Charlie explains the whole Fermi debacle in detail

[Dark4ng3l]

http://www.semiaccurate.com/2010/03/28/why-nvidia-hacked-gtx480/

He likes to reference himself allot and I haven't reread all of his articles that he links to to make sure that he is not twisting his own words to make himself look good. Though what he says sounds pretty much spot on as far as what I was thinking was the most probable cause for the heat, power and cut down reality of the Fermi cards.

I know we had a thread detailing if each of his articles were accurate or not post launch but I felt that this summed the entire saga up nicely and we can discuss whether he was actually right or he is trying to make himself look better than what he is.

 

[DefRef]

Charlie is such a sanctimonious jackhole that it's stomach-turning to try and slog thru his bile and gloating to learn anything. A legit site like Ars Technica could've told this story without the juvenile crap, but that's why they're legit and Charlie's just an ATI fanboy troll site.

 

[brybir]

http://www.semiaccurate.com/2010/03/28/why-nvidia-hacked-gtx480/

He likes to reference himself allot and I haven't reread all of his articles that he links to to make sure that he is not twisting his own words to make himself look good. Though what he says sounds pretty much spot on as far as what I was thinking was the most probable cause for the heat, power and cut down reality of the Fermi cards.

I know we had a thread detailing if each of his articles were accurate or not post launch but I felt that this summed the entire saga up nicely and we can discuss whether he was actually right or he is trying to make himself look better than what he is.

My only complaint with SA articles is that they are written as if everything nVidia does is a personal attack against Charlie. nVidia chip yield, binning and marketing decisions are a decision made collectively by the engineers, marketing and other types in the organization based on what can be done and what they believe is in the best interest of the company. These are the same things that AMD, Intel etc do every single day at various stages of product production. Charlie can take something like Fermi production and make its production timeline sound like a soap opera and it is just sensationalist when it is just what must occur in any chip producing or designing business.

Its not as if nVidia is alone in having duds based on production or design. One can go and look at many of Intel's problems scaling the netburst up to the clock speeds that it needed with its deep pipeline, or AMD's bumble on its first phenom generation or the 2XXX series of cards, or even nVidia itself back in the 5XXX days.

Bottom line: Designing and building complex IC's is really really hard work. It takes tons of planning and a single design decision that seems wise can end up killing your product. (go read anand's interview with the guy from ATI about how things can go wrong). To make it seem so darn dramatic is just a disservice to the incredibly hard work that goes into designing and making these chips that are planned for years ahead of time.

Judge a chip on its performance, its price and its abilities, but this dramatic overplay of corporate intrigue is just silly, and detracts, in my opinion, from the excitement that the advance in technology brings as we march into the future.

 

[Phil1977]

Love him or hate him, but he has been 100% correct about fermi. Perfromance, yield issues, heat, noise was all laid out quite well.

Nobody should have been surprised when the reviews did come out. I wasn't, I believed his articles 100%.

 

[phaxmohdem]

I've been following SemiAccurate throughout the whole Fermi ordeal. My personal opinion on the matter is that Charlie D. deserves a little more credit for his predictions and reporting than he got for this round.

People will always hate the man, and he does tend to self-congratulate and overstate things too much (His claim of Fermi being unmanufacturable for example. Inefficient/unprofitable to manufacture perhaps, but not completely and totally 100% unmakable.)

However, throughout the process the numbers he threw out seem to align pretty closely with reality in the end... and I agree with his conclusion that it is probably not worth it at this point to salvage Fermi 1 with a respin, but rather focus their engineering efforts on Fermi 2 instead.

From my own point of view however, it sucks that Fermi cannot be made profitably, as that means nVidia has no price leverage against ATI this round, and both companies will command high prices for their products... which means I won't be upgrading

 

[zagood]

Bottom line: Designing and building complex IC's is really really hard work. It takes tons of planning and a single design decision that seems wise can end up killing your product. (go read anand's interview with the guy from ATI about how things can go wrong). To make it seem so darn dramatic is just a disservice to the incredibly hard work that goes into designing and making these chips that are planned for years ahead of time.

