Samsung and GLOBALFOUNDRIES Forge Strategic Collaboration to Deliver 14nm FinFET

[raghu78]

Glofo is in chains as well, and it seams the one customer you are talking about only contributed 23% of the total sales in 2013. It just shows that GloFo has more customers than people here were made to believe.

only 23%. frankly for a fab company like GF, a single customer at close to 1/4th the total revenues is a lot of risk if that customer were to go to a competing fab. But luckily AMD is tied to GF with a WSA which runs till 2024. if AMD were to move completely to TSMC that would devastate GF. It would also help AMD as TSMC are executing much better at the leading edge process node. GF licensed Samsung's 14 finfet because that's the only way they were to get to 14 FINFET. their 14XM was just marketing and ppt slides. GF's execution at 28nm bulk was horrible. In fact Kaveri's delays are directly linked to the GF 28nm SHP being delayed. Luckily Samsung licensed their 14 FINFET and thus GF now has a chance to remain a viable foundry.

 

[mrmt]

Luckily Samsung licensed their 14 FINFET and thus GF now has a chance to remain a viable foundry.

Since when posting losses of about 25% of your revenue is symptom of viability?

 

[krumme]

Glofo is in chains as well, and it seams the one customer you are talking about only contributed 23% of the total sales in 2013. It just shows that GloFo has more customers than people here were made to believe.

Well except with the addition AMD is the only one not paying the market price for the product they get as evident from the loss gf generates.

Seriously would you make a kaveri on gf? Its a expensive mess taking forewer to execute with gcn beeing tuned for tsmc. Look at the rest of the product produced at gf. Second class former generation design.

Forcing an advanced product like kaveri into that production technology is just plain stupid business on the long run. Shows Mubadala fumdamental lack of understanding what they have in their hands. Its only a political pipe dream only possible with flowing oil.

This is the same story i wrote 3 years ago. Mubadala have executed perfectly.

I think there is an end to it. As Intel gigantic mobile loss have to end. 2 years and we have radical new business ownership structure imho. I dont beliewe aliances, partnering and licensing can cut it.

 

[Idontcare]

Before IBM made it clear they are planing to sell their fab business, people were saying that Samsung and GloFo cant create their own process because they luck the knowledge and R&D.
Now Samsung creates the 20nm process in collaboration with GloFo(they are after all in the Common Platform alliance) and sudently people react like GloFo just entered the Fab business and they know shit.
Not only that, but because GloFo licensing 20nm from Samsung, suddenly people believe GloFo will start from scratch in 2014 to develop 10nm. I wonder why they joint the Common Platform Alliance in the first place


http://www.commonplatform.com/index.html
You do see the Common Platform badge in the upper right hand corner. Nah, they just put it there because it looks nice with the rest of the colors of the slide

The 14nm process that GF is licensing from Samsung did not come from the common platform alliance, that is why GF has to license it from Samsung rather than using it as freely as Samsung gets to use it.

The 14nm-XM process that was a product of the common alliance is so badly broken that it is being shelved, to be replaced by Samsung's internally developed process tech.

I thought this was common knowledge (not being facetious) but maybe it isn't - Samsung maintains a fully internal process node development team that develops process nodes in parallel to the R&D team they field in NY for the IBM fab consortia. Samsung joined the consortia simply to hedge its bets and ensure that it would always have a choice of what it would put into production - be it the homegrown node or the collaboratively developed one involving IBM and the common alliance.

The dots are there if people want to connect them, you don't have to know anyone with industry connections to spell it all out. Look at the timeline of when the first rumors went up for IBM looking to get out of the fab business, the timing of when GF announced they were licensing Samsung's independently developed 14nm process while shelving their own common-alliance developed 14nm-XM, and the timeline of how far back those licensing discussions between Samsung and GF must have been initiated.

The common alliance platform died some 6-9 months ago, but no one in the public domain need know about it at the time while the decision makers at the big three looked to take action so that they had solutions in place (the GF licensing deal) when they finally came to the public with the info.

The stop-gap to replace it has just been announced, and the end-game announcement is probably another 12-15 months away.

 

[NostaSeronx]

The CPA is much still alive; the Samsung 14nm FinFETs are still based on the IBM's 20nm FinFETs.

You also have official and unofficial members of the CPA;
Samsung
GlobalFoundries
UMC
Renesas
Toshiba
CEA-Leti (STMicroelectronics / GlobalFoundries)
Infineon
Avago

All controlled by IBM R&D spooks.

