-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
ATI cuts 6950 allocation
- Thread starter An Exception
- Start date
Good to know there will be plenty stock of the 6970. For those thinking will the 6970 be faster than a 580, I got news for you, how does 30-40% faster at stock sound.
That will be great news. People are getting a bit tired of Nvidia and ATI releasing newer cards that are marginally faster than the previous generation of cards.
It will be crucial for ATI to provide functional and optimized drivers for reviewers before Cayman is released to actually show how well it will perform. I had myself convinced to wait for 28nm before upgrading but it seems ATI and Nvidia are doing their best to change my mind.
>Tinfoil hat on.
Did your source really mean compared to a 480? or was he/she really talking about the 6990?
Even with those corrections 40% would be impressive.
Associate
- Joined
- 10 Jan 2010
- Posts
- 480
But how much OC headroom will that leave the 6970?
If similar default clocks to 580, then it will indeed be an impressive achievement.
But factor in higher clocks & IQ tweaks and we have another story.
I suspect the 580 could have been launched with higher default clocks but what makes a good enthusiasts product ......good overclockability
If similar default clocks to 580, then it will indeed be an impressive achievement.
But factor in higher clocks & IQ tweaks and we have another story.
I suspect the 580 could have been launched with higher default clocks but what makes a good enthusiasts product ......good overclockability
The 580, the optimization on the shaders on the 69** has yielded impressive gains over the 68**, factor that in with the 70%+ shaders over a 6870 and you got 580 busting performance.
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
So in other words, the world renowned GPU manufacturer with the best brains in the business screwed up, big style
So whats changed other than time
So will the chip suceed on 28nm
You are effectively implying that after this process, Fermi is dead, back to the drawing board.
Surely this can't be the case
Roll on 28nm/2011
Doesn't matter what brains you have working at a company when the CEO insists on making the biggest and best GPU physically possible, even if yields will be in the toilet.
I have no idea what you are asking has changed other than time, the 580gtx is a BASE LAYER RESPIN, of 480gtx, then you ask whats changed other than time, I bolded it in this very sentence to give you a clue.
A respin can be significant, or insignificant, you might add in one feature, fix one small cluster of transistors that weren't yielding, or change the entire metal layer, you can use a more expensive metal, in the base layer or other layers, you can do many things. I don't know what they did, no one knows, but the architecture, from the top level view, is IDENTICAL to the GF100, Nvidia call it a GF100B internally, it acts exactly like a GF100b + 10% clocks and full shader compliment.
The last part makes no sense either, Fermi as is, is a 40nm design, it won't be used at 28nm, so yes, its dead, whats new in that? A 3870 is similar, but its NOT a 2900xt, a 4870 is similar but its NOT a 3870.
At 28nm Nvidia should be aiming at arounda 1000 shader(the natural aim would be 1024 obviously) the problem is, again that a 1024 shader "fermi" design simply expanded, will end up as a 500mm2+ chip, its size was its problem at 65nm, at 55nm(it was what, 8 months late), there was a plan for a 40nm 285gtx with a few more shaders, they couldn't get that working due to their lack of knowledge of the process, size wise on 40nm, that chip should have been aiming at 350mm2 or so and they still couldn't make it. Right now the GF104, what 18 months or so after the first 40nm chips started sampling(4770) the GF104, a 380mm2 chip STILL couldn't get full yields. Architecture isn't the problem, with a full base layer respin they've finally got a full yield of a 530mm2 part.
Its there problem hitting new process's in general, but the bigger the core, the worse these problems are. 28nm, sure makin ga 512sp "fermi" will be easy, but by then, that will also be the mid end, the same way you'd expect a 1200-1600shader mid range AMD part at 28nm, it will be half the size.
The problem at 28nm is you expect to double shader size, and this maintains that high die size, and thats fundamentally where Nvidia's problems stem from.