Judge a chip on its performance, its price and its abilities, but this dramatic overplay of corporate intrigue is just silly, and detracts, in my opinion, from the excitement that the advance in technology brings as we march into the future.

Well said.

 

[Hauk]

I'd really like to know if this part, the part about insiders at nVidia is true..

"When Nvidia got back A3 just before Christmas 2009, it was described to SemiAccurate by insiders at Nvidia as, A mess."

 

[Phil1977]

I'd really like to know if this part, the part about insiders at nVidia is true..

"When Nvidia got back A3 just before Christmas 2009, it was described to SemiAccurate by insiders at Nvidia as, A mess."

Fermi is like Vista.

Internally everyone knows it sucks and they are quickly working on a replacement.

But to the public it's all smiles...

 

[ScorcherDarkly]

I've only gotten familiar with Charlie in the last 9 months or so, so I don't have any insight on his track record beyond his Fermi coverage. IMO, his information has been pretty accurate compared to what has transpired. There's no question he's biased, and I agree, it would be easier to believe him if he didn't always sound like he was ranting. But if you know he's biased, and you read more than one source, then the information can still be useful. Its no different than political news coverage. When you watch Fox News you know its slanted to the right. When you watch MSNBC you know its slanted to the left. Both sources can still be useful, if you filter intelligently before you swallow the information. Charlie isn't any different.

 

[Lonyo]

Fermi is like Vista.

Internally everyone knows it sucks and they are quickly working on a replacement.

But to the public it's all smiles...

And while it's not *actually* terrible in terms of function and features, it requires a lot of power to run (in the case of Vista it was hardware power not electrical power)

 

[Genx87]

Fermi is like Vista.

Internally everyone knows it sucks and they are quickly working on a replacement.

But to the public it's all smiles...

I'd love to know your idea of "sucks".

 

[Zebo]

LOLZ he says if you have not pre ordered one you will not get one cuz nV only has 10K or so... He will be wrong here. What's nV going to do? Just quit selling? Let AMD have the market?

 

[ScorcherDarkly]

LOLZ he says if you have not pre ordered one you will not get one cuz nV only has 10K or so... He will be wrong here. What's nV going to do? Just quit selling? Let AMD have the market?

10k 480s and "twice that" 470s. And really it looks like he was already wrong over his initial 5-8k prediction, though if one was to defend him they'd say he was only talking about the 480 and he was only off by a few thousand at worst.

 

[sandorski]

I wouldn't be surprised if sometime soon NVidia starts leaking him False Information for Vengeance sake.

 

[gorobei]

10k 480s and "twice that" 470s. And really it looks like he was already wrong over his initial 5-8k prediction, though if one was to defend him they'd say he was only talking about the 480 and he was only off by a few thousand at worst.

actually the point of the article was that he was right about the numbers when nv was shooting for the 512@1600 and 480@1200 targets. if they had stayed with a 512 part then the yield would only have been a few 1000s, but by lowering the spec/binning the yield increased. If fermi required 512shaders to run at 1200, the number of qualifying cores would be around his 2% estimate (too expensive to sell)

even I as a layman, when I saw the lower shader counts on the 480 and 470, knew that nv was re-binning and the total number of parts released would go up.

 

[blanketyblank]

LOLZ he says if you have not pre ordered one you will not get one cuz nV only has 10K or so... He will be wrong here. What's nV going to do? Just quit selling? Let AMD have the market?

Let's wait and see. I'm not sure why he would make such a statement this late in the game, but if he's actually right about this it would go a long way to his credibility. Of course the real question is when and how are these numbers verifiable.

 

[bryanW1995]

I'd love to know your idea of "sucks".

