[SemiAccurate] Nvidia's Fermi GTX480 is broken and unfixable

[IntelUser2000]

You know what guys? Fermi, even with the delays might exceed performance of the then-ATI card to become the fastest single-GPU card, but one can't help but think what the side effect of the delay might have been.

Who knows if shaders are the only things cut? What if they cut something else like image quality and stability? Doing those right actually takes time, and while it might be ok on a not-delayed part, as your team gets desperate trying to get it out on time with pressure from management and marketing, they might be forced to cut and hope it works out well.

All the pressures caused by the delay can't be good for the mentality of the team as well. Maybe losing focus due to that might cause engineers to let bugs slip aside on accident.

Other than that I think its great Nvidia went all out and created something that's not a derivative. HD 5xxx seems less of a innovation and more of a heavily modified 4xxx. Maybe on the next gen they'll learn from their mistakes and create a great card.

 

[cbn]

I am very interested in Fermi, but I would also like to know what other techs Nvidia has up their sleeve that help regular folks justify expensive video card purchases.

Gaming on these cards is obviously the primary purpose, but maybe Nvidia needs to go beyond that as well?

 

[gorobei]

even if charlie's take is a worst case scenario and the yields aren't a financial disaster, the whole fermi gen is the very definition of pyrrhic victory.
The best analogy would be if the intel core2 dev team and the P4 team were two separate companies. One side going headlong into the megaHz wall regardless of power, heat, money issues; the other choosing to try to get more efficiency out of a more rational resource budget. Going big is rapidly becoming not a viable option.

 

[disappoint]

Pardon my ignorance as I'm just a consumer and not an industry insider but if TSMC is having so much trouble with the 40nm process why doesn't Nvidia go to Intel to fab the chips? Surely Intel's fab has better yields? They can even do 32nm can't they?

 

[Lonbjerg]

Because so far he was mostly correct about Fermi so far.

Bollocks.

I can only repeat what i have said before:
http://forums.anandtech.com/showpost.php?p=29122105&postcount=11

"Even a blind hen can find corn in time.

Neither of Char-lie's (or your guesswork) was based on fact(s)?
Why can I say this?
Because no one knew how A2 or A3 (or even if there where to be an A3 version) silicon would behave before they where out."


I find it pretty saying that LIE is part of his name...

 

[ViRGE]

Pardon my ignorance as I'm just a consumer and not an industry insider but if TSMC is having so much trouble with the 40nm process why doesn't Nvidia go to Intel to fab the chips? Surely Intel's fab has better yields? They can even do 32nm can't they?

1) Intel won't fab other people's chips. You need an independent foundry like TSMC to do it. TSMC is basically #1 for this thing right now (Global Foundries is still getting in to position).

2) Even if Intel would do it, you can't simply take a design to another fab. You basically have to rework your design to move it to another fab or another process, which takes several months to do.

 

[Kuzi]

Is it just me or were there some comments from Nvidia recently about the Fermi refresh still being slated for 2010 despite how late Fermi is? If that's in fact true, then Nvidia might very well end up doing exactly what Charlie mentions at the end of his article: Soft-launch Fermi 1 just to say they got it out the door while plugging away behind the scenes to get the "refresh" out later this year.

The thing is no FAB at this time has a 28nm process ready for production, TSMC only recently improved yields on their 40nm process, and I don't believe they can produce any 28nm chips till sometime in 2011.

My guess is both NV and ATI will use the 28nm process of GlobalFoundries, but even their 28nm process is not ready yet. GF said they will start 28nm production in H2 this year, though that doesn't mean we will see Fermi v2 or Northern Islands in 2010. The process has to be tested first, so maybe we'll see low end cards from NV/ATI come out in Q4 this year, then the high end cards in Q1 2011 at the earliest.

 

[Lonyo]

Pardon my ignorance as I'm just a consumer and not an industry insider but if TSMC is having so much trouble with the 40nm process why doesn't Nvidia go to Intel to fab the chips? Surely Intel's fab has better yields? They can even do 32nm can't they?

Intel manufactures Intel. TSMC manufactures whoever pays them.

 

[Mr. Pedantic]

You make a good point about Crossfire/SLI scaling, but lots of experienced people say they shy away from these solutions because of "microstutter".

Microstuttering doesn't happen so much anymore. The last card I've seen people have major problems with is the 9800GX2. CF/SLI both still have problems with GPU-GPU communication, but they're a lot more tolerable and a lot more manageable than in the GX2 days.

If you read the Anandtech RV870 article it explains why this happened.

"Sunspot" the internal name for Eyefinity was kept a complete secret up until a few weeks before release. The software guys at ATI didn't know about it till the last minute.