 

[erunion]

IDC, is this 14nm licensing the cause of your return to the forum? Or were we just lucky to have you back during this major event?

 

[Idontcare]

IDC, is this 14nm licensing the cause of your return to the forum? Or were we just lucky to have you back during this major event?

Just a coincidence. Happen to have two weeks with an unusually open schedule, and at the same time some interesting news finally makes it to the public domain.

 

[MisterMac]

Just a coincidence. Happen to have two weeks with an unusually open schedule, and at the same time some interesting news finally makes it to the public domain.

Hopefully you'll have less hectic schedules in the future - it's good too see your view again

 

[NTMBK]

Thanks for the insight, IDC Quick question- what do you make of the IBM fab sell off? Who do you think might end up with them? (Latest rumours said that it was between GloFo and Intel, with TSMC out of the race.)

 

[NostaSeronx]

I have done an in depth analysis of what occurred.

The Common Platform in 2012 releases information about the 14nm FinFET node. GlobalFoundries called their version 14nm-XM and Samsung called their version 14nm-LPE. In 2013, Samsung changed the design rules and found that the node could achieve higher density/performance for the same cost. In secret during the later 2013 half, the CPA decided to follow suit with Samsung's changes. Even though the 14nm-XM node has been cancelled. The 14nm-LPE/LPP node are still design compatible with 14nm-XM. No one prototyping with 14nm-XM has lost their IP.

This comes with a delay though;
14nm-XM's volume production would have started in Q2 2014. With customers releasing production or engineering devices in Q3 2014.
14nm-LPE/LPP collaboration volume production starts in Q4 2014. With customers releasing production or engineering devices in Q1 2015.

There could be risk production devices but with FinFETs you have a slower engineering to production timeline. Customers would devote all risk availability in getting the device to production ready.

 

[Idontcare]

Thanks for the insight, IDC Quick question- what do you make of the IBM fab sell off? Who do you think might end up with them? (Latest rumours said that it was between GloFo and Intel, with TSMC out of the race.)

All bets are off as to who will end up with them, but it is safe to rule out those companies that simply don't have any need for them.

That would be TSMC who thus far has out-trumped, out-R&D'ed, and out-maneuvered IBM and its entire fab consortia at every stage of the game.

So, the question really is "who most desperately needs IBM, and is IBM willing to sell them the specific parts of IBM that are most needed by the desperate party in question?"

At the moment that would appear to be GF, but can they get what they need from IBM, or will IBM prevent their crown jewels (IP) from being part of the deal?

 

[tarlinian]

All bets are off as to who will end up with them, but it is safe to rule out those companies that simply don't have any need for them.

That would be TSMC who thus far has out-trumped, out-R&D'ed, and out-maneuvered IBM and its entire fab consortia at every stage of the game.

So, the question really is "who most desperately needs IBM, and is IBM willing to sell them the specific parts of IBM that are most needed by the desperate party in question?"

At the moment that would appear to be GF, but can they get what they need from IBM, or will IBM prevent their crown jewels (IP) from being part of the deal?

The difference between common platform and TSMC has mainly been one of approach. TSMC has slavishly copied Intel process integration schemes for quite some time. See HKMG, which is RMG with High-K first at 28 nm, just like Intel 45 nm, the use of eSiGe for S/D uniaxial stressors instead of biaxial strain with a SiGe cap on the PMOS channel. IBM et al., have tried to come up with their own ideas which usually have not done so well or been very useful in the long term (e.g., SOI).

 

[Ajay]

[redacted]

 

[Fjodor2001]

Samsung joined the consortia simply to hedge its bets and ensure that it would always have a choice of what it would put into production - be it the homegrown node or the collaboratively developed one involving IBM and the common alliance.

Interesting. I've heard from a colleague that Samsung uses a similar policy when developing mobile phones. They have two totally separate teams that each develop the next version of a phone, and then they choose the product that turns out the best and scrap the other. Cannot say for 100% certain that this is for sure though, since I heard it through a second source.

 

[NostaSeronx]

I just noticed the marketing for the 14-nm LPE node from GlobalFoundries and Samsung. Is the exact same for 14-nm XM node from just GlobalFoundries. It completely confuses me why people would think this collaboration didn't exist before now. Since the first articles for 14nm FinFETs stated that 14-nm XM conformed to Samsung's FinFETs already. 14-nm XM was never cancelled but simply name changed to 14-nm LPE to have synergy with Samsung.

