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

[tviceman]

And why did you not address the tens of other claims he made and I just listed?

I think we need to take a step back and look at the whole picture. Charlie has probably been right about some of what he's said, but he's always talking worst case scenario and leaves nothing else good to say about Nvidia in general.

Out of the top 12 articles, which were more about the process of creating Fermi than the architecture of Fermi itself, I addressed 7 as being sensationalist, misleading, biased, and possibly based on opinion or false information. 7/12 articles that are possibly/likely completely spun and/or false. That is terribly bad/inaccurate reporting/journalism - even for an opinionated journalist such as him.

Fudzilla and BSON have both reported the 2% yield as wrong and 9% yields for A3 sound ridiculously low too. To illustrate how off a current 9% yield likely is, right now 94 gf100's can fit on a single wafer. A wafer costs them around $5000. 9% of 94 is 8.5. So if nvidia is averaging 8 1/2 chips per wafer @ $5,000, each chip is costing them $588 right now. Add in PCB, ram, heatsink, fan, manufacturing, and packaging, then you are looking at a $750 price just to break even on a consumer version of GF100.

GF100 may not be profitable in the consumer market in it's current iteration, but to be losing $250-400 per sale is not something Nvidia will do - limited supply or not. Nvidia is not a charity.

 

[Tempered81]

IMO, those quotes referencing "CPU" are initial developments what you now see in the compute centric design of GF100. I wasn't quoting them suggesting a typical x86 "AMD/Intel" type CPU, since legally Nvidia cannot manufacture that. After all, Jensen did say Fermi was 5-6 years in the making.

 

[bryanW1995]

hey guys im new here.so i read this article about nvidia's new fermi architecture falling on its face.some ree-ree sitting in the back of a small yellow bus wrote it,his name was charlie.look

guys,this article is crap,everything charlie writes/says...is crap.obviously weve all seen the demos of fermi spanking some ass @ CES 2010.anyone anticipating this launch should not be scared

away by the likes of charlie.i watched the farcry 2 demo ---> here http://www.youtube.com/watch?v=yNbY-m28bUk where "Fermi" the gtx 480 got 84ish fps with everything maxed out on a 24"

resolution.i watched the HD5970 benchmark ---> here http://www.youtube.com/watch?v=K8gjrgMOEqQ almost the same exact specs except fermi only had 4x AA and the 5970 had 8x.big

whoopdy wow wow.the 5970 got 80 ish fps.if you pay any kind of attention youll see the fermi had higher minimum fps. fermi had 64fps min and 5970 had 52 fps min.cmon guys,not to use cliches

but....the proof is in da puddin.yes,im an nvidia fanboy.i love what ATi is putting on the market right now A+ to them.to each their own.before anyone posts something stupid like im a "Nub" or

that "I dont know anything about computer hardware"...i built a 20,000 3dmark06 computer with a pair of 8800GT's and an E8500.thanks.



this isnt my best score just some proof http://service.futuremark.com/resultComparison.action?compareResultId=11226543&compareResultType=14

I'm surprised nobody spotted the obvious: if you compare 4xAA vs 8xAA then it is highly reasonable to assume that min fps will be slightly lower and avg fps will be slightly lower on identical cards. In fact, if you were to compare both of those hypothetical cards at 4xAA and 8xAA, I wouldn't be surprised to see the fermi win at 4x and the 5870 win at 8x. That result seems highly suspect to me, as even a shitty/hot/late fermi should at least be 10-15% faster than 5870.

If 5870 ends up being faster in 8xAA than gtx 480 how much will amd raise their gpu prices??? It's sad when even fanbois are predicting abject failure for your company.

I would hate to be an nvidia engineer right now.

 

[Soleron]

Fudzilla and BSON have both reported the 2% yield as wrong and 9% yields for A3 sound ridiculously low too. To illustrate how off a current 9% yield likely is, right now 94 gf100's can fit on a single wafer. A wafer costs them around $5000. 9% of 94 is 8.5. So if nvidia is averaging 8 1/2 chips per wafer @ $5,000, each chip is costing them $588 right now. Add in PCB, ram, heatsink, fan, manufacturing, and packaging, then you are looking at a $750 price just to break even on a consumer version of GF100.

GF100 may not be profitable in the consumer market in it's current iteration, but to be losing $250-400 per sale is not something Nvidia will do - limited supply or not. Nvidia is not a charity.

I see the problem. You can't believe yields could be that low. That's fine - it is implausible that an advanced engineering company could screw up that badly. However there are good reasons why low yields could possibly happen in this case, due to the size of the chip, the 40nm issues and how Nvidia hasn't dealt with them like AMD has.

