a Gigabyte card

IMG_3329-2.jpg


IMG_3336.jpg


It is widely known that Tweaktown's reviews are very good to compare AMD cards with each other and Nvidia cards with each other. Nvidia/AMD comparison is a no go.

That being said, I find these results very strange. The 670 seems to be way too close to the 680 <-(edit). Having in mind the one less SMx, I expected much less.

The author of the article seems to be worried that this card will bring a lot of pressure on AMD. Imagine the 660 Ti! :D

Unfortunately these results suggest that the card will probably not come cheap. :(

Thanks for the news.
 
Last edited by a moderator:
That being said, I find these results very strange. The 670 seems to be way too close to the 690. Having in mind the one less SMx, I expected much less.

The results would line up with the assumption that 680 is rops-limited and shader-heavy. Which really shouldn't be that shocking, considering the amount of extra flops it tosses on over the past gen without a corresponding improvement on the rops. So if your shader units are not getting fully utilized, cutting of one of the shouldn't have that much effect.
 
Not only is the 670 way too close to the 680 overall, it's even on par with it in several tests, which it shouldn't considering it also has a ~100 mhz clock disadvantage.
Looks like some overclocked non-ref card, even though gpuz says otherwise.
 
The 670 has exactly the same amount of memory bandwidth as the 680. It makes perfect sense that it would perfom like a 680 if the test is bandwidth limited. And as we know, bandwidth is one of the 680's weaker points.
 
Unfortunately these results suggest that the card will probably not come cheap. :(

I think somebody at GAF already saw one at a Fry's for 399.99. they wouldn't sell it to him because it was before street date.

But that seems to confirm the $400 rumors.

Good luck finding one in stock anytime soon, anyway.

Will be interesting to see how AMD's pricing reacts, on the surface the 7970 will eventually have to end up at 400 or below. I have a feeling AMD will be stubborn about denying the inevitable for a while again though.
 
Back
Top