I did read that. That doesn't mean it's any less AMD's fault. If anything, all it does is focus the blame on one person for being so paranoid.

they will answer them because they are legitimate reporters. Charlie got blacklisted because he is a crazy vindicative hack. Why should nvidia have to deal with such a person? you get a restraining order. Just because the nutcase can come up with a valid question once a blue moon doesn't mean you have to talk to him. If it is such a good question someone more cordial will ask it and you answer that person.

Nvidia employees are people, and they shouldn't have to deal with jerks like charlie.

Well, I want those questions answered because they're legitimate questions, and I don't really give a fuck who asks.

You know what guys? Fermi, even with the delays might exceed performance of the then-ATI card to become the fastest single-GPU card, but one can't help but think what the side effect of the delay might have been.

Who knows if shaders are the only things cut? What if they cut something else like image quality and stability? Doing those right actually takes time, and while it might be ok on a not-delayed part, as your team gets desperate trying to get it out on time with pressure from management and marketing, they might be forced to cut and hope it works out well.

Fermi may very well perform better than Cypress. But in doing so it may very well also run disproportionately hotter and be more power hungry doing so, and will definitely be more costly to produce, both because of larger die size and lower yields. So either price the cards high and squeeze people out of NVidia-ism, or lower the price and lose lots of money. In either case, lower-end derivatives won't be seen for a long time yet and NVidia will keep losing money while doing so, so I think they should have (as many others undoubtedly think) kept GT200b, worked more on GT212, and not EOL'd both for some mystical reason.

And delays don't occur because of sudden pressure from management and/or marketing. It comes about because of bad planning and/or lack of foresight. I don't think the NVidia engineering department are absolved of blame; if anything, if this was forced upon them, they should have spoken up and announced that it could not be done (yet).

 

[Kuzi]

You know what guys? Fermi, even with the delays might exceed performance of the then-ATI card to become the fastest single-GPU card, but one can't help but think what the side effect of the delay might have been.

If the rumors are true about the initial Fermi release having cut back SP's and lower clocks, then I'd say there is no chance it can be faster than a 5970 for gaming.

Other than that I think its great Nvidia went all out and created something that's not a derivative. HD 5xxx seems less of a innovation and more of a heavily modified 4xxx. Maybe on the next gen they'll learn from their mistakes and create a great card.

Because of Fermi's large die size and the long delay, NV will want to release Fermi v2 as soon as possible, so it will most likely be just a shrink to 28nm. Similar to how Nvidia went from the GTX280 to the GTX285. Northern Islands on the other hand might be a new architecture with bigger improvements in GPGPU capability when compared to Cypress.

 

[blanketyblank]

What I never really understood is why AMD/ATI didn't use global foundry more than TSMC considering it owned it and I think still owns a large chunk of it(I also think it's losing money).

 

[Lonbjerg]

NVIDIA kinda replies...via employes on twitter:
http://twitter.com/Igor_Stanek

RT @RS_Borsti: Oh Charlie... that just another hilarious post me: I think with this post Charlie totally destroyed his credibility

@RS_Borsti I want to see how he is going to explain his article in March.... looks like biggest lie and mistake of his life

RT @madciapka: @Igor_Stanek there was any sign of credibility before that? I don't think so :me right, you are right..

 

[PingviN]

NVIDIA kinda replies...via employes on twitter:
http://twitter.com/Igor_Stanek

RT @RS_Borsti: Oh Charlie... that just another hilarious post me: I think with this post Charlie totally destroyed his credibility

@RS_Borsti I want to see how he is going to explain his article in March.... looks like biggest lie and mistake of his life

RT @madciapka: @Igor_Stanek there was any sign of credibility before that? I don't think so :me right, you are right..

I don't know what you think, but I'd say it would be odd for Nvidia employees to go on Twitter and say they fucked up just as badly as Charlie claims. Don't you?

I don't believe Charlie's doomsaying, but taking an Nvidia employees word for it is equally messed up.

We still have to wait for the end of March to know whether he got this one right or not.

 

[Borealis7]

so what gets released first? Fermi or Larrabee?

for me, this product family from NV is not very interesting as the 58xx family allows maxing most games and certainly the ones i play.
i'm not in the market for a card right now, but if i were, i would spring for a Radeon.
actually, i'm GLAD i'm not looking to upgrade now. because later either prices will drop or new hardware with more features will be released and DX11 performance will improve as the technology matures.

 

[HurleyBird]

My guess is both NV and ATI will use the 28nm process of GlobalFoundries, but even their 28nm process is not ready yet. GF said they will start 28nm production in H2 this year, though that doesn't mean we will see Fermi v2 or Northern Islands in 2010. The process has to be tested first, so maybe we'll see low end cards from NV/ATI come out in Q4 this year, then the high end cards in Q1 2011 at the earliest.

