Samsung and GLOBALFOUNDRIES Forge Strategic Collaboration to Deliver 14nm FinFET

[piesquared]

I think intel is going to need to cough up a lot more 'contra revenue'. lol

 

[NostaSeronx]

Their 28nm HPP was somewhat late but it has actualy better perfs than TSMC s 28nm and judging by real numbers , not fud or hearsay, it s no worse than Intel s 22nm.

None of the GlobalFoundries Nodes were late, instead customers were late at adopting the node. Probably do to bad marketing and bad cost management on part of GlobalFoundires.

That is NOT a marketing failure, but an engineering failure. GLF could not sell what they had developed because it sucked, not because the sales team didn't do its job.

Globalfoundries was stuck between a rock and a hard place with AMD and IBM incentivizing 32nm SOI. AMD and IBM being the only two secured customers for GlobalFoundires, GF gave in.

28nm HP and 32nm SOI, mass production: 2H 2010.

The whole 28nm volume was pushed towards Fab 8, then IBM wanted Fab 8 for its 32nm SOI processor. 28nm volume then dropped like a rock.

GlobalFoundries made the mistake of prioritizing 32nm SOI then going on to say 28nm yields were bad. When they meant to say was 32nm SOI yields were bad.

 

[Hitman928]

It was a lack of marketing really;
28nm-HP was capable of 100% parametric yields in 2010. So, whoever got those Cortex-A9 chips they were the most perfect chips ever made.

- No defects
- No variation
- No yield issues

A modern process with zero variation does not exist, even for small arm chips.

 

[Abwx]

Now that GloFo will be adapting Samsung's 14 nm, it also means that AMD can release 14 nm chips at a time quite close to Intel (perhaps only ~1-1.5 years after).

What is published this day was started long ago , such agreements are not signed overnight , there has been early info that AMD was fully aware from their waffer supplier changing roadmap and that they had this new process caracteristics modeling (transistors Spice modeling and substrat parameters ) used in their future designs.

Recent info suggest that they ll likely use TSMC s 20nm for GPUs and GF licenced FinFet process for APUs/CPUs.

“....we are transitioning to both 20nm and to FinFETs over the next couple of quarters in terms of designs. So we will continue to do that across our foundry partners. […] We will do 20nm first and then we will go to FinFETs,” said Lisa Su, senior vice president and general manager of global business units at AMD

http://www.xbitlabs.com/news/other/...nm_FinFET_Chips_Within_Next_Two_Quarters.html

 

[NostaSeronx]

A modern process with zero variation does not exist, even for small arm chips.

No variation as in pretty normal negligible variation. For what came out it was a non-problem. No defects was from no metal defects within the BEOL and FEOL. The yield ratios were near perfect to perfect with lower variation. The lower variation was a big surprise for GF, IBM, and STMicro.

All the massive engineering papers for 28nm(G(HP) and LP(SLP)) were all with very high praise. It is weird that AMD and IBM instead of following the advice of GlobalFoundries to skip 32nm. Decided to pay extra and get 32nm SOI anyway.

 

[Hitman928]

No variation as in pretty normal negligible variation. For what came out it was a non-problem.

That's very different from what you originally said and even this term, "normal negligible variation," is very ambiguous and incomplete. To be honest, I'm really not sure what you're trying to say about the process. Let me put it this way, if I were a designer with a company trying to decide if this was a process we wanted to pursue, what could you tell me about the process variation/yield?

 

[bullzz]

@piesquared - considering intel is making a profit after all that contra revenue, i think they are fine

 

[NostaSeronx]

I'm really not sure what you're trying to say about the process.

I'm talking about Ion and Ioff variation. For which ever design you go for High Speed or Low Power. You wouldn't have much variation between the high speed devices. While also if you went for low power you wouldn't have as much variation between the low power devices. When comparing to 32nm SOI and TSMC's 28nm node.
---
Lower Variation = Higher Parametric Yields, for what ever design you want.

Yields in generally were very high not relating to parametric yields.

 

[erunion]

Dresden and Singapore foundries from GlobalFoundries are going for 14nm FDSOI. This will be announced later this year. So, you have GlobalFoundries off-sourcing;

Samsung's 14nm FinFETs
and
Samsung's 14nm FDSOI.

No FDSOI. Finfets are FD by themselves, FDSOI wafers would be redundant.

 

[NostaSeronx]

No FDSOI. Finfets are FD by themselves, FDSOI wafers would be redundant.

FinFETs and FDSOI are two different things. They aren't redundant to each other and can be produced equally.

 

[lefty2]

Hmm. Gloalfoundries 14nm FinFET is "up to 20 percent higher speed, 35 percent less power and 15 percent area scaling over industry 20nm planar technology"
Meanwhile, TSMC 16nm FinFET plus is 40% faster than TSMC 20nm planar (mentioned in latest earnings call )

 

[Hitman928]

FinFETs and FDSOI are two different things. They aren't redundant to each other and can be produced equally.

