News - It's Coming, NVIDIA announcement at 03:30 April 29th

[watercooled]

I'd really love to cling to the hope that this will be an accouncement of a GK110-powered card, but I get the horrible feeling that it's going to be a dual GK104-powered card.

It's pretty much confirmed to be a dual-GPU by the looks of it.

Don't get me wrong, the GK104 is a powerful card in its own right, but however you look at it, it was meant to be a high mid-end chip, not the highest-end chip. AMD did a lot of good catching up with Tahiti but I think a lot of people were hoping the performance gains would be more substantial.

I feel like a broken record today.
GK110 may be a larger GPU based on the same architecture but that doesn't imply at all that it was intended to be released at the same time. Yields are far greater with smaller dies and considering the trouble Nvidia have been having with GK104, GK110 would've been a nightmare on the still relatively new 28nm process. This dual-GPU release goes further to confirm that - Nvidia would much prefer to offer a single-GPU which beats everything on the market over a dual-GPU card along with its associated driver/microstutter/heat problems. This 'the 680 is only a mid range card' is just fanboy/Nvidia created nonsense - the 680 *is* their current fastest card, regardless of what numbers have been splashed about regarding other dies, and we almost certainly won't see a bigger die until the 780; 680/dual GPU naming doesn't leave much space - calling it the 695 makes it sound like an improved 690 and considering it will probably perform between the two, a 685 name would make more sense but wouldn't be terribly good for marketing. And we likely won't see the 780 until around the time AMD have their next line out, unless something massively improves yields over the next few months. Otherwise, you have a ton of partly failed dies, which would have to be sold under yet another name, (GTX682.5?) if Nvidia had intended to release GK110 soon after the 680 they would have left some naming room for it and the part-disabled dies.

[shaithis]

Well, call me a fanboy...but he is right. The 680 was not supposed to be a range-topper. Just look at all previous releases, model numbers and specs and then compare. It's blatantly obvious (at least to me!) this was supposed to be a 660Ti.....the biggest 2 clues are GK104 and 256bit memory interface.

Why release a card for £400 with a small margin when your competitors allows you to release the £200-250 card @ £400 and milk the market for months?

Or do you believe nVidia have decided to give the middle finger to high-end gamers with large/multiple screens by limiting bandwidth in their newest range?

Look at the evidence. It's why I have refused to buy any of these new cards, all of them are horrendously over-priced.

[watercooled]

Early intentions mean nothing - fair enough Nvidia may have intended it to be a mid-range card (we don't know that for certain) but that doesn't mean they were capable of releasing GK110 either way - as I said, look at the yield problems they're having with the much smaller GK104 die. And as I also said, they wouldn't have known before release how AMD's 7970 would perform, they wouldn't risk cancelling production of their high end part and produce a mid-range chip based purely on speculation. People seem to think it takes 5 minutes to program a new die into a production machine and press print - the timeframe between AMD's release and Nvidia's is far too small to have changed the silicon last minute. All the naming implies is they intend to release a larger chip with a similar architecture at some point. Far too much is being read in to what is just a name, for all we know Nvidia could have changed it like that for marketing purposes? And as for the 256-bit interface, maybe a wider bus didn't add much relative to cost/die size. Looking at actual evidence rather than rumour/speculation/marketing is how I drew this conclusion.

If die naming tells us so much, wouldn't 110 mean it's the next gen over the GK100 series, like how the GF110 was the improved GF100 die (580 vs 480)? I.e. Major.minor.range.

Either way it makes no difference, the 680 simply is the high end GPU for the 600 series, whether it was intended or not.

[Hicks12]

exactly watercooled, i wish people would let the whole "not high range" issue go... THERE was an intention to release a higher chip, but thats the same as aiming high and only achieving that 80%... as WC says it could be purely marketing as they changed the name! But There is way to many variables to account for it, yeilds is a strong reason and time frame is another and it may have simply been that Nvidia engineers couldnt do it, feasibly impossible with the die size and power requirements that they were allowed to play with.