You're most likely wrong on both counts there. GF 28nm process (at least, the performance process -- the unsuitable 'super low power' process may be) won't be online soon enough for the new video cards Dirk Meyer said would be launching in 2H 2010. Northern Islands is almost definitely going to use GF's 32nm SOI process, which is already up and running in a limited fashion. Only makes sense given the timetable and the fact that the Llano APU is made on 32nm SOI so ATI already has experience on that process. Nvidia, on the other hand, will probably stay with TSMC at least a little longer, firstly because of the connection between AMD and GF, secondly because they are worse at migrating to new processes and would not have the AMD CPU team's help like ATI did.

 

[Will Robinson]

Interesting analysis by Charlie there....wonder if Fermi's reviews will prove him wrong.

 

[Lonyo]

What I never really understood is why AMD/ATI didn't use global foundry more than TSMC considering it owned it and I think still owns a large chunk of it(I also think it's losing money).

Many reasons, but probably mainly process type and capacity issues I would guess, both of which GF due to ATIC are on the way to solving with the acquisition of Chartered and the investment of lots of money, but which only now are solving the "problem".

 

[Kuzi]

You're most likely wrong on both counts there. GF 28nm process (at least, the performance process -- the unsuitable 'super low power' process may be) won't be online soon enough for the new video cards Dirk Meyer said would be launching in 2H 2010. Northern Islands is almost definitely going to use GF's 32nm SOI process, which is already up and running in a limited fashion. Only makes sense given the timetable and the fact that the Llano APU is made on 32nm SOI so ATI already has experience on that process. Nvidia, on the other hand, will probably stay with TSMC at least a little longer, firstly because of the connection between AMD and GF, secondly because they are worse at migrating to new processes and would not have the AMD CPU team's help like ATI did.

That's also a possibility, that Northern Islands would use a 32nm process, although SOI would be more expensive, so they might use 32nm Bulk which GF has been working on for a while too.

As for Nvidia staying with TSMC, it could prove the wrong decision, because TSMC's 32nm process does not have HKMG, but their 28nm process will have it. While GF uses HKMG for both their 32nm and 28nm processes. So ATI would have an advantage by going with GF, unless NV uses TSMC's 28nm HKMG process.

I could be wrong but I believe GF are ahead with their 32nm/28nm HKMG processes, so that "could" also mean more delays for NV if they go with TSMC.

 

[IntelUser2000]

*snip*

And delays don't occur because of sudden pressure from management and/or marketing. It comes about because of bad planning and/or lack of foresight. I don't think the NVidia engineering department are absolved of blame; if anything, if this was forced upon them, they should have spoken up and announced that it could not be done (yet).

I'm saying marketing and management pressure because it was later than they expected. That pressure probably won't really help with the product's health.

 

[Jacky60]

Why if Fermi is so kick ass and releasing so soon have Nvidia not given performance figures yet. They are losing a high volume of potential sales to ATI and performance figures even without pricing would persuade a significant percentage of prospective buyers to hold and wait for Fermi (if it really can compete).
Charlie may not be much liked by the Nvidia fanboy community but he seems to be getting pretty close to the truth withnhis Fermi reporting. If he's right I doubt anyone here will give him the credit he will deserve. If he's wrong and Fermi really smokes the opposition I will be please but only to get lower prices-personally my money is on him being right.

 

[Lonyo]

Why if Fermi is so kick ass and releasing so soon have Nvidia not given performance figures yet. They are losing a high volume of potential sales to ATI and performance figures even without pricing would persuade a significant percentage of prospective buyers to hold and wait for Fermi (if it really can compete).
Charlie may not be much liked by the Nvidia fanboy community but he seems to be getting pretty close to the truth withnhis Fermi reporting. If he's right I doubt anyone here will give him the credit he will deserve. If he's wrong and Fermi really smokes the opposition I will be please but only to get lower prices-personally my money is on him being right.

Well AMD remained tight lipped when it came to the HD48xx launch, and that pretty much slapped a lot of people hard in the face because they weren't expecting it.
Keeping quiet isn't always done just because there are problems, there have been times when people have kept quiet about performance and then just come out of left field with something great.

 

[ayabe]

Sounds really bad, even if only half of that is accurate. Glad I stopped waiting.

 

[Juncar]

Just wait and see in a month.

 

[br0wn]

Just wait and see in a month.

We have waited in October 2009 and then in December 2009 when NVIDIA said they will release Fermi then. We are losing confidence with NVIDIA's execution.

 

[Stoneburner]

Interesting analysis by Charlie there....wonder if Fermi's reviews will prove him wrong.

I don't remember him ever questioning the performance (potential?) of Fermi, rather its viability as a marketable product.