But suppose it is true. What would we see? Nvidia couldn't cancel the product, since that would be very bad PR. We would be likely to see an initial small batch, then a tiny trickle of parts while they work on a major, 6-month respin to fix that. We would also see them release no or few derivatives as possible and try and get by on their current products.

So if we see that, and Fermi never really becomes available during its lifetime, would you accept that yields could be that low?

I would be prepared to accept the converse; that yields aren't that bad if we do see availability and timely derivatives.

--

I happen not to trust Fudzilla and BSN, as they have contradicted themselves numerous times (BSN puts the Fermi tapeout last March [impossible.] and has "leaked" specs for the "GTX380" and "GTX360" from a few months back which are far too optimsitic (1700MHz shaders, 512 shader count, 870 GFLOPS which is higher than Nvidia's public Tesla specs. Fudzilla has stated, with certainty, about 8-10 different release dates for Fermi so far.) But that's irrelevant to the above, just tell me whether you'd ever, in theory, accept a 9% claim if the evidence was there.

I do agree Nvidia is not a charity, so if I see parts in volume I'll know Charlie's wrong.

--

Charlie has probably been right about some of what he's said, but he's always talking worst case scenario and leaves nothing else good to say about Nvidia in general.

Again, what if it's true? What if it IS that bad? Would you expect him to be optimistic if he knew it was that bad? If it does come out and has a 280W power draw and underperforms etc. etc. then his comments wouldn't look sensational at all. You're just judging it based on some assumption that it can't be that bad. It might be. (as I say, we'll find out soon).

 

[Piotrsama]

Fudzilla and BSON have both reported the 2% yield as wrong and 9% yields for A3 sound ridiculously low too. To illustrate how off a current 9% yield likely is, right now 94 gf100's can fit on a single wafer. A wafer costs them around $5000. 9% of 94 is 8.5.

You have to consider how many of the "failed" dies will they be able to harvest.

I'm surprised nobody spotted the obvious: if you compare 4xAA vs 8xAA then it is highly reasonable to assume that min fps will be slightly lower and avg fps will be slightly lower on identical cards. In fact, if you were to compare both of those hypothetical cards at 4xAA and 8xAA, I wouldn't be surprised to see the fermi win at 4x and the 5870 win at 8x. That result seems highly suspect to me, as even a shitty/hot/late fermi should at least be 10-15% faster than 5870.

If 5870 ends up being faster in 8xAA than gtx 480 how much will amd raise their gpu prices??? It's sad when even fanbois are predicting abject failure for your company.

Why not wait for an apples to apples review?

 

[extra]

You have to consider how many of the "failed" dies will they be able to harvest.

Yep. Regardless, the 9% or whatever number seems absurdly low, I really doubt that. I bet the first yields are somewhere in the 40-60% range, but who knows.

Still, does the approach nvidia is taking remind anyone of what happened to ati with the 2900?

The other question really is how well will nvidia be able to get a competitive low end part out of this? For non-fanbois, we need both companies to have competitive offerings so that prices are good--we don't want 2008-2009 prices to be an anomaly and go back to having to fork out $400 to be able to game well.

 

[tviceman]

I see the problem. You can't believe yields could be that low. That's fine - it is implausible that an advanced engineering company could screw up that badly. However there are good reasons why low yields could possibly happen in this case, due to the size of the chip, the 40nm issues and how Nvidia hasn't dealt with them like AMD has.

But suppose it is true. What would we see? Nvidia couldn't cancel the product, since that would be very bad PR. We would be likely to see an initial small batch, then a tiny trickle of parts while they work on a major, 6-month respin to fix that. We would also see them release no or few derivatives as possible and try and get by on their current products.

So if we see that, and Fermi never really becomes available during its lifetime, would you accept that yields could be that low?

I would be prepared to accept the converse; that yields aren't that bad if we do see availability and timely derivatives.

--

I happen not to trust Fudzilla and BSN, as they have contradicted themselves numerous times (BSN puts the Fermi tapeout last March [impossible.] and has "leaked" specs for the "GTX380" and "GTX360" from a few months back which are far too optimsitic (1700MHz shaders, 512 shader count, 870 GFLOPS which is higher than Nvidia's public Tesla specs. Fudzilla has stated, with certainty, about 8-10 different release dates for Fermi so far.) But that's irrelevant to the above, just tell me whether you'd ever, in theory, accept a 9% claim if the evidence was there.

I do agree Nvidia is not a charity, so if I see parts in volume I'll know Charlie's wrong.

--



Again, what if it's true? What if it IS that bad? Would you expect him to be optimistic if he knew it was that bad? If it does come out and has a 280W power draw and underperforms etc. etc. then his comments wouldn't look sensational at all. You're just judging it based on some assumption that it can't be that bad. It might be. (as I say, we'll find out soon).