GlobalFoundries 14-nm XM

Samsung & GlobalFoundries 14-nm LPE

----
20-nm LPM -> 14-nm LPE
20% Performance Increase + 35% Power Decrease
- Various additive costs not included

20-nm LPM -> 14-nm FDSOI
20% Performance Increase + 30% Power Decrease
- Various removal of costs not included, no FBB or RBB optimizations included either.

I totally don't see any reason for any fabless semiconductor to move to 14nm LPE. Unless, they get a huge paycheck from sugar daddy Samsung.

On the AMD side, it is kind of enjoyable to see the internal strife on which nodes to use;
India R&D going for 14nm FDSOI high performance cells and the USA R&D going for 14nm LPE high performance cells.

 

[NTMBK]

Hah, good spot on the marketing slides What dates were those shown on? Was it long before the announcement?

 

[mrmt]

Hah, good spot on the marketing slides What dates were those shown on? Was it long before the announcement?

Q: What's the difference between Santa Claus and GLF?

R: Once a year Santa Claus might deliver what his marketing department told you he would.

 

[NostaSeronx]

Hah, good spot on the marketing slides What dates were those shown on? Was it long before the announcement?

For 14-nm Class FinFETs; Late 2012

For the 14-nm XM; November 2013

For the 14-nm LPE; April 2014/May 2014
http://globalfoundries.com/docs/default-source/PDF/samsung-globalfoundries-14nm-collaboration.pdf

https://semiaccurate.com/2012/03/19/ibm-shows-off-14nm-wafer/
https://semiaccurate.com/2012/03/20/samsung-shows-14nm-and-20nm-wafers/
https://semiaccurate.com/2013/01/10/global-foundries-shows-off-20nm-and-14nm-wafers/
https://semiaccurate.com/2013/05/28/samsung-talks-about-their-14nm-finfet-process/

It seems total hilarity realizing this is just a Common Platform thing. Everything GlobalFoundries said for 20nm LPM and 14nm FinFETs were told by Samsung and IBM earlier in 2012. Nice job marketing squads making us think 20nm LPM by GlobalFoundries wasn't 20nm LPP by Samsung.

It's quite clear that GlobalFoundries drove off the cliff not using their own INTELLECTUAL PROPERTY and PATENTS for FDSOI w/ FUSI. Nice job jumping off a proven boat GlobalFoundries. You are simply riding on the hope that your marketing is good. To hide the fact that your nodes you are producing are not your own.

GlobalFoundries CEO/Board; Lets clearly drop our superior FDSOI technology and adopt more expense and more complex Bulk processes.
GlobalFoundries Workers; What?
GlobalFoundries CEO/Board; They are questioning our authority give out the pink slips.

All that research from 1999 to 2007 on FUSI FDSOI completely wasted. For some odd reason they then decided to adopt CEA-Leti's FDSOI. Which is slightly more expensive than their own FUSI FDSOI property.

It's really funny when AMD's CEO went to GlobalFoundries then completely forgot his objective being at GlobalFoundries. Which was to get money and R&D teams to get AMD the FUSI FDSOI node they wanted in 2004-2006. What really throws sand on the wounds is that of the last FDSOI document. Made by AMD's Foundry was only five months before the spin off.

GlobalFoundries with its superior FDSOI IP and experience mind all of you. Could have hit off the SOI explosion way before the CEA-Leti one did.

 

[Idontcare]

Hah, good spot on the marketing slides What dates were those shown on? Was it long before the announcement?

There is unwritten history here, of course, as the public domain only sees the tip of the iceberg.

That said, in truth the marketing slides tell absolutely nothing in regards to similarities or dissimilarities between GloFo's 14XM and Samsung's 14nm.

The reality is that 14XM was nothing like Samsung's 14nm, completely different PDK.

Samsung was ready to let GF die on a vine but Apple made them an offer they couldn't refuse.

There is a such an interesting situation with Samsung to tell, but I dare not tell the tale in public because its just the kind of stuff that gets people fired.

Suffice to say there is no conspiracy going on, GF really did their absolute best to work with IBM to develop 14XM, but Samsung's independently developed 14nm (the one developed in Korea, not the common alliance one developed in NY, Samsung has two independent R&D teams for process nodes) was the only thing that has a ray of hope of yielding and delivering the required reliability in time to be commercially relevant to the customers who are currently being courted for TSMC's 16FF+.

 

[Idontcare]

All that research from 1999 to 2007 on FUSI FDSOI completely wasted. For some odd reason they then decided to adopt CEA-Leti's FDSOI. Which is slightly more expensive than their own FUSI FDSOI property.

