Nvidia's Fermi GTX480 is Broken and Unfixable

[The_DJax]

SemiAccurate :: Nvidia's Fermi GTX480 is broken and unfixable

Number one on Nvidia's hit list is yields. If you recall, we said that the yield on the first hot lot of Fermis that came back from TSMC was 7 good chips out of a total of 416 candidates, or a yield of less than 2 percent.

Why are things this bad? The simple answer is that Nvidia didn't do its homework. Just like the failures that lead to the bad bumps, Nvidia simply didn't do test, instead it tried to brute force things that require nuance and methodical forethought. ATI did test. ATI put out a full production run of HD4770 (RV740 silicon) GPUs and used that as a test of TSMC's 40nm process. TSMC looks to have failed, badly, but crucially, ATI learned why and how the parts failed to yield. That learning was fed back in to the Evergreen 5000 series GPUs, and those come off the line at perfectly acceptable yield rates now.

Really interesting read of why NVidia is really hurting with Fermi.

[Afrodisiac]

Good read, and I recall something JJ said a while back which I had trouble buying into until now; he mentioned that this year might very well be Nvidia's last.

[orangegator]

Good read, and I recall something JJ said a while back which I had trouble buying into until now; he mentioned that this year might very well be Nvidia's last.

Let's hope not. The last thing I would like to see is Nvidia going bankrupt and then possibly being bought by Intel or just disappearing altogether. We need good competition and price wars.

[The_DJax]

Let's hope not. The last thing I would like to see is Nvidia going bankrupt and then possibly being bought by Intel or just disappearing altogether. We need good competition and price wars.

Agreed. I would hate to see them go down the drain, so to speak. I think they might need a restructuring inside the company, perhaps, but I don't want them out of business. That would be bad news for all consumers.

[Holy_COW]

Not bad - terrible!

If Intel buys nVidia may be fun at the beginning, Intel will finally have a decent GPU to include in a CPU but will damage a lot Ati/AMD, and since AMD is not so well on the finances...

We do need someone to keep the price wars and technology/innovation war, if not...

[Holy_COW]

p.d.

and please dont tell me they made 3Dfx broke and then bought them to just die years later!

3Dfx >>> All

[m25]

Let's hope it does not happen but it's really tough for nvidia. their stubbornness in any circumstance has been devastating the last couple of years and it's hard to recover when things are going so bad you still claim to be number one and what your competition came up with is just a toy, because you,... you have the real thing coming
Nvidia has to take some serious action, steer tighter than AMD did in 2007 or they're going to get that big iceberg they have in front, I'd not really like a world shared by Intel and AMD, but sadly that is where we're heading for the moment.

[The_DJax]

I would've never seen this coming a year ago. AMD was in trouble back then, but look where it's landed itself now.

Spun off Globalfoundries, went Fabless, got $$ from Intel in settlement, and produced two solid series' of graphics cards (4000 and 5000 series).

I think Nvidia should be fine if they just tighten their belts and don't give in.

[CapnBFG]

Notice who write this...

Nonetheless, he echoes some of the things that Anand pointed out in his article a couple days ago. This is what I was getting at with this post a while back:
GT300 / Tesla / Fermi Board Fake?

A proper learning vehicle (in ATi's case it was 4770) is needed to understand how to make things yield, to know which ground rules can be pushed and which cannot. Pushing against ground rules is bad for yield, making a huge chip is bad for yield, and not using redundant vias is ****ing terrible for yield. I was talking with a senior engineer about this today, and he said that he first started using redundant vias in his layout in the mid to late '80s. He was absolutely baffled that it is not standard practice.

Regarding the author's comment that TSMC failed badly, I don't agree with that. Clearly their process works, you just have to know how to design for it.


[EDIT]
Didn't finish reading the whole article yet because I need to go to sleep, but I just want to comment on a couple things he said:

1) reworking metal levels doesn't help yield
That's partly true. If their problems are with vias and leaky devices like everyone is saying, then no, it won't help. If they fixed wires that were undersized, then yes it might help, either for timings or end-of-life degradation. Charlie was right that metal level fixes are normally for logic patches, but from his wording I got the impression that he didn't really understand what he was saying.

2) a full relayout is needed to fix the problems
If what has been claimed about their specific issues is true, then yes. If they did in fact build everything to the min size possible and didn't leave any fat, and if they really did use single vias to save space then it is presumable that it is the case, then there is no way to drop redundant vias in without relaying most of the circuits. That is the sad truth.

[The_DJax]

I actually found this article linked on Anandshimpi's Twitter page.