Its incredibly hard to just 'change' silicon at any time, it takes alot of time and Nvidia had no real time, maybe we will see it at the end of the year or next gen. I dont think Nvidia are milking this gen by much, AMD and nvidia are pretty much neck and neck in terms of price vs performance and they both have had heavy dev cost and really its been more than last gen so its understandable that its getting less cost effcient in process jumps. I feel we are hitting the dimishing return, much like audio... a £20 headphone would sound ~ double that of a £10 but a £50 headphone wouldnt generally be 5 times better. We have hit that point so its really just a throw more money at it, charge a bit more and perform more , unfortunate for us but competition is reducing profit margins and bringing it slowly down to a similar level.

The gtx680 is a high end chip, their naming structure could of changed but regardless it meets the main rules of a high end gpu, its an expensive chip to make and sells for a high amount and it has the highest performance that Nvidia could achieve in a single gpu design at the time of release, its yields were sufficient but werent amazing (due to chip size and new process). its performance competes with the competitiors flagship card, it is a flagship card.

Can we leave the its not the flag ship ideal now? Yes Nvidia may have plans for a better gpu but its a long way off.

Shaithis id say the reason why you can avoid this generation is that the gtx580 is a solid card regardless! Alot of tech in that chip and you got it when costs were lower . Its hard to compete with that as this current gen is expensive to make and until yields improve and it becomes more mass market and cheaper to use it wont be changing much, all thats happening is AMD and Nvidia profit margins getting squeezed much more to make more competitive.

[shaithis]

256bit bus.

If you still think this was meant to be the high-end card....well, you don't understand the need for bandwidth in large resolutions.

But hey, if you think it fine and normal to charge £400 for a 1080p graphics card, be my guest.....but please do not infer it's designed to be flagship....., or call nVidia buffoons for not seeing that there was something so obviously wrong with their design.

Either nVidia are milking a mid range card @ high end prices....or they completely messed up and got the spec wrong...yet still made the fastest single GPU in the world with the wrong spec......

[kalniel]

256bit bus.

If you still think this was meant to be the high-end card....well, you don't understand the need for bandwidth in large resolutions.

Bandwidth. Not bus size. I've yet to see anything suggesting it matters how you achieve that bandwidth. (For eg, the bandwidth of the 680 is no less than the 580).

But yes, it's clear this wasn't their 'biggy', and it's also clear that in the absence of any biggy, you judge based on what's out there today.

[shaithis]

Same bandwidth + faster core = bottleneck waiting to happen.....and I am 100% sure nVidia do not need me to tell them that.....

Or you judge that they are fleecing you and wait.

People buying 680s are going to be ill the day the real 680 gets released......the price drops are going to have to be insane.

[watercooled]

Shaithis id say the reason why you can avoid this generation is that the gtx580 is a solid card regardless!

Exactly, it's very rarely worthwhile upgrading from one generation to the next, less so for high end end less again when you picked up a high end card for a good price later on. I'm in a similar boat with my 5870 which I picked up for about £140, actually less than the ~£160 I paid on my previous 4850 so it was incredibly good value for me, but even now cards costing about £150 aren't much faster, more expensive ones don't offer relatively good value and TBH even the high end ones aren't going to make a difference like I observed with my last upgrade.

256bit bus.

If you still think this was meant to be the high-end card....well, you don't understand the need for bandwidth in large resolutions.

Bandwidth. Not bus size. I've yet to see anything suggesting it matters how you achieve that bandwidth. (For eg, the bandwidth of the 680 is no less than the 580).

Couldn't have put it better myself. For example, adding bus width increases die size, increasing clock speed does not. Nvidia, or any chip MFR, would prefer a less costly method of increasing bandwidth, the only downside of lowering a number on a spec sheet is you get people complaining they've taken a step back.

