This would appear more likely in a vacuum, but the tone of the marketing campaign suggests something big.
iMacmatician
Regular
I had a thought that one major reason for the supply drought is because they're stockpiling GK104's for the 690 but I doubt that since I presume the 690 would be sold in vastly lower quantities than the 680.
From looking at past dual GPUs and asking elsewhere, they could likely get a dual-GPU consisting of two 680s into 375 W. Assuming linear scaling, two 680s at 875 MHz, 1344 CCs, and 4.57 GHz memory would get 289 W and that's before any other power savings the dual GPU provides. Considering the 5870 ran at 188 W and a lower clocked version of it formed the 5970, it shouldn't be too hard for a 195 W 680 to do the same. However I think they will go straight to a 375 W dual GPU, (among other reasons) so the 7990 won't beat it when it comes out (or if the 7990 wins, then not by much). There could even be two dual-GPU models with two different sets of specs (like the 3850 X2 and 3870 X2) but that would be a first for NVIDIA.
I dont think it will make a big difference in term of availbility, like you said, the ask about thoses type of product is really low.
Without saying, shops dont need to have big stock for them.. ( the price will end a little bit higher too of what GTX590-6990 was ).
Silent_Buddha
Legend
Well that really only leaves a dual GPU card, IMO. I'm expecting whitepapers on GK110 some months before its launch similar to what was done for Fermi in order to prep the HPC and GPGPU guys.
So, for a LAN event, I'd agree that midrange is probably out, but you never know. Hence, I'm thinking it'll be a dual GPU card at an exhorbitant price.
But who knows maybe they'll pull a surprise out of their bag.
Regards,
SB
iMacmatician
Regular
NVIDIA has listed the GT 630 (OEM), the GT 640 (OEM), and the GT 645 (OEM) on geforce.com.
The GT 630 is GK107 DDR3. The GT 640 comes in 3 versions: an apparently renamed GT 545 DDR3 (except for a tiny memory clock change), a GK107 DDR3 version, and a GK107 GDDR5 version. The GT 645 looks like a renamed GTX 560 SE.
Fudzilla claims that the GTX 670 will launch at the same day as the GTX 690, May 10 (they still say an April 28 announcement for the 690).
There's also some stuff (I don't know about its reliability) floating around (such as this).
The GT 630 is GK107 DDR3. The GT 640 comes in 3 versions: an apparently renamed GT 545 DDR3 (except for a tiny memory clock change), a GK107 DDR3 version, and a GK107 GDDR5 version. The GT 645 looks like a renamed GTX 560 SE.
Fudzilla claims that the GTX 670 will launch at the same day as the GTX 690, May 10 (they still say an April 28 announcement for the 690).
There's also some stuff (I don't know about its reliability) floating around (such as this).
That's about the same drop in TDP, percentage-wise, as the GTX 450 was over the GTX 460 [1 GB] and the GTX 550 Ti was over the GTX 560 Ti. If that's the case I hope the 660 has a higher increase in clock speed over the 680 than the 550 Ti had over the 560 Ti….Google Translate said:
Last edited by a moderator:
So we should expect GTX 650 @ >1GHz, since GT 640 GDDR5 is already very high clocked with 950MHz.
And why is GT 640 DDR3 slower than GT 630?
The chinese GK106 news blurb is just a repost of this old speculation.
And why is GT 640 DDR3 slower than GT 630?
The chinese GK106 news blurb is just a repost of this old speculation.
iMacmatician
Regular
From NVIDIA: "The Impact of Process Technology on Kepler's Efficiency."
Well that was a whole lot of words to say very little. It seems to be a trend in corporate blog posts.
A1xLLcqAgt0qc2RyMz0y
Veteran
Actually there is some nice details in there if you read past the first three paragraphs:
I did, and the figures are interesting—if somewhat misleading—while the rest is pretty much expected.
I don't think you are part of the main target audience of that article.Alexko said:
UniversalTruth
Veteran
lol, guys, seriously.
To me the first thought after reading the title and before clicking the link, or my expectation was that NV were not happy with the new process and they would complain, again.
Those numbers look so low...
A1xLLcqAgt0qc2RyMz0y
Veteran
Explain?Those numbers look so low...
That is a fantastic number. Most of the time leakage increases when going to a new node.
Leakage was getting pretty bad at 65nm and below.
The 45/40nm generation was pretty much at the wall when it came to having a practical process in terms of leakage.
It follows that leakage improved so much with the jump to HKMG, which provides an initial improvement. Without additional technologies, the next node would have an increase again (Intel keeps adding new process tech nodes ahead of the others, and enjoys serious benefits, for a reason).
