Jen Hsun Huang on Fermi's Failure, in the Development Stages

[cadaveca]

Let me put his own words here:

Imagine if we were all processors, and we all just works. But we can not talk to each other. We are all deaf and dumb.

- Huang in reference to Fermi's SPUs


Google Translate from German:

http://translate.google.com/transla...=org.mozilla:en-US:official&channel=s&prmd=iv

Original in German:

http://www.golem.de/1009/78179.html

 

[Mr McC]

All deaf and dumb? Why didn't they use written signs?

 

[afw]

video direct link ---> http://www.youtube.com/watch?v=RPUaSd4KM70

Well .. at least he admits the truth .... and the lesson learned from this will ensure it doesn't happen again (hopefully) ... but its kinda lame to put the blame on TSMC though ...

 

[Benetanegia]

but its kinda lame to put the blame on TSMC though ...

He has in no way put the blame on TSMC (did you actually watch the video you linked??). He doesn't even mention TSMC or 40nm manufacturing process once. The only time he mentions the fabs is to say that when the chips arrived from the fabs they found they didn't work, because the two Nvidia teams had messed up something "easy", because of a typical mistake with "easy" things that fit in between two important things and "anyone" can take care of it: no one does it in the end. It happens everywhere, it happened to NASA (km -> miles messup), it (supposedly) happened with CIA/NSA/FBI with the 11 S... Nevertheless a BIG mistake and Jen-Sun is admitting it to be his own fault.

 

[Completely Bonkers]

Er, what is this thread about?

Huang does not admit Fermi's failure. Far from it. He explains why there was a delay: because of a problem with the prototypes, where the interconnects between the 512 processors were defective.

He also uses the "dumb and blind" symbolism to draw a parallel between the defective prototype and also his division an nV where there were 2 departments not talking together (effectively).

He then goes on to say how brilliant Fermi is and that the industry should continue to focus on innovation, not integration of IGP onto CPU which is a totally proprietary land-grab and doesnt benefit the customer.

 

[Bjorn_Of_Iceland]

yeah thats what I remember as well as to the delay of the chip. A bug with the multithreading.. I think they revamped the gigathread engine for it.. thats what he was stating broken

 

[yogurt_21]

Er, what is this thread about?

Huang does not admit Fermi's failure. Far from it. He explains why there was a delay: because of a problem with the prototypes, where the interconnects between the 512 processors were defective.

He also uses the "dumb and blind" symbolism to draw a parallel between the defective prototype and also his division an nV where there were 2 departments not talking together (effectively).

He then goes on to say how brilliant Fermi is and that the industry should continue to focus on innovation, not integration of IGP onto CPU which is a totally proprietary land-grab and doesnt benefit the customer.

obviously otherwise the gtx470 sitting in my rig couldn't even render based on the description of the issue. He admits the mistake in oversight (ie who's in charge of the final product) and that was the casue for the delay of fermi. while there are alot of people on here who think fermi was a failure I disagree, my gtx470 works fine. The issue is that it was too little, too late. Had it launched on time I think it would have been a different story. But to launch 6 months after ati's variants, and then still have the heat and power consuption issues was a bit of an insult to nv's customers.

but if we change that timeline and were to place fermi alongside the cypress launch and then put the gf104 launch 6 month's out we would be telling a different story as the heat and power issues could be attributed to a rushed launch and customers typically forgive a company for that. But a late launch, power and heat issues, and the fact that performance wasn't near what we thought it would be has caused the vista effect with fermi.

 

[afw]

He has in no way put the blame on TSMC (did you actually watch the video you linked??)

Well to be honest I dont knw much about making a gpu ... I said that cos I read an article stating so ---> http://www.tweaktown.com/news/16876/nvidia_blames_tmsc_for_fermi_s_failures/index.html (original source is semiaccurate though )

 

[JATownes]

I don't think Fermi is a failure. The GTX460 & GTX470 are REALLY fast video cards, that have been marketed at a competitive price. Not to mention that the GTX480 is still the single GPU champ. The only problem was power consumption and heat. The heat issues has (somewhat) been worked out with aftermarket coolers, etc., but they are still a little power-hungry. I think Nv really has done something right, simply due to the fact that they have a GPU that can withstand that much heat and power. We all assumed, me included, that Fermi would have a massive failure rate due to heat/power issues, and that has not happened. I say Fermi is a good architecture, it justs needs a little tweaking (die shrink) in its next iteration.