EDIT: Haha looks like they'll have to do another re-branding!

[joejamesatou]

I'm waiting for their ****y CEO to spin this with bravado.

[billdcat4]

Nvidia isn't going any where

The revenue from Desktop gaming GPUs is a small part of their total revenue.

They sell IGPs, discrete GPUs for Laptops, and other devices (Tegra)

Having the fastest GPU is more of a status issue than a pure-business motive

Dell, HP, Lenovo, and the other companies that are selling Nvidia GPUs in their notebooks don't really care if Nvidia's architechture is old. As long as they keep rebranding the parts, the average consumer, where Nvidia gets most of their money from, will never know the difference.

[Viperabyss]

I wouldn't put too much weight into this article, as the name of the site as well as the author don't usually rhyme well with facts and professional journalism.

With that said however, I believe that Nvidia may be pulling their R600 with the launch of Fermi. I could be wrong, but it certainly looks that way from all the information we know so far.

[Simul4ti0n]

Wow, that sounds really bad but who knows for sure? Nobody but nvidia and TSMC.

[logain]

Nvidia's arrogance is finally catching up to them. It seems it happens to every company that is on top for too long.

[fidgewinkle]

Notice who write this...

Nonetheless, he echoes some of the things that Anand pointed out in his article a couple days ago. This is what I was getting at with this post a while back:
GT300 / Tesla / Fermi Board Fake?

A proper learning vehicle (in ATi's case it was 4770) is needed to understand how to make things yield, to know which ground rules can be pushed and which cannot. Pushing against ground rules is bad for yield, making a huge chip is bad for yield, and not using redundant vias is ****ing terrible for yield. I was talking with a senior engineer about this today, and he said that he first started using redundant vias in his layout in the mid to late '80s. He was absolutely baffled that it is not standard practice.

Regarding the author's comment that TSMC failed badly, I don't agree with that. Clearly their process works, you just have to know how to design for it.


[EDIT]
Didn't finish reading the whole article yet because I need to go to sleep, but I just want to comment on a couple things he said:

1) reworking metal levels doesn't help yield
That's partly true. If their problems are with vias and leaky devices like everyone is saying, then no, it won't help. If they fixed wires that were undersized, then yes it might help, either for timings or end-of-life degradation. Charlie was right that metal level fixes are normally for logic patches, but from his wording I got the impression that he didn't really understand what he was saying.

2) a full relayout is needed to fix the problems
If what has been claimed about their specific issues is true, then yes. If they did in fact build everything to the min size possible and didn't leave any fat, and if they really did use single vias to save space then it is presumable that it is the case, then there is no way to drop redundant vias in without relaying most of the circuits. That is the sad truth.

I'm pretty sure nVidia knows how to put the right number of vias in their design. In any case, this would be fixed in the first metal fix. As for power problems, it is possible that they could have crosstalk problems that could be fixed and dramatically improve transistor output loads and improve performance at a lower power consumption.

If their problems are in their transistors or substrate isolation, then they are screwed.

[CapnBFG]

I'm pretty sure nVidia knows how to put the right number of vias in their design. In any case, this would be fixed in the first metal fix.

I'm sure they know now. Why would you assume they knew it previously? ATi didn't know it until after their first chip on this process. And yes they would fix it during their metal respin, but that is assuming they could fit it with the current circuit placement, which is not always the case.

If their problems are in their transistors or substrate isolation, then they are screwed.

True.

[Mandrake]

This probably explains why Nvidia needed another lot of silicon for Fermi and why the launch has been put off for so long. Whatever the reason, I refuse to buy a new card until prices come down. If I'm stuck with my museum piece 8800 GTS for another few months then so be it.

Good read, and I recall something JJ said a while back which I had trouble buying into until now; he mentioned that this year might very well be Nvidia's last.

I think this is news to Nvidia themselves!

NVIDIA - News Release

[joejamesatou]

This probably explains why Nvidia needed another lot of silicon for Fermi and why the launch has been put off for so long. Whatever the reason, I refuse to buy a new card until prices come down. If I'm stuck with my museum piece 8800 GTS for another few months then so be it.



I think this is news to Nvidia themselves!

NVIDIA - News Release

Museum piece? What are you talking about? To Nvidia that is the core of a brand spanking new GTS 250! Someone should make a joke press release about Fermi being another rebaged G92 core.

[DaSickNinja]

Fermi broken? Most likely yes. But unfixable? Technically, the 5870 is a fixed 2900XT. So no, nothing is unfixable.