The peak (edit: active) number didn't reduce by so much. This is part of the reduced voltage scaling we are seeing, and also why you don't get as much mileage out of a shrink if the design doesn't change too much.
The decision to peel off the additional scheduling layer in the 680 is one way to improve overall peak power by reducing the amount of active silicon needed for that part of the pipeline.
The 45/40nm generation was pretty much at the wall when it came to having a practical process in terms of leakage.
It follows that leakage improved so much with the jump to HKMG, which provides an initial improvement. Without additional technologies, the next node would have an increase again (Intel keeps adding new process tech nodes ahead of the others, and enjoys serious benefits, for a reason).
The peak (edit: active) number didn't reduce by so much. This is part of the reduced voltage scaling we are seeing, and also why you don't get as much mileage out of a shrink if the design doesn't change too much.
The decision to peel off the additional scheduling layer in the 680 is one way to improve overall peak power by reducing the amount of active silicon needed for that part of the pipeline.
Last edited by a moderator:
Looking at that blog, I'm totally underwhelmed.
http://www.realworldtech.com/page.cfm?ArticleID=RWT012707024759&p=4
When Intel went from 65nm bulk to 45nm HKMG (both using strained silicon), they got:
30% lower active power
upto 5X lower sub-threshold leakage
10X lower gate leakage
When I make that comparison, it seems like TSMC+Nvidia did alright with active power. However, the leakage improvements were not nearly so large as I would have expected.
I wonder what is going on there...
DK
I didn't see what Nvidia's methodology was for calculating its numbers.
It's possible they are making a comparison between a test architecture or theoretical equivalent to a design like Kepler between the 40nm and 28nm nodes.
Hot-clock was tossed, but now the global clock has risen significantly, which could eat into savings compared to a design that did not increase transistor switching speeds.
Perhaps in addition to this, TSMC may have had to make some tradeoffs Intel did not.
The last update I saw on numbers for its process was probably the old article on RWT, and perhaps reality has intruded since then.
Perhaps they couldn't get Nx savings in various forms of leakage at the same performance, and TSMC and Nvidia had to eat into savings to prevent a regression.
It's possible they are making a comparison between a test architecture or theoretical equivalent to a design like Kepler between the 40nm and 28nm nodes.
Hot-clock was tossed, but now the global clock has risen significantly, which could eat into savings compared to a design that did not increase transistor switching speeds.
Perhaps in addition to this, TSMC may have had to make some tradeoffs Intel did not.
The last update I saw on numbers for its process was probably the old article on RWT, and perhaps reality has intruded since then.
Perhaps they couldn't get Nx savings in various forms of leakage at the same performance, and TSMC and Nvidia had to eat into savings to prevent a regression.
DavidGraham
Veteran
Are you saying we are advancing all too quickly ? should we research more before announcing a new node ?
It's not a question of being too fast, just that Intel has had the luck, resources, and hard work invested in putting in technologies like HKGM and multigate transistors before scaling smacks into a wall.
So far, each of these is a mostly single-shot improvment. Everyone else's 28nm is HKMG, but without something like trigate, it's just putting them somewhat back on the same curve.
Its process competitors eventually follow suit after a lag time that so far has only shown signs of increasing. So far, the outcome is that they get these absolutely necessary fixes only after the problems have already afflicted them, sometimes badly.
It's sort of the difference between ducking 1 second before and 1 second after someone swings their fist at you.
So far, each of these is a mostly single-shot improvment. Everyone else's 28nm is HKMG, but without something like trigate, it's just putting them somewhat back on the same curve.
Its process competitors eventually follow suit after a lag time that so far has only shown signs of increasing. So far, the outcome is that they get these absolutely necessary fixes only after the problems have already afflicted them, sometimes badly.
It's sort of the difference between ducking 1 second before and 1 second after someone swings their fist at you.
DavidGraham
Veteran
Still I find it hard to believe Intel maintains it's luck every time , If it's a question of resources , then why don't the other foundries throw more money to acquire the necessary edge .
I am not a conspiracy theorist myself , but sometimes I have to question Intel's constant higher technological capabilities , not to mention it's past questionable practices that were not actively punished or recalled in any way that matters , Sometimes I feel that the company has ties to certain governments that grants them resources and preferential treatment!
I am not a conspiracy theorist myself , but sometimes I have to question Intel's constant higher technological capabilities , not to mention it's past questionable practices that were not actively punished or recalled in any way that matters , Sometimes I feel that the company has ties to certain governments that grants them resources and preferential treatment!