 

[wolf]

The only 'faulire' Fermi endured was the launch itself, like yogurt 21 said, if it launched alongside the 5000 series the story would have been much different, with them equaling/beating high end 5800 single GPU's on the spot at the cost of a little more power and heat. that would have been easily forgiveable all on its own, but 6 months late was the real nipple twist.

If it had have launched there and then they most likely wouldhave gone into GF104/106/108 development and production straight away, instead of spending that 6 months sorting out wtf happened to GF100.

now having said that they cost themselves 6 months, I believe the company is made up of some great minds, and they have a plan to, catch up/stay on track/maybe even take the lead in the near future, who knows.

I find myself reiterating very often that a market with no competition is not what we want at all. so to those who hate Nv and love AMD, it's still in your best interests to have Nv competitng very well.

 

[Benetanegia]

Well to be honest I dont knw much about making a gpu ... I said that cos I read an article stating so ---> http://www.tweaktown.com/news/16876/nvidia_blames_tmsc_for_fermi_s_failures/index.html (original source is semiaacuate though )

Yeah, he's linking to the latest Demerjian's gem. Oh good ol' Charlie, it's always Charlie and his hordes of brainwashed sheeps. Apparently not even news editors can think for themselves anymore... I suggest you always go to the source, but oh yeah finding the source in SA is usually a dead end, that ends in another SA "article".

But you have to give credit to Charlie though, not everyone is able to manipulate so many people in that way so that they don't think for themselves at all and resort to agreeing with and replicating "articles" in which he takes 25 seconds from an interview 5 mins long, puts the quote completely out of context, makes his own conclusions of what Jen-Hsun wanted to say, because he sees something between the lines, takes it way out of proportion, distorts the consecuences of the implications he just fabricated, regurgitates it all on his hate for Nvidia and finalizes with a self-inflicted egomaniac orgasm... Genius!

 

[CDdude55]

Pointless thread is pointless...

He didn't mention anything about Fermi being a failure, more of just what was wrong with it during development and how they did their best trying to fix it. And as said, it's all in the past.

And of course considering i own a 470, i can say, it's an awesome card.

 

[the54thvoid]

Jeez, lock the post before the hurting starts.

"He then goes on to say how brilliant Fermi is and that the industry should continue to focus on innovation, not integration of IGP onto CPU which is a totally proprietary land-grab and doesnt benefit the customer."

Physx for f*cks sake, talking about pot calling the kettle black.

"Oh good ol' Charlie, it's always Charlie and his hordes of brainwashed sheeps"

I read S/A, and Anand and Hexus and here and Bit Tech and X-Bit Labs and wherever i can source english matter. Yes, Charlie's a shit journalist in that he is very opinionated and for that loses the middle ground. He also has a vendetta against NV which makes him very partial and prone to rants but ocassionally he hits home and usually the NV camp dont like it.

So, does anyone have a 512 core 480, like Charlie said wouldn't work?
But you do have excellent 460's which he stayed very quiet about.
There you go, two sides, he's right, he's wrong or he stays quiet.

Lets get rational people.

 

[{JNT}Raptor]

He admits a production defect due to non-communication and the OP posts like he admitted Fermi itself is a failure??......this thread is flamebait from the start and should get renamed or binned.

 

[the54thvoid]

it (supposedly) happened with CIA/NSA/FBI with the 11 S...

Watch that area Ben, we've just had a nutter Iranian call it a US conspiracy, best leave any 9/11 anecdotes out. If someone wants to kill you enough, they will. Civil liberties allow certain freedoms, that was the main culprit for 9/11. It's hardly comparable to the Fermi issues.

Please lets not over react.

 

[kid41212003]

Do we even need to read this? I'm using the GTX480. It's a fantastic product .

 

[cadaveca]

It takes a far stronger man to admit defeat, and learn from past mistakes, than it does to deny there's a problem at all.


While alot of you seem nV fanboyish, I'm gonna say that Mr. Huang does admit there was a fault between the communication between teams within nV. He even says that Fermi was 256 units, but that the crosstalk was too high, and that lead to the delay, and the disabling of units.

I don't realyl think Fermi is such a failure either...many of you have seen me complaining about my 5-series cards, so it's not like nV is the only one with issues this gen...

However, my respect for Mr. nV has gone up a notch...he used to gloss over failures in the past, and his staff are a great bunch of people. I think that they are capable of more than what Fermi is now, and Jen Hsun admits that(in the link is not the full video, BTW.)