While we refer to planar fully depleted transistors as FDSOI, erunion is correct, Finfet's are fully depleted as well. They do, obviously, have different characteristics and could potentially serve different markets, it doesn't look very good for FDSOI right now though, maybe that will change, but it doesn't look that way currently.

 

[Hitman928]

I'm talking about Ion and Ioff variation. For which ever design you go for High Speed or Low Power. You wouldn't have much variation between the high speed devices. While also if you went for low power you wouldn't have as much variation between the low power devices. When comparing to 32nm SOI and TSMC's 28nm node.
---
Lower Variation = Higher Parametric Yields, for what ever design you want.

Yields in generally were very high not relating to parametric yields.

OK, to put a direct question, have you seen relevant data that shows the performance of GF's 28nm process or is this more of something you've gathered across internet rumblings / second hand info?

 

[NostaSeronx]

OK, to put a direct question, have you seen relevant data that shows the performance of GF's 28nm process or is this more of something you've gathered across internet rumblings / second hand info?

The STMicroelectronics and CEA-Leti comparison of 28nm FD to "28nm-G / 28nm-LP" was to GlobalFoundries Nodes.

28FD = 28nm FDSOI from CEA-Leti.
28nm-G = 28nm HP from GlobalFoundries.
28nm-LP = 28nm SLP from GlobalFoundries.

In the 2011/2012 presentations. You can extract inexact performance from the comparisons.
In the 2013/2014 presentations. The node comparison was to 28nm-HPP and 28nm-LPS from GloFo.

Measuring the growth in performance from 28nm FD between the two generations of G and LP. You can factor that inexact information to find out 28nm-HP was a great but costly node.

28nm-HP would have been ~>1.5x the cost of 28nm-FD, if both nodes were launched within six months of each other in 2010/2011.
Right now, 28nm-HPP is only ~1.2x the cost of 28nm-FD.

 

[erunion]

FinFETs and FDSOI are two different things. They aren't redundant to each other and can be produced equally.

Maybe I misunderstood you, I assumed you meant some finfets would be on FDSOI and some not.

Your point was that Samsung would produce planar and finfet transistors both called 14nm? That's amazingly unlikely from a marketing standpoint.

 

[NostaSeronx]

Maybe I misunderstood you, I assumed you meant some finfets would be on FDSOI and some not.

FinSOI at the 14nm node so far is only being looked at by IBM. CEA-Leti is waiting for that stepping stone at 7nm.

Your point was that Samsung would produce planar and finfet transistors both called 14nm? That's amazingly unlikely from a marketing standpoint.

No, Samsung's nodes are;
14nm-FD => FDSOI
14nm-XM/LPE => Low Power Enhanced, relative to 20nm LPM "Low Power Mobility/Low Power Manufacturing"
14nm-LPP => Low Power Performance. The "G" node.

Of those being shared;
14nm-FD and 14nm-XM.

 

[mrmt]

None of the GlobalFoundries Nodes were late, instead customers were late at adopting the node. Probably do to bad marketing and bad cost management on part of GlobalFoundires.Globalfoundries was stuck between a rock and a hard place with AMD and IBM incentivizing 32nm SOI. AMD and IBM being the only two secured customers for GlobalFoundires, GF gave in.

28nm HP and 32nm SOI, mass production: 2H 2010.

Maybe you could explain to us how GLF developed 28nm in 2010, less than one year. Because 28nm wasn't in AMD's plans, they were always to follow IBM with 32nm and then 22nm.

There isn't the slightest chance of GLF developing a node to a productive state in less than a year after finishing the spin off from AMD. Show your sources, show your links, cut the crap.

 

[NostaSeronx]

http://www-03.ibm.com/press/us/en/pressrelease/27222.wss
http://tof.canardpc.com/view/59183502-956c-44f2-aaa8-5f7add29e397.jpg

GlobalFoundries had an internal node 28nm HP, while it shared 28nm SLP with everyone else.

Edit: http://www.xbitlabs.com/news/other/...ther_28nm_Fabrication_Process_to_Roadmap.html
Edit2: http://www.xbitlabs.com/news/graphi...d_28nm_Graphics_Chips_at_Globalfoundries.html

 

[mrmt]

http://www-03.ibm.com/press/us/en/pressrelease/27222.wss
http://tof.canardpc.com/view/59183502-956c-44f2-aaa8-5f7add29e397.jpg

GlobalFoundries had an internal node 28nm HP, while it shared 28nm SLP with everyone else.

That's the node the common platform was developing together, you know, the one that Samsung could only put together for a product in 2013. How could GLF have this node ready in 2010, three years before Samsung? You know, that's the node one AMD was supposed to use for Krishna and Wichita, the chips crapped because Globalfoundries didn't have the node ready.

And given that a foundry takes 4-5 years to develop a node, that press release of yours fits perfectly the time frame for Samsung to launch a product in 2013 with the node they started to develop in 2009, but it wouldn't fit the schedule needed for Globalfoundries to have the node ready by 2010, because that would imply that they would have started R&D around 2006, and we know they didn't because at the time there was no Globalfoundries to talk about, just AMD, and their policy was to license IBM high performance nodes and be done with it.