6 1/2 months late, enormous power draw, enough respins to send wreckage to counseling, after all that doesn't compete any better vs the current gen amd card than gtx 280 did (with the gtx 470 being significantly less competitive than gtx 260 was), still getting its ass kicked because amd's dual gpu card has been out for a LONG time already.

2900xt was the definition of "suck". fermi is not that bad, but it does look suspiciously like the modern day nvidia 5800 series. I want a new card but I'm not going to spend $300-$400 to get it. Hopefully nvidia will be able to run some timely respins and/or get the next gen out quickly because my gtx 260 isn't going to last forever.

 

[T2k]

Love him or hate him, but he has been 100% correct about fermi. Perfromance, yield issues, heat, noise was all laid out quite well.

Nobody should have been surprised when the reviews did come out. I wasn't, I believed his articles 100%.

I've been following SemiAccurate throughout the whole Fermi ordeal. My personal opinion on the matter is that Charlie D. deserves a little more credit for his predictions and reporting than he got for this round.

People will always hate the man, and he does tend to self-congratulate and overstate things too much (His claim of Fermi being unmanufacturable for example. Inefficient/unprofitable to manufacture perhaps, but not completely and totally 100% unmakable.)

However, throughout the process the numbers he threw out seem to align pretty closely with reality in the end... and I agree with his conclusion that it is probably not worth it at this point to salvage Fermi 1 with a respin, but rather focus their engineering efforts on Fermi 2 instead.

From my own point of view however, it sucks that Fermi cannot be made profitably, as that means nVidia has no price leverage against ATI this round, and both companies will command high prices for their products... which means I won't be upgrading

This... if you get past the sensationalist style - which is obviously deliberate, to be unique among plenty of other sites - you will recognize his reasonable logic and his pretty awesomely correct predictions as conclusions.

Unlike anybody else this guy has sources everywhere and not afraid to publish his info. He's raw and sometimes unpolished but certainly smart and he was pretty much 100% correct on Fermi.

Give the credit when its due...

I'd really like to know if this part, the part about insiders at nVidia is true..

"When Nvidia got back A3 just before Christmas 2009, it was described to SemiAccurate by insiders at Nvidia as, A mess."

You'll never know that - in fact he deletes every sensitive email after a day or two, to make sure no subpoena can open them.

 

[gorobei]

... Charlie can take something like Fermi production and make its production timeline sound like a soap opera and it is just sensationalist when it is just what must occur in any chip producing or designing business.....

....To make it seem so darn dramatic is just a disservice to the incredibly hard work that goes into designing and making these chips that are planned for years ahead of time.

while charlie's writing style goes into the hysterics range at times, part of the point of NV 'issues' is their operating style. NV has typically gone brute force in design(more transistors, faster speeds, bigger die). while this worked well enough for them before, they started to hit the limits of the node processes and power/thermal envelopes.

g200 marked the point where going big wasnt delivering the returns. gf100 had even bigger problems. If charlie is to be believed about the behind the scenes stuff, NV responded by screaming at it's suppliers and stonewalling its AIB partners.
This is relevant as it suggests that there may have been a few more disgruntled employees among NV's partners this time than there were before. More of them, and more likely to leak real info. Hence charlie's more accurate info.

 

[T2k]

while charlie's writing style goes into the hysterics range at times, part of the point of NV 'issues' is their operating style. NV has typically gone brute force in design(more transistors, faster speeds, bigger die). while this worked well enough for them before, they started to hit the limits of the node processes and power/thermal envelopes.

g200 marked the point where going big wasnt delivering the returns. gf100 had even bigger problems. If charlie is to be believed about the behind the scenes stuff, NV responded by screaming at it's suppliers and stonewalling its AIB partners.
This is relevant as it suggests that there may have been a few more disgruntled employees among NV's partners this time than there were before. More of them, and more likely to leak real info. Hence charlie's more accurate info.

Did you see this piece?

Amazing arrogance and stupidity...