Anyway, what he's saying is that they know exactly what went wrong with Fermi, and that they won't be repeating what was kind of an amateur mistake...


I'm sorry if you guys don't like nV put into a bad light...personally I could care less aqs it's business, but really, this shows more character and honesty than Mr nV has ever shown...


And that's a good thing. Only those ignoring the facts see it as bad. Nothing wrong with humility, and being humble.

 

[Benetanegia]

He also has a vendetta against NV which makes him very partial and prone to rants but ocassionally he hits home and usually the NV camp dont like it.

Even a broken clock gives the correct time twice a day. And a coin will be accurate 50% of times if you throw it enough times.

So yeah, sometimes he is right, but considering the kind of claims he makes, a 50% accuracy is guaranteed, I mean it's pure statistics, a coin can guess the same ammount as him.

Let's recoupd some of his latest Nvidia related claims:

- GF100 hot and power hungry. true (yay!)
- GF100 unmannufacturable, 2% yields, 8000 cards at launch... false, 400.000 where shiped in the first month, 20x as many as Cypress at launch.
- GF104 as hot and power hungry as GF100. false
- GF104 exactly the same architecture as GF100, 3 blocks being put in a row. false
- no one wants Fermi Tesla cards. false (unless IBM, Cray, US goverment are "no one")
- no one wants Tegra 2. false (here the entire article was about how it takes months to create embedded products and how no one was creating them at the time, so no Tegra2 devices, that was a fact, from his sources. well unless you have to be an inter-galactic imperium to be "someone" according to Charlie... HTC, Motorola, LG, Samsung (20+ more vendors)... idk they look like "someone" to me, but hey who am I? I'm not uber-Charlie so I must be wrong, he must be right)

EDIT: And all that is irrelevant since we have a much better proof of how Charlie really is, in this latest interview. Mr. Huang is saying one thing, CD is saying a completely different thing, something 100% fabricated, like always, but now is evident, CD really exceeded this time on psichotropics and has been caught as the liar he really is. Even some long time readers are telling him to STFU on this one on SA. The fact that still has some sheep left defending him is amazing though, and I can only humbly give credt to him again for his mastery.

 

[CDdude55]

It takes a far stronger man to admit defeat, and learn from past mistakes, than it does to deny there's a problem at all.


While alot of you seem nV fanboyish, I'm gonna say that Mr. Huang does admit there was a fault between the communication between teams within nV. He even says that Fermi was 256 units, but that the crosstalk was too high, and that lead to the delay, and the disabling of units.

I don't realyl think Fermi is such a failure either...many of you have seen me complaining about my 5-series cards, so it's not like nV is the only one with issues this gen...

However, my respect for Mr. nV has gone up a notch...he used to gloss over failures in the past, and his staff are a great bunch of people. I think that they are capable of more than what Fermi is now, and Jen Hsun admits that(in the link is not the full video, BTW.)


Anyway, what he's saying is that they know exactly what went wrong with Fermi, and that they won't be repeating what was kind of an amateur mistake...


I'm sorry if you guys don't like nV put into a bad light...personally I could care less aqs it's business, but really, this shows more character and honesty than Mr nV has ever shown...


And that's a good thing. Only those ignoring the facts see it as bad. Nothing wrong with humility, and being humble.

But no ones denying anything, it's the way the thread comes across ''OMG jen says Fermi is fails!!!1!!'' when it fact, he said nothing about it being a failure, the discussion was about the manufacturing issues they had and how they the problems got fixed one way or another. How is anyone ignoring facts?

 

[cadaveca]

If you say so. I actually posted this to show nV is on the ball, and isn't hiding the mistakes they made...and that's not inaccurate. It's your own interpretation of my post that has you seeing it that way.

I even said "In the PAST TENSE"...in other words, they've moved on from that period of time. I haven't skewed anything around here...I even put the quote, and not my own thoughts, for a reason. Where's my biased view?

Jen Hsun says:

"We found a major parasitic problem between the foundry, the tools, and reality. Everything appeared to be working, but really, things weren't connected at all...imagine we are a bunch of processors, but we are all deaf and dumb..we can't talk at all..the fabric...it was completely broken"(refering to first Fermi samples)


That's the failure. But in the end, they fixed it, mostly. And I find it pleasantly shocking that Huang admits this.

