[confirmed] Industry Wide Implications: Samsung, TSMC 16nm FinFET Extremely Low Yields For Apple, Others

[patrickjp93]

http://www.fudzilla.com/news/36648-intel-could-save-apple-s-bacon

It turns out despite Samsung having 16/14nm fabrication technology, and despite TSMC's boastful claims of 16nm FinFet being ready by Q1 2015, that just may not be true.

It turns out Apple gave production contracts to both companies for its 16nm A9 chips based on tri-gate (FinFET) designs, but neither can produce these transistors in remotely tolerable yields. Furthermore, as some analysts are suggesting, it is possible Apple could drop business with these two companies and work out a solution with Intel (having mastered 3D transistor production) in an effort to stay on the bleeding edge of technology.

While this news in and of itself is inconsequential for most here, we can see the potential harm this does to competition between Intel and AMD in 2016, when Zen is due to be released on a 16/14nm FinFET process. We all know how long it took Intel to get yields up, and to have expected less struggle across the industry would have been a tad foolhardy, but this means no one apart from Intel will have 14nm working for some time, and construction of the Israel 10nm foundry has already begun.

Furthermore, this could spell disaster for Nvidia which has gambled on saving the Pascal architecture for the 16nm FinFET TSMC process, leaving AMD a clear and open field for an entire year or more on the GPU front.

This also means Intel's Atom processors which it is still trying to push into the phone market will gain even more process advantages and continue to shrink power and heat envelopes, meaning 2015-2016 could be the years Intel takes a foothold.

These foundry problems have widespread implications, some demoralizing, some positive depending on who you root for in this giant game.

Update: TSMC and Samsung are both trying to save face without revealing their own low yields. http://appleinsider.com/articles/14/12/30/conflicting-reports-claim-tsmc-samsung-each-have-upper-hand-for-building-apples-a9-chip

Update 2: Discussion on Anandtech (with included sources and press releases) just last week on the subject of Samsung's yields: http://forums.anandtech.com/showthread.php?t=2414548

[LukaP]

Im not sure how much of an impact it will really have. Intel will not have 10nm functional before 2016 anyway, if they are still building the fabs. so Zen will be competitive (and even if intel does get a process node advantage in the future, that was always there anyway)

 

and considering AMD Nvidia thing. neither of them will have access to 16/14 FinFET unless they make a deal with intel, which i dont see happening with AMD, and ive no idea how NV could make it happen, apart from maybe some GPU IP deal. so they will be on the same node regardless, and if needed, to beat amd, Pascal will be renamed and moved onto 22nm TSMC for the time being

[ilikemacandpc]

TSMC is terrible. Intel (as fanboy as this sounds) is awesome. 

[patrickjp93]

Im not sure how much of an impact it will really have. Intel will not have 10nm functional before 2016 anyway, if they are still building the fabs. so Zen will be competitive (and even if intel does get a process node advantage in the future, that was always there anyway)

 

and considering AMD Nvidia thing. neither of them will have access to 16/14 FinFET unless they make a deal with intel, which i dont see happening with AMD, and ive no idea how NV could make it happen, apart from maybe some GPU IP deal. so they will be on the same node regardless, and if needed, to beat amd, Pascal will be renamed and moved onto 22nm TSMC for the time being

Cannonlake is still due for release in summer 2016, so Zen (designed for the 16nm FinFET process) may have even less time to dance with Skylake than AMD would have wanted.

 

AMD has 20nm GPUs coming out right after CES. Nvidia still has Maxwell only on 28. That means they have to do a decent amount of redesigning to get Pascal onto the 20nm process.

[patrickjp93]

TSMC is terrible. Intel (as fanboy as this sounds) is awesome. 

Even so, Samsung's additional failure spells disaster for Global Foundries. If Samsung, which built the process tech, can't master making 3D transistors on it, will Global Foundries really fair much better with any palpable ease?

[LukaP]

Cannonlake is still due for release in summer 2016, so Zen (designed for the 16nm FinFET process) may have even less time to dance with Skylake than AMD would have wanted.

 

AMD has 20nm GPUs coming out right after CES. Nvidia still has Maxwell only on 28. That means they have to do a decent amount of redesigning to get Pascal onto the 20nm process.

Im not sure how much AMD really plans to dance with Skylake (apart from PR stuff). neither am i sure how on time can Intel be with Cannonlake. they have 1.5 years to build and perfect an entirely new process node. thats not an easy task, and there is one more thing. mobile. i can assure you the first thing to get 10nm, and for a long time will be mobile and Atoms. look at broadwell, its been out for a while now, with no real sign of desktop chips.