Edit Reply:
Not necessarily a bottleneck, it depends on the rest of the architecture. Sure, if you'd just bolted a load of new cores onto a Fermi die it would be an obvious problem, but Kepler is different enough to need to be reconsidered. Memory may be a bottleneck in some applications, as it will with any card. If there were never any bottlenecks, you'd have infinite performance...

You always get fleeced with high-end cards on release. But I feel this is all wasted text considering you're still referring to 'teh realzz 680'.

[watercooled]

But hey, if you think it fine and normal to charge £400 for a 1080p graphics card, be my guest.....but please do not infer it's designed to be flagship....., or call nVidia buffoons for not seeing that there was something so obviously wrong with their design.

Either nVidia are milking a mid range card @ high end prices....or they completely messed up and got the spec wrong...yet still made the fastest single GPU in the world with the wrong spec......

What is it then, a measly 1080p card or the fastest card in the world? As I said, it doesn't matter if it was designed to be flagship or not; ATM, it simply is. They are milking a high-end card at high-end prices as is usual...

[shaithis]

But it's really a 660Ti lol

Now, how much would a 660Ti have cost? £250?
670 @ £320?
680 @ £400+

nVidia have just finally got around to giving nVidia users a version of eyefinity yet at the same time designed the card to be poor with large resolutions and multiple displays. For a very profitable company with the fastest card around....they sure do mess up don't they?

Not necessarily a bottleneck, it depends on the rest of the architecture. Sure, if you'd just bolted a load of new cores onto a Fermi die it would be an obvious problem, but Kepler is different enough to need to be reconsidered. Memory may be a bottleneck in some applications, as it will with any card. If there were never any bottlenecks, you'd have infinite performance...

Did you even look at the benchmarks? It is a serious bottleneck. Look at resolutions 1200p and below and compare with previous generation. Then look at 1080px3 or 1600p benchmarks and compare with previous generation.

Now, go and look up the bus width and bandwidth on all those cards and tell me what the deciding factor is in performance trailing off so badly at those higher resolutions.

Now define "high end gpu"....because to me a 1080p/1200p card is NOT high-end.

Kind of makes sense why they are releasing this "half fat" 690 instead of the Gk110 as well. If they released the GK110, the GK104 would have to be slashed (and many of the cards in the 5xx range as well).....yet releasing this 690 allows them to carry on fleecing people for 680s...and double-fleecing 690 purchasers.

[watercooled]

I've seen plenty of benchmarks and know where the bottlenecks lie, but it doesn't change anything, I could say the same about any card, take the 7970, e.g. it's mostly frontend bound so beef it up a bit and performance would increase by a fair amount because of the excess of shader power, especially at lower resolutions.

But that's all besides the point.

A high-end GPU is one which is sold as a high-end GPU, i.e. marketed/named/priced as one. This fluff about 'real 680' is purely academic, we have what's available on the market - making buying decisions based on what's planned for future makes as much sense as ever.

[CAT-THE-FIFTH]

Thanks goodness I don't play BF3 - its saved me a lot of money!!

[Hicks12]

Shaithis can i just comment on the last part of your post

yet releasing this 690 allows them to carry on fleecing people for 680s...and double-fleecing 690 purchasers.

Have a quick look on scan bud and you will see that is a bit silly, there is ONE gtx680 instock... Nvidia have next to no chips shipped stock is terrible and so why would reducing their sales volume by half help? It doesnt, there is no new card coming any time soon and Nvidia are just trying to hype up their brand.


You can keep saying that bandwidth is an issue but that is a limitation, you dont know how much it costs for Nvidia or AMD to make their chips, its not a cheap as you would think and coupled with poor stock theres isnt much in it at the minute, cost is effecting what the companies can do with their chips and if AMD threw more money into their 7970 it would be even bigger and badder but no they have to keep it profitable they're a business, they dont want to make a loss (as AMD have been for awhile!) just as much as Nvidia dont, they were most likely given strict guidelines and both companies did the best they could with each resource.

