[Ryan Smith]

Originally Posted by kalelovil

All above the K2000M are N14E chips, so I assume cut down GK104s.

Indeed. And it's nothing out of the ordinary for NVIDIA to use severely cut-down GPUs in Quadro parts, though this traditionally applies to Gx100 parts.

[Psycho]

GTX 660 (OEM) turning up in retail. Looks like 192 bit confirmed.
http://www.proshop.dk/Stationaer-com...r-2344418.html

[Homeles]

Originally Posted by Psycho

GTX 660 (OEM) turning up in retail. Looks like 192 bit confirmed.
http://www.proshop.dk/Stationaer-com...r-2344418.html

I suspected it'd be that way from the beginning. The 670's neutering was very modest... there needs to be something more substantial to create a gap between the 670 and 660 (Ti or non-Ti).

[Man from Atlantis]

well, if 1152CCs and 192 bit is truth for full blown GK106 not 768CCs/192bit then something should have happened like Barts.. there were 2 Barts and one was canceled.. NV said they were disappointed by Tahiti but i think they wouldnt expect Pitcairn will be that strong too.. it would give hints about long delay as well..

[CarstenS]

Sorry for the self quote, but Steam seems to confim those numbers for their june data set.

Originally Posted by CarstenS

Well, not being steam surveys and thus having a rather limited number of participants are two quickpolls in german enthusiast forae are those two quickpolls. But apart from having a strong tendency of over emphasizing the high-end portion of the market, both communities see GTX 680 roughly on par with HD 7970 despite being on the market for a significantly shorter amount of time.

http://www.forum-3dcenter.org/vbulle...d.php?t=528393
http://extreme.pcgameshardware.de/qu...ni-2012-a.html

NVIDIA GeForce GTX 680 - - 0.31% 0.56% 0.74% +0.18%
ATI Radeon HD 7970 0.31% 0.38% 0.41% 0.50% 0.50% +0.00%

This does definitely not look like no one out there wasn't able to get a 680 in the past month.

[no-X]

Man from Atlantis: I don't quite believe the story about much higher expectations regarding Tahiti's performance. GF104 "done right" (fully enabled and with GF114-like clocks) would be also as fast as HD 5870. Both GF104 and GK104 were targeted relatively to the same performance level as competitiors high-end GPU. It's hard to believe, that Nvidia would be surprised just from one of these GPU.

[UniversalTruth]

Originally Posted by no-X

Man from Atlantis: I don't quite believe the story about much higher expectations regarding Tahiti's performance. GF104 "done right" (fully enabled and with GF114-like clocks) would be also as fast as HD 5870. Both GF104 and GK104 were targeted relatively to the same performance level as competitiors high-end GPU. It's hard to believe, that Nvidia would be surprised just from one of these GPU.

WoW, sorry to disturb, but I had much higher expectations about Tahiti and to me it is underwhelming. Why wouldn't NV think so too, without of course these technical details you are trying to protect your point? They are meaningless from time frame point of view.

[leoneazzurro]

Originally Posted by CarstenS

Sorry for the self quote, but Steam seems to confim those numbers for their june data set.

NVIDIA GeForce GTX 680 - - 0.31% 0.56% 0.74% +0.18%
ATI Radeon HD 7970 0.31% 0.38% 0.41% 0.50% 0.50% +0.00%

This does definitely not look like no one out there wasn't able to get a 680 in the past month.

It is a little strange the 7970 gained no share, especially with price going quite down on standard editions (in Europe, you could find a SE 7970 for around 50-60€ less than the 680) and the GHz edition launch. As usual, these results should be taken with a grain of salt... (and, as usual, there is no sign of 670 and lower Kepler parts in the list...)

[LordEC911]

Originally Posted by UniversalTruth

WoW, sorry to disturb, but I had much higher expectations about Tahiti and to me it is underwhelming. Why wouldn't NV think so too, without of course these technical details you are trying to protect your point? They are meaningless from time frame point of view.

There is simply no way in hell that Nvidia was "surprised" by Tahiti.
We knew almost exactly what Tahiti was about a year ago, July 2011.
Does anyone really believe Nvidia planned to compete against a full GPU family with a single ASIC for six months with a capacity constrained fab on a new process?

GK104 is Nvidia's RV670?

[no-X]

UniversalTruth: Well... We can't ignore, that GF104 had potential comparable to Cypress. Of course it came underclocked and partially disabled due to some complications, but that doesn't change anything on the fact, that GF104 was designed as a product offering some level of performance - in fact comparable to AMD's high-end.

The same applies to GK104 with the only exception - GK104 doesn't have GF104's issues. Performance target of GF104 and GK104 (relatively to AMD's high-end) is almost identical. It's really hard to accept, that Nvidia could be surprised in one case and not in the other one.

[lanek]

Originally Posted by no-X

UniversalTruth: Well... We can't ignore, that GF104 had potential comparable to Cypress. Of course it came underclocked and partially disabled due to some complications, but that doesn't change anything on the fact, that GF104 was designed as a product offering some level of performance - in fact comparable to AMD's high-end.