I find amusing the notion that someone actually believe these Globalfoundries fairy tales, especially in the month we see Globalfoundries trying to acquire IBM R&D team, and in the very day GLF announced to the world that they were flushing their R&D pipeline and buying someone else's process. Oh, the irony!

 

[NostaSeronx]

Just to point out the ATR roadmap is full on volume production. GlobalFoundries had 28nm before pretty much anyone else.

People keep thinking customers need to wait for volume, they don't customers can start at risk.

In regards with AMD for their early 28nm; CPUs, APUs, and GPUs. It was AMD's fault for the lack of such designs. At the time it was the interim-CEO's decision to save money from R&D and delay all projects.

in the very day GLF announced to the world that they were flushing their R&D pipeline and buying someone else's process.

14nm-XM prototyping started in Q1 2013. March 2014 was the CPA risk production deadline for 14nm-XM.

 

[Idontcare]

That's the node the common platform was developing together, you know, the one that Samsung could only put together for a product in 2013. How could GLF have this node ready in 2010, three years before Samsung? You know, that's the node one AMD was supposed to use for Krishna and Wichita, the chips crapped because Globalfoundries didn't have the node ready.

And given that a foundry takes 4-5 years to develop a node, that press release of yours fits perfectly the time frame for Samsung to launch a product in 2013 with the node they started to develop in 2009, but it wouldn't fit the schedule needed for Globalfoundries to have the node ready by 2010, because that would imply that they would have started R&D around 2006, and we know they didn't because at the time there was no Globalfoundries to talk about, just AMD, and their policy was to license IBM high performance nodes and be done with it.

I find amusing the notion that someone actually believe these Globalfoundries fairy tales, especially in the month we see Globalfoundries trying to acquire IBM R&D team, and in the very day GLF announced to the world that they were flushing their R&D pipeline and buying someone else's process. Oh, the irony!

That's OK, Intel has had 10nm fully developed (complete with 100% parametric yields as well) since 1995, but customer demand was lacking so they just kinda shelved it for the time being and instead decided to roll out the lower performing and lower yielding stuff since that was where the demand was at.

Its true, I read it on the internet.

Anyone who can claim with a straight face that GloFo had a high-performance 28nm node in 2010 but AMD walked from it and didn't use it is just being silly.

 

[AtenRa]

Anyone who can claim with a straight face that GloFo had a high-performance 28nm node in 2010 but AMD walked from it and didn't use it is just being silly.

IDC, the IBM Press Release is talking about Low-Power 28nm HKMG process not High-Performance. The same press release is talking about risk-production in 2H 2010. That is the same time-line early volume production of 32nm HKMG SOI was started that lead to AMD Llano official release in H1 2011(Shipping Q1 2011). Since 28nm is Half-Node of 32nm it is easy to understand that 28nm HKMG Low-Power was ready for volume production around H2 2011 and H1 2012. That is also the same time-frame of TSMCs 28nm Volume production.
The problem was not if the process was ready or not, the problem with IBM (and the rest of the alliance FABs) was they had customers with 32nm(Samsung also used 32nm themselves) and GloFos 32nm SOI for AMD.
32nm At the time was starting volume production and and it was occupying the bulk of those Fabs production capacity.
TSMC had 28nm HP earlier than IBM/GloFo/Samsung and that two are the reasons most of the companies went to TSMC.

It is like today's situation with 20nm/16nm. Companies went from 40nm to 28nm because 28nm was better than 32nm. Today companies are going from 28nm to 16nm FinFets because the jump in performance and especially power is higher than to go from 28nm to 20nm.
But as i have said before, Today things are also different for GloFo and Samsung because they are only the two of them now that make the decisions, they dont have to follow IBM any more. Both GloFo and Samsung have the same process needs for their customers so they both need to have a competitive process against TSMC and also GloFo has a new Fab in New York that is capable of doubling its capacity. Now Glofo can continue producing 28nm in Germany and have enough capacity for 14nm in New York to satisfy their customers needs.

ps: IBMs 28nm is Half-Node of its 32nm process. TSMCs 28nm is a Full-Node process over its 40nm.

 

[SammichPG]

Why do you hate Samsung?

He hates anyone who competes with Intel, simple as that.

 

[NTMBK]

That's OK, Intel has had 10nm fully developed (complete with 100% parametric yields as well) since 1995, but customer demand was lacking so they just kinda shelved it for the time being and instead decided to roll out the lower performing and lower yielding stuff since that was where the demand was at.

Its true, I read it on the internet.

Anyone who can claim with a straight face that GloFo had a high-performance 28nm node in 2010 but AMD walked from it and didn't use it is just being silly.

Good to have you back, idc

 

[shady28]

What do you mean by "just late"? In the semiconductor industry time frame is everything. Take Bulldozer, it was a disaster when it was launched in 2011, but it would be a nice processor if AMD could get that level of performance in 2008.

If AMD had released Bulldozer in 2008, it would have not only crushed the C2D line that was current at the time it would have also crushed the soon to be released Nehalem core series. AMDs FX line got one-upped by Sandy bridge in early 2011, and Phenom was one- upped by Nehalem.