Yeah, I can definitely admit that if the card is never in stock and there are no lower end consumer derivatives by the end of summer then it's yields are worse than awful and Nvidia only released a few thousand to save face. So, beyond that, we're just arguing what if scenarios. Likely, it's somewhere in between the optimistic reports and unoptimistic reports.

 

[Soleron]

Yeah, I can definitely admit that if the card is never in stock and there are no lower end consumer derivatives by the end of summer then it's yields are worse than awful and Nvidia only released a few thousand to save face. So, beyond that, we're just arguing what if scenarios. Likely, it's somewhere in between the optimistic reports and unoptimistic reports.

OK, agreed.

 

[tviceman]

You have to consider how many of the "failed" dies will they be able to harvest.

That is true; I was just under the assumption that the 2% or 9% accounted for all functional chips.

 

[Dark4ng3l]

The way I see it the 2% is probably compared what they were expecting to ship. ie the full 700/1400 MHz 512 SP parts. When you lower it to the rumored specs of 448 SP and 600/1200 MHz for the high end part and probably lower for their lower end part they can probably get a reasonable amount of usable cores (like 30%+) Now the performance of the downclocked/ lower SP parts they would have to release could be pretty bad VS what they expected to be able to ship.

 

[blanketyblank]

So if he's right what we'd see is Nvidia would release very few high end 512 SP parts maybe just a few thousand. They would claim victory at having the fastes single gpu since this part is faster than the 5870. However the only part we'd actually see on the market is a much weaker part with 448 SP or less and performance maybe similar to or slower than the 5870.

This actually sounds about right, and would be very difficult to verify since even if they were released in mass they'd likely run out early. Just looking at the 5970s we can see there isn't much stock compared to demand. We'd probably have to wait for sales figures or at least have some retailers willing to tell us about how much stock they are getting.
Otherwise all we'd know is they aren't available anywhere for anything close to MSRP.

 

[Rebel44]

Fermi performance (info from S|A)
http://www.semiaccurate.com/2010/02/20/semiaccurate-gets-some-gtx480-scores/

.......
The GTX480 with 512 shaders running at full speed, 600 or 625MHz depending on which source, ran on average 5% faster than a Cypress/HD5870, plus or minus a little bit. The sources were not allowed to test a GTX470 which is likely an admission that it will be slower than Cypress/5870.
.......

 

[tommo123]

is that true what he says in the article about cutting anand etc out of reviewing nvidias hardware in the past?

 

[zerocool84]

is that true what he says in the article about cutting anand etc out of reviewing nvidias hardware in the past?

Yes which is sad.

http://www.theinquirer.net/inquirer/news/1051123/nvidia-cuts-reviewers-gts250

 

[blanketyblank]

Fermi performance (info from S|A)
http://www.semiaccurate.com/2010/02/20/semiaccurate-gets-some-gtx480-scores/

guess we'll see. My guess is if heat is really the big problem Charlie claims we'll see GTX 480 Ultras with water cooling that run at 700+ so nvidia can sell them over a 5870 and have larger margins. Also assuming the sources are accurate sounds like there are a lot of driver problems still so there is the possibility the card can improve dramatically once those are ironed out.

 

[HurleyBird]

Also assuming the sources are accurate sounds like there are a lot of driver problems still so there is the possibility the card can improve dramatically once those are ironed out.

That's not a very safe assumption, at least regarding the supposed tessellation bugs. Any fix for artifacting or stability is most likely to A. employ costly workarounds or B. decrease the workload on the GPU, decreasing performance, not increasing it. But yeah, increase in performance is also possible, and gets more and more likely as they have more time to optimize for the architecture.

 

[blanketyblank]

That's not a very safe assumption, at least regarding the supposed tessellation bugs. Any fix for artifacting or stability is most likely to A. employ costly workarounds or B. decrease the workload on the GPU, decreasing performance, not increasing it. But yeah, increase in performance is also possible, and gets more and more likely as they have more time to optimize for the architecture.

I agree. I'm not saying it's likely or easy, but I do think nvidia can work out at least marginal performance increases for the card with driver optimizations just like how ATI is getting like 5%+ boosts in select games with driver releases in the 5xxx series. As for tesselation since they are getting double frame rates (since Charlie is being positive about it and we've seen CES demo, it has to be true) I think they have some extra performance they can sacrifice to get it working.

 

[Piotrsama]

That's not a very safe assumption, at least regarding the supposed tessellation bugs. Any fix for artifacting or stability is most likely to A. employ costly workarounds or B. decrease the workload on the GPU, decreasing performance, not increasing it. But yeah, increase in performance is also possible, and gets more and more likely as they have more time to optimize for the architecture.

