GF100 (Fermi) previews and discussion

Page 14 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Toggle sidebar Toggle sidebar
Status
Not open for further replies.
But they didn't do their homework while ATI did.

ATI knew almost a year ago that TSMC's 40nm process was whacked in serious ways. That was much of the reason why the 4770 was created - to figure out the issues.

Almost a year ago. Nvidia *still* don't have a 40nm card anywhere near as powerful as the 4770.

All of the 40nm Nvidia cards have lower clocks and perform terribly - soundly beaten by their equivalent ATI parts. One of the reasons is, Nvidia didn't do their homework - they didn't double up the via's on the chips and that is almost an unforgivable error.

Nvidia are in serious, serious trouble if this stuff can happen because it's bad decision making and bad engineering. Not even their marketing department can save that.
 


That wasn't what Charlie was implying, he was implying that they literally have no idea how to build a chip or respin one, which just isn't true.

It isn't that nVidia didn't know about any of this or how to fix it, it is that they chose to try to take the easier (arguably lazier) and riskier route, which proved to be a bad decision. That isn't lack of homework, that is bad leadership, there is a distinct difference.
 
If the vias are that screwed, and nVidia mentioned this awhiles back, complaining about TSMC, and they need to double them up as well, itll be delayed, and no, they may still launch as Charlie says, but with very few parts, like the 7800GTX 512, only worse.
If they need to remake it, its a rock and a hard place here, as doubling the vias, and doing whatever they need to do to allow for the trannys to "even out", well, it cost ATI almost 15% die space, which was reducing features etc on 5870.
So, either add 15% tp die area, or remove even more features/shaders etc from Fermi.
If hes right, its 28nm or bust for Fermi, with low low numbers for the 40nm version
 


I could make a case for both.

Well it is certainly the final call of a manager, no one in a position to make a call like that would choose to ignore advice of scientists. There must have been a t least a couple engineers that were pretty high up saying all will be well, if this ends up being true.

Whether they chose to ignore facts due to ego, or just made the wrong call, we won't know unless they get fired.

This is very interesting.. true or not.. We will find out soon enough if this has any weight.

We keep hearing Fermi 2 info.. is that to Fermi what the 285 was to the 280? Or is Fermi 2 the code name for an entirely new architecture? Id imagine either way it will be on 40nm though.. I can't see them wanting to push out a new GPU on 28nm after all this.. Don't suppose we will see 28nm cards until a shrink of the 6000 series. (Right now I figure 6000s at the end of 2010 on 40nm, with a shrink for the refresh.. but if 6000s don't hit until Q1 11 we might see a shrink of the 5000s first.)
 
Remember, the 40nm was to be here as early as qtr4 08, with risk wafers anyways, and nothing til the 4770 which was April, so 6+ months right there. TSMC did the pooch here, and yes and no, nVidia should have been paying more attention in what ATI was doing, and no, TSMC is mainly to blame here, as Im sure they told them itd be fixed.

Maybe it is fixed, and this will work itself out, no thanks to TSMC, but even so, itll be delayed for numbers.
So, being as late as a year on process, the R&D doesnt stop at nVidia or ATI, and their plans for their next gen/iteration is ready, regardless of process, so time will tell, R900 vs Fermi2 anyone?
 
Well, all those posts about nVidia huge die costs are true, and is why they never lowered their prices, and now, theyre running low on higher performing cards, and cant compete using such huge dies on old processes in price or perf.
They needed to quit renaming and saying thats good enough, should have embraced DX10.1, made more cards, bigger better cards, instead of those puny OEM DX10.1 cards.
Theyve essentually been sitting on Fermi, and now, it appears theres just more and more setbacks, some of which we know, and some of which I hope are just rumor.
Oh, and MM, you could just go out and buy a cheaper better performing ATI card, its what I did when I bought my old 320GTS
 
well, we will see this in about 6 weeks or so when Fermi is launched or delayed, though i hope for the best (competition is always good)

i am also staying away from Gen-1 DX11 cards as well, got a little screwed by my 8800GTS640 (price wise), luckily i already have a decent card (4870) and that works fine for now
 
http://forum.beyond3d.com/showpost.php?p=1395767&postcount=1351
TSMC has had via issues before. The first case I know of is back at 130nm or 90nm.

