NVIDIA Approaching Other Foundries than TSMC for 28 nm Production

well...sometimes breakneck pace just isnt good enough. there is a serious shortage of 680s though I admit

what about intels fab?

Last I heard they did contact intel, but no word on what came of it, would be sweet though.

Good! I think that Nvidia + Samsung would do a much better team for producing my GPU than Nvidia and this joke of a manufacturer that has been proven unreliable throughout these years. I am someone who still remembers the 800 series problems and how everybody blamed Nvidia for it although it was clearly a manufacturing problem and TSMC were the ones responsible for it. Good riddance I say!

Would changing fab's do anything to the performance?

Delta6326 said:

Would changing fab's do anything to the performance?

Click to expand...

It shouldn't have an effect on the actual Performance, it's more about availability. What good is Nvidia having the most powerful single GPU if they can't stock shelves ANYWHERE with it? They are still delaying releasing the rest of their line as well, because of a lack of supply from TSMC. The only problem with this is that to my knowledge TSMC is really the only company with any capacity for 28nm products. I know GloFo is working on it, but they are nowhere near what TSMC can offer.

EDIT: Just read the end of the article, was unaware Samsung had a 28nm line.

Katanai said:

I am someone who still remembers the 800 series problems and how everybody blamed Nvidia for it although it was clearly a manufacturing problem and TSMC were the ones responsible for it.

Click to expand...

800 series?

btarunr said:

Intel lacks 28 nm. There's 25 nm (NAND flash, IMFlash JV), and there's 22 nm (processors). Optical-shrinking Kepler to any of those will take another 4 months (sampling, testing, qualification, moar testing, mass production).

Besides, NVIDIA will not risk giving its designs to Intel (a GPU competitor, which could steal its designs for processor graphics).

Click to expand...

I just imagined an Intel CPU paired with Kepler based IGP, wow that'd be awesome for notebooks

radrok said:

I just imagined an Intel CPU paired with Kepler based IGP, wow that'd be awesome for notebooks

Click to expand...

For Nvidia GPU to be an IGP in Intel's CPU, Intel would have to buy NVIDIA (just as AMD had to buy Ati). Nvidia's market value is about $7.3 billion and Intel although worth $143 billion has about $7.5 billion in cash, so right now Intel would have to get into substantial debt for this transaction. It could be happening in couple of years if Intel continues to grow at this rate.

BiggieShady said:

For Nvidia GPU to be an IGP in Intel's CPU, Intel would have to buy NVIDIA (just as AMD had to buy Ati). Nvidia's market value is about $7.3 billion and Intel although worth $143 billion has about $7.5 billion in cash, so right now Intel would have to get into substantial debt for this transaction. It could be happening in couple of years if Intel continues to grow at this rate.

Click to expand...

true...nvidia is still relatively healthy and it would be pricey....long term though it might be worth it. Some of intel latest actions(or could be my overactive imagination) lead me to believe they have or are on the verge of making some serious headway in the graphics department.

Isenstaedt said:

800 series?

Click to expand...

It was a typo. I hate this keyboard!!!

Saw this coming. All those problems TSMC has had. Was waiting for one company to get smart and think..."gee, maybe we can get this made elsewhere without the hassle and problems?" Go NV!

why you wait so long nv?

i mean its not like tsmc have been faultless in the past as most of nv issues the past 5 years have been caused at the fab.

btarunr said:

Intel lacks 28 nm. There's 25 nm (NAND flash, IMFlash JV), and there's 22 nm (processors). Optical-shrinking Kepler to any of those will take another 4 months (sampling, testing, qualification, moar testing, mass production).

Click to expand...

AFAIK changing to another foundry also requires more testing and qualification. And Samsung's 28nm is gate-first, unlike TSMC's gate-last process, so that surely means a lot of work to change.

If Nvidia is decided to change to another foundry, it means it's ready to loose a few months testing the new process and if that is the case, I'm sure they wouldn't mind spending a few extra weeks if that means making their GPUs on a much smaller and reliable process, such as Intel's 22nm.

Besides, NVIDIA will not risk giving its designs to Intel (a GPU competitor, which could steal its designs for processor graphics).

Click to expand...

I don't think that's a real problem right now. First of all making the hardware is not the most difficult task for a company doing processors for 40 years, it's the drivers what really makes the difference between a CPU an a GPU. HD3000 and specially HD4000 are already a very decent piece of hardware for their size.

Second and most important, Nvidia already shared their patents when they settled the lawsuit with Intel, so I don't think there's much more secrets to be found in the "silicon". IMHO any secrets/tricks that might be found on silicon (power reduction, lower latency, higher clock...) probably Intel knows better*.

And in fact, Nvidia CEO already called Intel to start making ARM chips on contract. Newer Tegras will have Kepler GPU inside so they surely aren't very concerned about Intel stealing anything if they want Intel to make future Tegras for them.

* BTW who's to say that a lot of the improvements in Kepler didn't come from the patents Intel shared as part of the deal? Even GPU Boost is similar to Turbo Boost, in the name too, where you might risk a lawsuit.

well, AMD didn't have any big problem with TSMC. it's kepler low yields that made 680 scarce..

Jonap_1st said:

well, AMD didn't have any big problem with TSMC. it's kepler low yields that made 680 scarce..

Click to expand...

You say AMD did not have any problems, however, cards release with 925 default clock when nearly every one hits over 1 GHz and then some. To me, this indicates wildly varied silicon quality, as the 1 GHz milestone is too big to skip over with a flagship GPU. They did not release these GPUs @ 1 GHz, to me, becuase they couldn't, and that indicates a problem.

Seems to me that AMD's success is related to the number of wafers they purchased, and nvidia cannot get enough wafers, so is looking elsewhere. It doesn't realyl indicate rel problems at TSMC, other than that they cannot meet consumer(OEMs are their consumers) demand.

If nVidia is sampling Samsung process already, they are running wafers out to verify yields.

cadaveca said:

You say AMD did not have any problems, however, cards release with 925 default clock when nearly every one hits over 1 GHz and then some. To me, this indicates wildly varied silicon quality, as the 1 GHz milestone is too big to skip over with a flagship GPU. They did not release these GPUs @ 1 GHz, to me, becuase they couldn't, and that indicates a problem.

Seems to me that AMD's success is related to the number of wafers they purchased, and nvidia cannot get enough wafers, so is looking elsewhere. It doesn't realyl indicate rel problems at TSMC, other than that they cannot meet consumer(OEMs are their consumers) demand.

If nVidia is sampling Samsung process already, they are running wafers out to verify yields.

Click to expand...

ahh.., so it's like AMD bought nearly an entire for what TSMC had selled on their store, and nvidia only get a tip of it. maybe in the future whoever got the contract faster and had succesfully made sure that their chip yields is on good amount have the bigger chance to avoid shortage like what happen to nvidia right now, they starting to loose interest from customer..

AMD/ATI has almost always been to new process first. They just capitalized on it this go around, and Nvidia didn't let the same mistake with power gating occur again.

Delta6326 said:

Would changing fab's do anything to the performance?

Click to expand...

No, but it could have the effect of every chip coming of the production line like the best binned ones from a poorer fab - meaning lower stable voltages (and thus power consumption) or higher stable frequencies (and thus performance).