The same applies to GK104 with the only exception - GK104 doesn't have GF104's issues. Performance target of GF104 and GK104 (relatively to AMD's high-end) is almost identical. It's really hard to accept, that Nvidia could be surprised in one case and not in the other one.

You base yourself as the 560 was the initial design.. yes in term of SM, 1 was disabled.. but you do not know the clock speed they could have push 1 year before on Fermi fully enabled. And sorry but even in this case the 560 was not as fast of the 5870 .. clock to clock 1SM enabled on the 580 was lead only on a 1-2% increase on performance vs the 480... The real difference was in the clock speed.

If you base yourself on overclocked 560TI (not the second version ), you can too compare it with 570 so the 480 ... you was match easy a 570 with an overclocked 560.. so you think Nvidia have decided to design their 460 for overpass the 480-570 performance ?

Look the Asus GTX670 DirectCU II.... who is faster in 100% of the case of the stock GTX680 ....

[no-X]

reference GTX 560 Ti was comparable to HD 5870 (or even slightly faster):
1680 × 1050 + AA/AF: +4 %
1920 × 1080 + AA/AF: +1 %

source

There was no problem with clock-speed regarding 40nm process. This issue was related to inferior design of the metal layers as described by Jen-Hsun Huang. 40nm GF104 was clocked very closely to 65nm G94 / 9600GT - that could be hardly their initial target. 40nm process offered high clocks from the beginning, there was no significant uplift (850 MHz for Cypress, 880 MHz for Cayman 15 months later). GF104 was released almost a year after HD 5870, you can't blame TSMC for GF104's clocks.

Simply - if Nvidia didn't spoiled the desing and prioritized GF104 design over GF100, they would have direct competitor to Cypress. I believe they got it quite soon and decided to prioritize GK104 on this account.

[lanek]

I had look a too recent test maybe .. well different driver different games, different times

the 460 was run at 675mhz, on your link the 560TI at 822mhz ... You cant say the 460 with 384SP (full SM ) will have been set at 822mhz .... the card could have been set at 700-725mhz, who knows .. If cypress have been reused 1 year later, you could have got a different clock speed. instead of 850mhz, the card could have been released at 950-975mhz, as the leak in power due to production ( on same process ) will evoluate. AMD could have rework the design on transistor for get the same specification, but goes for higher clock due to tracking and modifiy transistor size, leak etc.

The 560TI was here for match a certain level of performance, at the time of the 460, it was aimed to another level. The 560TI was aimed at the 6870 performance, and the 6870 performance was around 5870 performance ( even if the design was a little bit different ). ofc the 6870 was not the 6970. Clockspeed is a good way ( when possible ) to match exactly the competitive part you are aimed at. 30mhz less and you are not looking competitive on review, 30mhz more and you look competitive.
Look the 7970 vs 680 .. if the card have been launch at 1000-1010mhz instead of 950mhz 3month before .. the 7970 vs 680 competition will have look totally different. at thoses speed the difference between both 680 vs 7970 will have been only of maybe 2-3% in average ( so impossible to say who is the best outside maybe TDP ). Even the driver change a lot the performance diffference, specially the ratio / resolution.. as some games where the 7970 was under at 680 release, the 7970 is now faster. and vice versa


The Sapphire OC run at 1000mhz .

[UniversalTruth]

Originally Posted by LordEC911

There is simply no way in hell that Nvidia was "surprised" by Tahiti.

Well, I agree NV had in their minds few expectations- like case A, B, etc. But what happened was one of the better possible scenarios for them.
Perhaps they expected GeForce 7 to GeForce 8 type of performance improvement from Tahiti over Cayman.

Originally Posted by LordEC911

We knew almost exactly what Tahiti was about a year ago, July 2011.

Did we know GCN efficiency (which is the surprise) and drivers maturity at launch, or what clocks it could have been pushed to...

[lanek]

Originally Posted by LordEC911

There is simply no way in hell that Nvidia was "surprised" by Tahiti.
We knew almost exactly what Tahiti was about a year ago, July 2011.
Does anyone really believe Nvidia planned to compete against a full GPU family with a single ASIC for six months with a capacity constrained fab on a new process?

GK104 is Nvidia's RV670?

Huum, the performance on games was absolutely unkown, but the GCN impressive list of computing aimed part should have not put Nvidia with a big smile on their face. the conference was more aimed at what GCN will capable of in term of power and raw power and on this part the card deliver totally.
With no error this is the first consumer card able to reach the 1Tflops theorical DP, and is actually the higher SP Tflops single parts never made.

GK110 is described on Tesla20 with 7.1billions of transistors and a 1Tflops DP ( non theorical i think ).. nothing about SP Tflops.. , but we dont know if the GK100 had the same spec. ( could have been less SP for less transistor at this time or Nvidia will have no problem to release it now ( and 7.1billions of transistor is really big numbers for a first 28nm cards ).