 

[blanketyblank]

Did you see this piece?

Amazing arrogance and stupidity...

I don't really see what's wrong with that unless you have something to show it's untrue or true for that matter.

It'd be nice to have some verifiable info. Although one thing that stands out is if he says there are twice as many 470s then how does he explain this ratio in the article?

 

[Dark4ng3l]

The problem IMO is that Nvidia were trying harder and harder to make the best compute chip that they can. The way I see it they think HPC market is their big thing and probably the future of the company. To be fair if they can develop that market they don't have any direct competition right now. But the problem is that some decisions (100% ''full shaders'', going to a more brute force things with computing power instead of building fixed function hardware for it(ie tessellation)) are making their efficiency as a video device suffer terribly.

Now Fermi is mostly a compute chip rigged to be a graphics device. Sure it's great if you want to do heavy computing tasks on it but it is probably far from an optimal device for gaming purposes. I think they should have gone with 2 different products entirely(but the rumors are that they cancelled the successor to GT200 so they were probably doing just that).

In the end it remains that Fermi's biggest problem is really the fact that it's too big for the process it's being made on and thus the yields are bad(so you need to up the voltage and cut some shaders out to get a good amount of functional hardware).

The problem for nvidia right now is that if the next generation of chips ends up being 40nm again ATI has room to tweak and add some die space if it's necessary. Even if we assume that the 40nm process matures quite a bit in this time frame Nvidia are still in a position where they would want to try and reduce the size of the chip and certainly not making it any bigger. In the end any effects of process maturity helps both sides and so Nvidia won't be able to pull ahead by banking on the process becoming much better.

 

[OCGuy]

They are going to sell a crapton of Fermi. (As much as TSMC allows, anyway)

Then they are going to sell a crapton more of the re-spin, which hopefully takes advanage of all clusters and draws less power.

It will also hopefully have higher memory clocks. nV has never used DDR5, and thier controller sucks. That is why the memory is clocked lower than spec.

All in all, I dont think nV is worried about much at the moment.

They do need to worry about the next generation though. The large-die strategy is showing it's weakness. Not with raw performance, mind you, but with thermal limitations.



ATi had a real chance here to steal a large amount of marketshare. However, their continued pricing of 5870 is preventing what could have been spectacular for them.

They are not gouging, they just cannot afford to price 5XXX like they did 4XXX. That was a loss-leader that everyone is paying the piper on now.

 

[T2k]

They have no choice.
NV has no license for x86 and last year they lost their entire chipset business (accounted for one-third of their revenue!), handheld/mobile chip sales make nothing compared to desktop especially WS sales and the discrete market is rapidly going away (into the CPU a' la Llano and Intel HD) sans niche, really high-end gaming cards... they have to create/carve out their new market, be it as small as it is right now.

 

[BenSkywalker]

ATi had a real chance here to steal a large amount of marketshare. However, their continued pricing of 5870 is preventing what could have been spectacular for them.

While the situation has been improving quite a bit, ATi keeping their margins high on the 58xx parts is the only smart thing to do given the supply available in the pipe. It makes little sense to have an incredible value that noone can buy. For once, ATi is in the position that nV regularly finds themselves in and are doing precisely what nVidia would do. That is a good indicator honestly, taking advantage of marketplace realities is key to operating a good business.

They are not gouging, they just cannot afford to price 5XXX like they did 4XXX. That was a loss-leader that everyone is paying the piper on now.

Both the 4xxx parts and the GTX2xx parts had margins in the 30%-40% range, in most industries that would be dancing in the streets stellar. They were not producing loss leaders and they could have reasonably priced them quite a bit lower still without them falling into that category. That said, eveyone in this segment shoots for Intel's 60%+ margins as an ideal, so it certainly wasn't like it was an insanely profitable part either(it was directly comparable to nV's margins, if they sold as many units as nV had they would have been more profitable as ATi has lower operating margins then nV).