And while you can go ahead and ignore some design rules, you need to be very very very careful. Everyone I've ever talked to about this was very clear - TSMC has the loosest design rules out there, if you ignore them, do so at your own peril. And pay close attention to their recommendations, or you're likely to be sorry.

David
__________________
www.realworldtech.com

Could be a possibility, as a few things fit too well
 

As I'm running an SLi rig I'll stick to Nvidia cards (and drivers [:mousemonkey] ) thanks, and as for DX10.1 by the time that was announced the G80 GPU's were a done deal, so unless it's a software not hardware tweak Nvidia would have had to scrap a shedload of stock whereas because ATi were late they were able to implement it IIRC.
 
You either go tall or go small. They decided to go this way, and gambled and are eating it now, whereas if theyd eaten it then and gambled, theyd have cards in the mix, instead of nada, which sounds better to you?
 
Theyd have spent more time doing the little things, which may have weeded out a few larger things, like ATI did.
Moving tp DX10.1 would have enabled them to also allow for a larger DX10.1 40nm ?test? chip, which they never did, and cancelled. If theyd already had DX10.1 at 55nm, then itd been just a cleaner easier path to 40nm testing
 
It'll be somewhere in between. Whatever version of Fermi the press gets will be fast enough to beat the 5870, and Nvidia will take a few $hundred million loss on this series while putting everything into it at 28nm.

There will probably be more than 10,000 also, but not a lot more I'd say. It hardly matters now, this series was lost ever since Cypress was released last September.
 
An example is, how many 210s or 220s are going to overheat? Whereas a larger/harder pushed chip will. If your voltage/amps are low on a low end card, the vias also may not be compromised as well.
Someone earlier in another thread wondered what card would ATI use for 28nm. Alot said another 4770 type chip, or around 800 shaders, I said no way, itd be more like the 5870/50, so theyd have a better idea of a card/chip that size or larger would perform/react
 

Then again they could always pad out the numbers a bit, like ATi did with the 2 million 5 series cards party and cake giving the impression that that it was all just retail desktop parts responsible for that figure.
 
Weird I was thinking about that exact thing earlier JDJ.

The 4770 was pretty much midrange, that helped ATI see issues scaling up and downward. The ultra low-end gt200's Nvidia released could not have shown them the potential issues with such a huge chip.

Imo, they were mostly hoping TSMC would fix the issues plaguing their 40nm and to a large degree they have, but Fermi is so huge that any weaknesses of a process will be amplified.
 

As I recall there was an article a short while ago where a bod from NV was giving it large on the PR front about how they needed a near zero defect process and were happy with TSMC's ability to deliver that, so I think they've taken that into account but only time will tell.
 

Were they all desktop parts then? are you sure about that?
 

Update February 11, 2010 23:01 UTC - Following the publication of this story, we got contacted by AMD and Bite PR, wanting to clarify the comments we gave out in the story. We spoke with Chris Hook, Director of PR at AMD, who told gave us the statement in regards to availability of Radeon HD 5000 series, which we all remember - was mostly unavailable after the launch.

"The 2 million DX11 GPUs shipped announcement back in January included mobile and desktop. Remember we announced the entire mobile family in early Jan, so chips would have had to ship to customers before then. So look beyond Cypress, Juniper and Hemlock when you’re looking at the numbers. We promise that they’re accurate.
BTW, no issues getting 5870 boards today – just check online!"
Click to expand...

Source

No mate it wasn't just desktop parts, so say the man from AMD.
 

I never said anything about IGP's, it's just that without a definite breakdown of the numbers the impression given by the cake in the shape of a Radeon card with the claim of 2 million sold written underneath was that they had sold two million graphic cards no mention of mobile chips, so a little misleading don't you think?
 
Status
Not open for further replies.

TRENDING THREADS

Latest posts

Moderators online

Share this page

COMPANY
RESOURCES
FOLLOW
Top Bottom