Interesting post, maybe we can draw a parallel with the TLB bug on the first Phenoms?

 

[Tempered81]

I'm guessin it's gonna be 20-30% faster than 5870. I'll wait til Anand's review (2 months or so) hahah.

 

[HurleyBird]

As for tesselation since they are getting double frame rates (since Charlie is being positive about it and we've seen CES demo, it has to be true) I think they have some extra performance they can sacrifice to get it working.

Yeah, no doubt. The big question is, does using tessellation on Fermi reduce shader performance as Charlie states, and if so, by how much?

Next interesting question would be: Is Fermi's tessellation prowess due to a fundamentally superior design, or is it due to tessellation being heavily cut in RV870 so it could fit into a smaller die size?

Third, an observation: Tessellation is unlikely to be heavily used for at least another year. It could be an ironic twist that even though the two are reportedly equal in games now, but in two or three years time Fermi could be at a significant advantage. Since people mainly look at current games to make purchasing decisions ATI won't lose out, and will have Northern Islands launched by the time tessellation becomes a factor. Seems like 1800/1900xtx vs. 7800/7900GTX with the companies reversed, no?

 

[Lonyo]

Yes which is sad.

http://www.theinquirer.net/inquirer/news/1051123/nvidia-cuts-reviewers-gts250

Early last week Charlie over at The Inquirer posted a story saying that a number of reviewers were cut out of the GeForce GTS 250 launch. We felt a bit hurt, by the time the story launched we weren't even asked to be briefed about the GTS 250. Cards had already gone out to other reviewers but we weren't on any lists. Oh, pout.

Magically, a couple of days after Charlie's article we got invited to a NVIDIA briefing and we had a GTS 250 to test. Perhaps NVIDIA was simply uncharacteristically late in briefing us about a new GPU launch. Perhaps NVIDIA was afraid we'd point out that it was nothing more than a 9800 GTX+ that ran a little cooler. Or perhaps we haven't been positive enough about CUDA and PhysX and NVIDIA was trying to punish us.

Who knows what went on at NVIDIA prior to the launch, we're here to review the card, but for what it's worth - thank you Charlie

http://www.anandtech.com/video/showdoc.aspx?i=3523&p=2

 

[zerocool84]

http://www.anandtech.com/video/showdoc.aspx?i=3523&p=2

A timed restriction is just as bad.

 

[Lonyo]

Yeah, no doubt. The big question is, does using tessellation on Fermi reduce shader performance as Charlie states, and if so, by how much?

Next interesting question would be: Is Fermi's tessellation prowess due to a fundamentally superior design, or is it due to tessellation being heavily cut in RV870 so it could fit into a smaller die size?

Third, an observation: Tessellation is unlikely to be heavily used for at least another year. It could be an ironic twist that even though the two are reportedly equal in games now, but in two or three years time Fermi could be at a significant advantage. Since people mainly look at current games to make purchasing decisions ATI won't lose out, and will have Northern Islands launched by the time tessellation becomes a factor. Seems like 1800/1900xtx vs. 7800/7900GTX with the companies reversed, no?

http://www.anandtech.com/video/showdoc.aspx?i=3721&p=2

NVIDIA’s geometry hardware is no longer bound by any limits of the pipelined fixed-function design (such as bottlenecks in one stage of the pipeline), and for better or for worse, they can scale their geometry and raster abilities with the size of the chip. A smaller GF100 derivative will not have as many PolyMorph or Raster units as GF100, and as a result won’t have the same level of performance; G92 derivatives and AMD’s designs both maintain the same fixed function pipeline through all chips, always offering the same level of performance.

Essentially an HD5850 and an HD5870 should have the same tessellator performance.
A GTX480 will have higher tessellation performance than a GTX470.
It's a fundamental design choice which scales up for NV, and scales down for ATI.

 

[HurleyBird]

Essentially an HD5850 and an HD5870 should have the same tessellator performance.
A GTX480 will have higher tessellation performance than a GTX470.
It's a fundamental design choice which scales up for NV, and scales down for ATI.

Yeah, I've read that. I know the difference in design, and I know that at the high end, Nvidia's design is faster. My main points of curiosity are how Nvidia's tessellation performance scales with other shader usage, and if the tessellation performance in RV870 was one of the parts of the chip that got put on the chopping block when they went for a smaller die.

 

[waffleironhead]

For all those claiming that nV never had a release date specified, that may be true, but they did have a release window.

http://www.pcworld.com/article/184908/amds_lead_with_directx_11_is_insignificant_nvidia_says.html

"This 60-day lag between these events of when our competition has DX11 and when we're coming to market will absolutely seem insignificant in the big picture,"

unless "coming to market" means a paper launch.