 

They dont have to do it. do you think they dont know the status of the nodes atleast a few months ahead of the rest of the world. they have probably already done most of the work to bring maxwell to 22/20nm, and will release the 3072 core version on that (as per that thread, i cant imagine that on 28nm). which means, yes the 390x will hold the crown for a bit, maybe even be as power efficient as the GM204, but in a few months, NV will release the Titan2 series (and later 1000) which will again take the crown. and then so on until who knows when, as it has always been.

[samcool55]

I'm not suprised that samsung does't want the contact, even if they could do it.

 

If they can fabricate 16nm stuff properly, they can use it for their exynos stuff and crush everything like they did with the S3, well i hope they will crush everything

So even if they have it, it would be logical if they put the new technology in their own stuff first.

[Guest]

I'm not sure why so many people assumed that Intel's >2 year fab tech lead on the competition had suddenly evaporated. I want competition as much as everyone else but I can't see it being likely that Samsung or TSMC will be able to overcome the technical issues of the lower node in much less time than it took Intel.

[patrickjp93]

Im not sure how much AMD really plans to dance with Skylake (apart from PR stuff). neither am i sure how on time can Intel be with Cannonlake. they have 1.5 years to build and perfect an entirely new process node. thats not an easy task, and there is one more thing. mobile. i can assure you the first thing to get 10nm, and for a long time will be mobile and Atoms. look at broadwell, its been out for a while now, with no real sign of desktop chips.

 

They dont have to do it. do you think they dont know the status of the nodes atleast a few months ahead of the rest of the world. they have probably already done most of the work to bring maxwell to 22/20nm, and will release the 3072 core version on that (as per that thread, i cant imagine that on 28nm). which means, yes the 390x will hold the crown for a bit, maybe even be as power efficient as the GM204, but in a few months, NV will release the Titan2 series (and later 1000) which will again take the crown. and then so on until who knows when, as it has always been.

The Quadro? It's a 28nm chip, massive thing too.

 

Also, Intel is skipping EUV for the first 10nm process and going with one more proven for yields but slower in production (something about multi-level patterning).

[patrickjp93]

I'm not suprised that samsung does't want the contact, even if they could do it.

 

If they can fabricate 16nm stuff properly, they can use it for their exynos stuff and crush everything like they did with the S3, well i hope they will crush everything

So even if they have it, it would be logical if they put the new technology in their own stuff first.

It's not a matter of samsung not wanting to. It failed to produce the chips in meaningful yields per wafer. Samsung got contracts for its processes as soon as they were online. So far Qualcomm and Apple both cannot get above a 30% yield on Samsung's 14/16nm FinFET process. If you think Samsung is fairing much better on its own chips you're out of your mind, because that means Samsung is sinking an already low profit margin business just to spite its competition at the cost of its reputation.

[LukaP]

The Quadro? It's a 28nm chip, massive thing too.

 

Also, Intel is skipping EUV for the first 10nm process and going with one more proven for yields but slower in production (something about multi-level patterning).

was it confirmed 28? the GM210 for the M6000? link?

 

that doesnt mean they will achieve great yields no problem tho and MLP (not the friendship is magic one) has a higher BOM cost as well as longer production time, so im not sure how keen intel is on using it

[patrickjp93]

was it confirmed 28? the GM210 for the M6000? link?

 

that doesnt mean they will achieve great yields no problem tho and MLP (not the friendship is magic one) has a higher BOM cost as well as longer production time, so im not sure how keen intel is on using it

GM 200 and yes you can find the process information on Nvidia's website. Also, confirmed by TechPowerUp 551 sq. mm die.

 

Krzanich and others said they have to use it because EUV isn't ready for their production schedule. Yes, it has higher upfront costs and longer production times, but if the yields are sufficiently high enough for it then the time to market could overall be unaffected or even be shortened vs. what could have been with EUV.

 

Edit: we will see EUV in the Israel fab, used for Cannonlake's successor. Intel already is refitting an older 40nm fab plant with 10nm technology.

[LukaP]

GM 200 and yes you can find the process information on Nvidia's website. Also, confirmed by TechPowerUp 551 sq. mm die.