Seriously sometimes i ask me if Nvidia have not let down the GK100 for this reason... Is really the GK100 was so different of the GK104? surely .. but im sure the card was coming with lower clockspeed, compute
parts enabled, maybe a bigger memory controller, but not so much more SP ... I have some problem to imagine a GK100, bigger, with more TDP, who goes at 1100mhz turbo boost speed.


The great part of the Nvidia and AMD mid range parts, is they are in general more balanced at gaming ( maybe not for other task, but for gaming ).. its where you get a 7870 who goes at 7950 level, or a 560TI who can reach 570 level performance just by up the clock speed... they can be clocked high.. really high in some case.

[Kaotik]

Originally Posted by UniversalTruth

Did we know GCN efficiency (which is the surprise) and drivers maturity at launch, or what clocks it could have been pushed to...

Yes, it was quite surprising how efficient it is in gaming considering how much GPGPU "bloat" it has.

[jimbo75]

Nvidia could just have got lucky with GK104 - that's also an option that nobody seems to mention.

With GK107 looking like poor competition, GK106 months late (possibly due to being poor competition), and GK100 nowhere to be seen...why assume otherwise?

But 1/4 ain't bad right? Except AMD is constantly releasing new series top-to-bottom before them, but nevermind that when you still hold the mindshare with a couple of cards that are actually really terrible value?

Is there any reason why GK104 didn't just get lucky, and Nvidia really are still an absolute mile behind everywhere else?

[lanek]

a good question can be while the GK104 was allready 3-4 month later of AMD product if this was "just" a middle range product ... why GK100 have never exist.. and if this was just cause Nvidia decided their high range sku is not needed for gaming why they have not release it for professional sku and instead use 2x GK104 with extremely poor performance in DP and SP as their first Kepler professional and computing sku ..

I cant believe they have miss by a so high margin the promise they was do and show us in the graphic as projection for show us Kepler 3x times faster in SP of Fermi .. Where is the Kepler they was promise to us ?
They can use and over use the marketing artifice we know from Nvidia, but where is the Kepler they was promise us ?



Look Maxwell and Fermi - Kepler in this graph, and look what is Kepler today, they dont really think we will believe they be able to multiply the performance by watt on maxwell with DP by 8x ? actually Kepler is under Fermi in DP rates by 2 times not 2x more ... Nvidia is absolutely not in track with what they was promise to achieve.. they are really far and far of it...

I wish allready a good luck to Nvidia to achieve 8x time the performance of Fermi in 2013 with Maxwell... They are late on tracks and they know they will never deliver what they was promise...

Dont try to make me believe, Nvidia have not keep the track on their professional tesla and computing parts and use the gk104 for them finally, cause AMD was not deliver ... its like said they think peoples who buys tesla computing parts are completely dumb and dont count for them .. . Nvidia dont care of them ... they better like release a good gaming card....

[I.S.T.]

The card you're looking for is GK110.

[sonen]

oh come on... that's an ancient pre-Fermi roadmap

http://www.nvidia.com/docs/IO/113297...ne11-Final.pdf
(page 14)

[lanek]

Originally Posted by I.S.T.

The card you're looking for is GK110.

The GK110 you are speaking about or what peoples think is the GK100 should have been first released in 2011 and now will not appear before 2013 .... the GK 104 as Nvidia have call it and have need be released in a dual version for ensure performance in SP computing parts based on Kepler, this is still not the GK100 Kepler promised ...... you are speaking as a gamer point of view who think Nvidia have cut down his Kepler version... ( as they was dont need more ( lol ) ) ... but why should dont need more of this shame part who is the GK104 on this aspect ? The GK104 need to be in a 2 cores version for be better on Tesla cores releasing 2years ago, and even there, the Tesla10 cant even match the performance in DP of the old tesla .. ( lets alone the AMD gaming parts with his 1Tflops DP )

I dont think the GK100 was made of 7.1 billions of transistor as will be the GK110 `(If this was the case, this explain why it have never been released, I dont think Nvidia want to try the Fermi adventure a second time, and in harder position ) ... impossible in a first attempt on 28nm ( and we know how bad have been by the past 10years Nvidia on node transition ) Where Nvidia is really good is to take an existant node architecture and take the maximum of it ... not create a new architecture on a new node at the same time .

[ninelven]

Originally Posted by lanek

The GK110 you are speaking about or what peoples think is the GK100 should have been first released in 2011

Nvidia can't really do TSMC's job for them. If you think GK104 had supply problems and was expensive for its die size... yeeesh.

[I.S.T.]

Pretty much. Right now TSMC doesn't have quite enough room for Qualcomm, AMD and Nvidia.

[Davros]

passive gtx680
http://www.theverge.com/2012/6/27/31...-graphics-card

[vking]

GK100 was canned very early in the cycle (much before anything about Tahiti was known) by Nvidia in favor of accelerating GK110 with full compute feature set.

Nvidia intended to fight AMD with GK104 until GK110 was ready. Nvidia got lucky due to the fact that AMD paid compute tax via Tahiti. Remove compute and AMD is actually slightly ahead of Nvidia on most metrics.