 

[CDdude55]

He admits a production defect due to non-communication and the OP posts like he admitted Fermi itself is a failure??......this thread is flamebait from the start and should get renamed or binned.

This^

 

[bpgt64]

obviously otherwise the gtx470 sitting in my rig couldn't even render based on the description of the issue. He admits the mistake in oversight (ie who's in charge of the final product) and that was the casue for the delay of fermi. while there are alot of people on here who think fermi was a failure I disagree, my gtx470 works fine. The issue is that it was too little, too late. Had it launched on time I think it would have been a different story. But to launch 6 months after ati's variants, and then still have the heat and power consuption issues was a bit of an insult to nv's customers.

but if we change that timeline and were to place fermi alongside the cypress launch and then put the gf104 launch 6 month's out we would be telling a different story as the heat and power issues could be attributed to a rushed launch and customers typically forgive a company for that. But a late launch, power and heat issues, and the fact that performance wasn't near what we thought it would be has caused the vista effect with fermi.

I agree, I have GTX 470s and there some of the best cards I have ever had...My wife loves her GTX 460. And the two comps I built for friends are using 460s in SLI, and they love the performance they get. I sold my HD 5870s, they were great cards, but Crossfire wasn't all that great

 

[cadaveca]

This^

If you watch the video...Jen Hsun was asked: "What broke in Fermi."

He responded "Well, I'll give you two answers, first the technical one, and secondly, the management one".

I'm not twisting nothing here. Go to 2:45 in the video and watch from there...you'll see that what went wrong in Fermi is exactly what that video is all about. He even repeats that exact question.

He then says:

"That's not really the question at all...the question is, how do you prevent this from happening again? Or next time, it will be something else. We don't know what we don't know. It turned out, the reason the fabric failed, wasn't because it was hard...but because it sat between the responsibility of 2 groups."(meaning two developmetn departments within nV).

Huang is being very candid about what went wrong during the development of Fermi, and that's pretty important, if you ask me.

It doesn't even have anything to do with the retail products...this is all before launch, when Fermi was in development. Like the link I posted about the 5-series development...it's this stuff that interests me.

 

[CDdude55]

If you watch the video...Jen Hsun was asked: "What broke in Fermi."

He responded "Well, I'll give you two answers, first the technical one, and secondly, the management one".

I'm not twisting nothing here. Go to 2:45 in the video and watch from there...you'll see that what went wrong in Fermi is exactly what that video is all about. He even repeats that exact question.

He then says:

"That's not really the question at all...the question is, how do you prevent this from happening again? Or next time, it will be something else. We don't know what we don't know. It turned out, the reason the fabric failed, wasn't because it was hard...but because it sat between the responsibility of 2 groups."(meaning two developmetn departments within nV).

Huang is being very candid about what went wrong during the development of Fermi, and that's pretty important, if you ask me.

Right, i think it's good to know what happened to Fermi during the developments phase as it would give us more insight into the delay. And luckily the majority if not all the problems have subsided.(besides the slight inefficiency of GF100). I think the substance of what you're trying to get at is important for not only us but also for Nvidia to know. And it's a learning process, shit happens and sometimes things don't turn out the way you hoped at first, but you just have to break through it, and that's what Nvidia did. Regardless of the delays, they bought it out and they brought some impressive power with this architecture, is it perfect?, of course not. I don't see what what needs to be talked about, i think they learned from developing each architecture that things don't always goes as planned and that sometimes you just have to make sure you get something out, and make sure that something is even remotely close to what the enthusiasts want, and i think in that regard alone, Fermi was a success.

 

[cadaveca]

Yes, I understand that, but that's why I put "In The Past Tense", and the quote about the SP's being "deaf and dumb"...clearly this is not the case in the retail products, or they wouldn't work.

Also, Huang doesn't normally talk about this sort of stuff at all...

It's a very cool video, showing alot of humanity from Mr nV, and that's why I posted the thread. I didn't add my own comments, I gave the title, and the info for you'all to watch.

Retail chips are NOT called Fermi, to me. Fermi was the chip's name while in development, but that chip, that is is retail cards, is now known as GF100, right?

Fermi has come and gone, to me, so to look back at what happened during that time is some of the best "news" there is...that's real journalism, and I love it!

And now we have the name for the next chip in development...I lok forward to seeing how it pans out, and then getting more synopsis like this one, and the 5-series one on Anand, some time in the near future. I think this is good publicity for nV.

If i thought GF100 was so bad, I wouldn't have two here on my desk.