Theres always going to be a better way to do something and so the gpus will always get better BUT its increasingly more difficult to do that and its the reason why we dont seem to get anymore of these 8800GT wonders anymore .


As an end to my post, i will say that if Nvidia really did have the higher model ready then they would release it, they wouldnt reduce the cost of the 680 as theres no need and they would simply charge this new high end card at a relative price (e.g. double the performance of a 680 then double the cost) and then they would be able to make a killing for maybe two generations if its really that 'OMFGZ AMAZING UBER KILLING GPU'.

Its a bad business strategy to be holding back that chip if they actually have stockpiles, and if they dont then its obvious the chip is no where near ready for consumer use.


PS: if the gk110 does come out (somehow) can i steal your 580? my 460 is getting a bit tired , i could do with a stronger card haha.

[aidanjt]

I already don't care.

[anselhelm]

I feel like a broken record today.
GK110 may be a larger GPU based on the same architecture but that doesn't imply at all that it was intended to be released at the same time. Yields are far greater with smaller dies and considering the trouble Nvidia have been having with GK104, GK110 would've been a nightmare on the still relatively new 28nm process. This dual-GPU release goes further to confirm that - Nvidia would much prefer to offer a single-GPU which beats everything on the market over a dual-GPU card along with its associated driver/microstutter/heat problems. This 'the 680 is only a mid range card' is just fanboy/Nvidia created nonsense - the 680 *is* their current fastest card, regardless of what numbers have been splashed about regarding other dies, and we almost certainly won't see a bigger die until the 780; 680/dual GPU naming doesn't leave much space - calling it the 695 makes it sound like an improved 690 and considering it will probably perform between the two, a 685 name would make more sense but wouldn't be terribly good for marketing. And we likely won't see the 780 until around the time AMD have their next line out, unless something massively improves yields over the next few months. Otherwise, you have a ton of partly failed dies, which would have to be sold under yet another name, (GTX682.5?) if Nvidia had intended to release GK110 soon after the 680 they would have left some naming room for it and the part-disabled dies.

I can see what you're saying there, but the fact is that GK110 is the Kepler generation and not next generation, so it's natural that people want what Nvidia has created.

The problem is just that the chip clearly isn't ready yet or is having terrible yields, hence why we got the mid-high-end chip GK104 being launched as its flagship.

I wonder what what happen if Tahiti had been so much noticeably better that Nvidia had had to respond in kind. I don't imagine the GK104 would be able to be branded as the GTX 680 if AMD's 7970 weren't quite so much of an architectural letdown, relatively-speaking.

For the record, my last three cards have been AMD, so I'm hardly an Nvidia fanboy, but when a mid-high-end chip can beat a flagship chip, there's a problem there. That Nvidia hasn't produced the GK110-based card speaks less about Nvidia and more about AMD, as Nvidia simply didn't need to get that out the door without the real competition.

[watercooled]

Competitor has a problem, blame AMD. That's something I heard yesterday too. xD

GK110 is, as far as we know, Kepler (as the K in GK would imply). But again, as far as we know and if they were continuing existing naming, GK110 would imply the second revision, like how GF110 was the second revision of Fermi (re-read my post, that's exactly what I said). We'll never know what would've been the case now if the GK104 couldn't compete, but I have no doubts whatsoever it still would have been the first chip released. Maybe the GK110 would have been rushed out shortly after, but even so it's unlikely what with yields even on the much smaller GK104, it would have been terrible for profits and image so I suspect they'd have settled with second best until TMSC could improve yields.

7970 is nothing like an architectural letdown, I'm not sure where you pulled that one from, it beats the previous generation by significant margins (similar to previous generations) in both performance and efficiency and is an incredibly powerful compute platform. Not to mention it was both the first 28nm GPU and the first based on a completely new architecture, for sake of comparison, look at the problems Fermi (480) had when it was released, and that used an already established production process.

But TBH I feel I could answer most of these replies by looping back to by previous responses, it all seems remarkably familiar.