FUSI turned out to be a dead-end. It kills the xtor reliability. If you research the technical journal publications you'll notice pretty much everyone stopped publishing FUSI efforts around 2008, and for a good reason. We all realized there was an intrinsic problem with FUSI and killed it.

Not everyone reached the point of having enough data to come to the that conclusion at the same time though, so a few companies persisted another year or so, but not much beyond that.

 

[Homeles]

Looks like I was right; FinFETs inherently suck at > ~1.2V, but are great for mobile -- now it's confirmed from another foundry, and not just an Intel thing.

This puts AMD in an interesting situation -- 32nm PDSOI clocked well, 28nm bulk sucks, 20nm bulk would be great (if a suitable HP process existed), but 14nm is probably going to suck for overclocking.

 

[Idontcare]

Looks like I was right; FinFETs inherently suck at > ~1.2V, but are great for mobile -- now it's confirmed from another foundry, and not just an Intel thing.

Just understand that you are right only because that is precisely the "pocket" that the process and electrical engineers have intentionally designed the finfets to fit into for commericial viability for the products being targeted by the design engineers.

The cause and effect here is not fully disclosed. If people wanted Finfet to deliver at silly high voltages (relative to the process node label) with crazy high drive currents and off currents then they could do just that.

These nodes, Intel's and everyone else, are specifically being designed and optimized (for years) at the low-leakage targets necessary for the mobile markets because that is the very market that commands high volume and high ASP.

There is nothing intrinsic to a finfet that relegates it to the mobile low-power realm. But if you want to go there with your finfet then you do so at the trade-off of giving up the high-power realm.

Finfets are not a "have your cake and eat it too" revolution. They are very much a "go where planar can't, if that is what you so desire" evolution.

But make no mistake, any process flow manager could elect to drive their finfet development team towards the development of a finfet that not only rivals, but bests, the high power capability of a planar CMOS xtor.

That has always been an option, but the market TAM for such a process node simply doesn't justify the R&D at this time. Very few people want 300W CPU's. That is the reality of the 21st century.

 

[Arachnotronic]

There is unwritten history here, of course, as the public domain only sees the tip of the iceberg.

That said, in truth the marketing slides tell absolutely nothing in regards to similarities or dissimilarities between GloFo's 14XM and Samsung's 14nm.

The reality is that 14XM was nothing like Samsung's 14nm, completely different PDK.

Samsung was ready to let GF die on a vine but Apple made them an offer they couldn't refuse.

There is a such an interesting situation with Samsung to tell, but I dare not tell the tale in public because its just the kind of stuff that gets people fired.

Suffice to say there is no conspiracy going on, GF really did their absolute best to work with IBM to develop 14XM, but Samsung's independently developed 14nm (the one developed in Korea, not the common alliance one developed in NY, Samsung has two independent R&D teams for process nodes) was the only thing that has a ray of hope of yielding and delivering the required reliability in time to be commercially relevant to the customers who are currently being courted for TSMC's 16FF+.

So...does this mean TSMC isn't going to be building Apple's chips at the 14/16nm generation?

Also, when do you expect products based on Samsung's 14nm to be in the market?

Thanks!

 

[witeken]

Thank you for your very interesting comments, Idontcare.

 

[Homeles]

Just understand that you are right only because that is precisely the "pocket" that the process and electrical engineers have intentionally designed the finfets to fit into for commericial viability for the products being targeted by the design engineers.

The cause and effect here is not fully disclosed. If people wanted Finfet to deliver at silly high voltages (relative to the process node label) with crazy high drive currents and off currents then they could do just that.

These nodes, Intel's and everyone else, are specifically being designed and optimized (for years) at the low-leakage targets necessary for the mobile markets because that is the very market that commands high volume and high ASP.

There is nothing intrinsic to a finfet that relegates it to the mobile low-power realm. But if you want to go there with your finfet then you do so at the trade-off of giving up the high-power realm.

Finfets are not a "have your cake and eat it too" revolution. They are very much a "go where planar can't, if that is what you so desire" evolution.

But make no mistake, any process flow manager could elect to drive their finfet development team towards the development of a finfet that not only rivals, but bests, the high power capability of a planar CMOS xtor.

That has always been an option, but the market TAM for such a process node simply doesn't justify the R&D at this time. Very few people want 300W CPU's. That is the reality of the 21st century.

I figured that was the case, but if no one ever targets high performance and power, then the result is effectively the same as if it weren't possible to do so in the first place. This is why I'm worried about TFETs and the like, which would only make the issue even more extreme.