Fermi's entire achitecture, and Nvidia's since the 8800, has been single shaders, each with its own core logic, each capable of one operation per clock, the reason AMD have a smaller core is they use a 5(or about to be 4) shader arrangement, it cuts down on core logic and size dramatically. 8800 through 580gtx are big, because of the single shader, simple structure design, you can't make simple single shaders small, you just can't, theres nothing wrong with it, its incredibly easy to leverage the power from, but the downside is size and how many you can fit. This wouldn't be a problem if TSMC could make any size core you want with high yields, but, they can't.
Architecture didn't change that much, but it DID change from 280, to 480, and the 285gtx was on a different process which takes a lot of work, yet all three have had yield problems, all three had issues, the very simple problem is they want to make a HUGE core, and TSMC can't make a HUGE core easily, or with high yields and especially early on in a process's lifetime.
Thats another change between the 480-580gtx in terms of time, TSMC improve a process from introduction to end of its lifecycle, though 90% of the improvement will come in the first few months, the last 10% can be the difference between a 480sp and 512sp Fermi yielding or not.
Its almost certain Nvidia will hit issues if they go for another 500mm2+ core at 28nm, and they'll likely do the same at every process, and AMD would have problems if they went with a core that big, thats why they simply decided not too. Remember Fermi/285gtx type problems were what AMD had with the 2900xt, their last big core, which is why they changed their design goal to never hit a huge core again, and since then, they've been on time, with high yields, incredibly competitive, cheaper and gaining market share, because of one decision, to not get caught up with TSMC sucking balls at every single process.
It's a guy that works for a very well respected AMD AIB, he has 6950/70 samples.
Yeah I said wouldn't but that kind of performance is just too much to ignore.
Last edited:
Sorry if I am miss understanding you but I don't think you can compare Nvidia/ AMD clock for clock.
Well if we see any fair/comparable benchmark showing fps 35% over a 580, I will put anything you want in my sig as long as it doesn't go against any board rules.
I'm pretty much set on getting a 6970 so I hope its true but I really find that hard to believe.
Associate
- Joined
- 10 Jan 2010
- Posts
- 480
Thanks drunkenmaster
That was a good, informative read.
It's a good job you don't write DIY manuals
You begin to wonder how Nvidia are still in business!!!
That was a good, informative read.
It's a good job you don't write DIY manuals
You begin to wonder how Nvidia are still in business!!!
so a guy you know, who works for a very respected AMD AIB, has passed some info on to you,
surely thats more than his jobs worth ?
Associate
- Joined
- 10 Jan 2010
- Posts
- 480
He's just started with them, they got some samples in and he has seen them benched, it was just a casual email from him pointing out some benchmark scores and I did the math. There is nothing I have posted to incriminate him, if I posted up some results then that might but obviously I will not.
Soldato
- Joined
- 16 Jan 2003
- Posts
- 10,819
- Location
- Nottingham
If this is true and a big if, it just shows how bad Nvidia are at designing GPUs if AMD can produce a GPU that is a fair bit smaller than Fermi and yet be equal or even faster.
- Joined
- 8 Aug 2010
- Posts
- 6,453
- Location
- Oxfordshire
Hmmm, maybe a R300 after all???
- Joined
- 8 Aug 2010
- Posts
- 6,453
- Location
- Oxfordshire
That's correct you can't...
Also clocks can be affected greatly by yields and not just architecture, so if yields are good then...
- Joined
- 8 Aug 2010
- Posts
- 6,453
- Location
- Oxfordshire
Depends how good yields are, and if TSMC have increased it's capacity to allow availability to be maintained that will actually enable AMD to gain market share by selling at lower prices.
If TSMC has indeed increased it's capacity and yields are good, then that set's the stage for a price war.
Soldato
- Joined
- 16 Jan 2003
- Posts
- 10,819
- Location
- Nottingham
Yields have to be good if they are reducing 6950 allocation, they don't want to be selling too many fully functional chips as 6950s if they can sell them as 6970s