 

Krzanich and others said they have to use it because EUV isn't ready for their production schedule. Yes, it has higher upfront costs and longer production times, but if the yields are sufficiently high enough for it then the time to market could overall be unaffected or even be shortened vs. what could have been with EUV.

where, all i can see is the K6000, and the article doesnt confirm jack shit. they could have samples of 22nm chips, or even full on production (if AMD can have it, so can NV)

 

I know, but i still dont see them building and perfecting it in time for summer 16

[Metrize]

I hope Intel don't partner with Apple. I'd rather see Apple slowly die, they smother innovation.

[patrickjp93]

where, all i can see is the K6000, and the article doesnt confirm jack shit. they could have samples of 22nm chips, or even full on production (if AMD can have it, so can NV)

 

I know, but i still dont see them building and perfecting it in time for summer 16

Can't be 22nm. the tapeout of GM 200 was before the 22nm process came online. http://www.guru3d.com/index.php?ct=news&action=file&id=7515

[patrickjp93]

I hope Intel don't partner with Apple. I'd rather see Apple slowly die, they smother innovation.

Apple doesn't smother innovation at all. It combines existing inventions in their highest quality to deliver an end product a cut or two above the rest. That said, they need 3 USB ports on their MBP line. Getting an expensive Thunderbolt dongle for 4 ports when I only need to use 1 extra is obnoxious.

[Metrize]

Apple doesn't smother innovation at all. It combines existing inventions in their highest quality to deliver an end product a cut or two above the rest. That said, they need 3 USB ports on their MBP line. Getting an expensive Thunderbolt dongle for 4 ports when I only need to use 1 extra is obnoxious.

 

It seems like that but they really they don't change much and slightly improve specs, like the iPhones, which are still stuck on dual cores, if they actually wanted to innovate I am sure that they could make a hexcore or octacore with all the money they have

[LukaP]

Can't be 22nm. the tapeout of GM 200 was before the 22nm process came online. http://www.guru3d.com/index.php?ct=news&action=file&id=7515

that link wants to download a pic for some reason

[LukaP]

It seems like that but they really they don't change much and slightly improve specs, like the iPhones, which are still stuck on dual cores, if they actually wanted to innovate I am sure that they could make a hexcore or octacore with all the money they have

the fact is, that A8 has a higher IPC than any other CPU (apart from Denver) and that 99% of mobile apps never utilise more than 2 threads anyway, so its useless to implement a cpu with 8 threads slower than a potato, when you can have one that gets utilised much better, with 2 threads with almost the speed of sandy bridge

[patrickjp93]

that link wants to download a pic for some reason

It does. That's the graph from TSMC's production schedule, courtesy of Guru3D

[patrickjp93]

It seems like that but they really they don't change much and slightly improve specs, like the iPhones, which are still stuck on dual cores, if they actually wanted to innovate I am sure that they could make a hexcore or octacore with all the money they have

Your forget all the dedicated hardware on the Apple chips. Also, the performance of those dual cores continues to kill the Qualcomm sanpdragon and exynos quads.

[patrickjp93]

the fact is, that A8 has a higher IPC than any other CPU (apart from Denver) and that 99% of mobile apps never utilise more than 2 threads anyway, so its useless to implement a cpu with 8 threads slower than a potato, when you can have one that gets utilised much better, with 2 threads with almost the speed of sandy bridge

Um, did you hit your head on something? Cyclone is still miles behind sandy Bridge. The performance scaling is not nearly as high as you think, and all that dedicated hardware is responsible for most of the SOC performance, not the CPU architecture itself. In reality Cyclone is equivalent to Kentsfield.

[LukaP]

It does. That's the graph from TSMC's production schedule, courtesy of Guru3D

no thats a few shipments of GM200 (not 210, as it is in the M6000) the last one being A1. that doesnt rule out them pushing the B1 (or A2) onto 20nm

[patrickjp93]

no thats a few shipments of GM200 (not 210, as it is in the M6000) the last one being A1. that doesnt rule out them pushing the B1 (or A2) onto 20nm

Yes it does. There's not enough time there to redesign the architecture for two different processes.  Production begins about 6 months before shipping. There's no possible way for GM 200 to be designed for the 20/22nm processes. And the Quadro is GM 200. 

[LukaP]

Um, did you hit your head on something? Cyclone is still miles behind sandy Bridge. The performance scaling is not nearly as high as you think, and all that dedicated hardware is responsible for most of the SOC performance, not the CPU architecture itself. In reality Cyclone is equivalent to Kentsfield.

im just calling it based on nvidia stating Denver to be as good as last gen desktop chips (and overestimating, so i called sandy, not ivy) and then used that, since cyclone+ is about the IPC of denver AFAIK