[Anandtech]: GlobalFoundries Stops All 7nm Development !!

[Hitman928]

As a consumer i couldn't care less about this WSA. It doesn't affect me why should I care about the financial relationship between AMD and GF.

As a consumer, you shouldn't care. We're just having a business (and slightly technical) discussion about the state of the industry and some of it's major players.

 

[piesquared]

Meh, nm.

 

[Qwertilot]

Presuming they can get the capacity they want anyway. Samsung too to help a bit of course.

If one of TSMC or SS drop out at some point then they might end up getting some stuff made by Intel
(The games consoles presumably absolutely require a second source?)

 

[Topweasel]

Presuming they can get the capacity they want anyway. Samsung too to help a bit of course.

If one of TSMC or SS drop out at some point then they might end up getting some stuff made by Intel
(The games consoles presumably absolutely require a second source?)

I don't think they require a second source. If I remember correctly for both consoles they are licensed IP that Microsoft and Sony then shop for manufacturing. But TSMC has poured more into CPU manufacturing than even Intel has in the last 5 years including a new 22 billion dollar factory (that's 10x what dresden cost AMD). Intel will fold their manufacturing arm before TSMC does.

 

[krumme]

I don't think they require a second source. If I remember correctly for both consoles they are licensed IP that Microsoft and Sony then shop for manufacturing. But TSMC has poured more into CPU manufacturing than even Intel has in the last 5 years including a new 22 billion dollar factory (that's 10x what dresden cost AMD). Intel will fold their manufacturing arm before TSMC does.

This. And its not good. Oil kept gf going way to far. Tsmc ecosystem and business model is far far superior to Intel. Business model always wins over tech and tsmc is no slouch either for process tech.
Samsung is china on wheels with woodscrews.
Intel is old school. But i much prefer that to some samsung future.

 

[NostaSeronx]

From VLSI_Research:

No where does GlobalFoundries state they aren't going to have 7FDX.

There might also be 12FDX+ later in the roadmap as well which would push 7FDX out: (From STMicroelectronics w/ GlobalFoundries Fab 1 as virtually their fab: Green is GF/Blue is Samsung)

With no FinFETs after 14LPP/12LP, while there definitely is a FDSOI after 22FDX. Means a big shift in GlobalFoundries strategy. As if shrinks are somewhat important, there is none after 14LPP/12LP for FinFETs.

If as a customer of 28nm bulk from GlobalFoundries. The only option that is safe and secure is thus the 22FDX -> 12FDX route. There being no technical issues with 7LP means that GF can safely pursue the 7FDX node. Since, FDSOI is two years behind but still leading edge in performance(FD > FF), power(FD > FF), yield(FD > FF) and cost(FD > FF). It is only natural for the 7FDX node not to be in roadmap as of yet. CEA-Leti/Fab 1 isn't done prototyping the 7nm FDSOI node.

12FDX has the same amount of power shrink as 7LP.
14nm -> 7nm = >55%
22FDX is 30% lower power with Cortex A35 than in 14LPP.
12FDX promises 47% lower power than 22FDX and 50% lower power than 14LPP(n16FF)/12LP(n12FF).
https://images.anandtech.com/doci/13171/15335754089771237328638.jpg
^ 7nm FinFET Power Efficiency.
Looking at STM slides:
22FDX is ~10% faster than 14LPP
22FDX+ is ~10% faster than 22FDX at minimum.
12FDX which uses some 10nm FDSOI boosters is >25% faster than 22FDX(actually: 22FDX+, but who cared before 7LP hold).
+10 +10 +25 => >45% so, 12FDX has performance on demand slightly higher than 7LP DUV.

Again, GlobalFoundries has been sandbagging their FDSOI benchmarks. As not to anger the 14LPP/7LP crowd. Except, the 7LP crowd left and half of GlobalFoudnries never went 14LPP. With majority of the new customers and old customers going 22FDX.


GlobalFoundries have three GTCs coming up and we should definitely see some more FDSOI focus in them;
2018 GLOBALFOUNDRIES Technology Conference on September 25, 2018
GLOBALFOUNDRIES Technology Conference on October 12, 2018
3rd GTC on November 1st, 2018

 

[amd6502]

I think 12nm finfet and will be a long lived node for mainstream, as well as for consoles. Pinnacles can be like the new FX piledriver. 7nm is perfect for high end and server (both cpu and gpu), since production cost is not so significant for these markets.

If they have the bulk of the volume of the production on 12nm (and 14nm) they may be good with the wafer agreement. RR is also mainstream and counts toward WSA with GF. A dual core ~100mm2 budget APU on 12nm would add to this mainstream volume. Perhaps they can negotiate with MS and Sony to make console production also count towards this.

The high cost of 7nm makes it seem like the wrong choice (at least over medium term) for middle of the road mainstream or lower (including console).

22fdx makes good sense for budget to mainstream markets if porting excavator+gcn 1.2 or creating a successor from old 28nm designs. I don't know how suitable zen2 would be for porting to 12fdx.... and what about on the gpu side, they would have to either use gcn 1.2 or adapt vega/navi to SOI.

 

[AkulaMD]

But only after selling 22,000,000 AMD shares on 7/30/18 ... http://ir.amd.com/ownership-summary

Nice info, thanks. I can see that most of the other shares holders was actually increasing their position in AMD.

 

[MrTeal]

No I am stating the obvious. That AMD isn't required to maintain wafer purchases in their desire to move to 7nm. We know AMD's financial situation. Paying the fines or suing to get out from under the WSA is not something that they can afford. So either they were released from their requirement or GF was not able to maintain a requirement that AMD placed on them for the agreement and that released them from their purchase requirement (at least as far as 7nm is concerned). That part should be obvious. Because that was pretty much all but confirmed yesterday. Trying to figure out if GF let them out rather than throw money at 7nm or AMD brought up that their desire to forgo 7nm would dissolve their purchase requirement is matter you and others can debate.

I just find it funny "Well the part you are describing is redacted. So since I can't read it I am going to ignore todays (yesterdays) events and proclaim that AMD can't do what they just said they are doing".

Obviously, we don't have the full copy. However I find it hard to believe that there are not terms in there that would allow AMD an out if GF is unable to produce a competitive node or abandons it completely. AMD was bound to produce with GF (or pay) and thus delays in their production would materially affect AMD's bottom line as is stated in their investor guidance. I just really find it a stretch that AMD wouldn't have identified a non-competitive GF as a massive risk factor and would not have structured in an escape clause of some sort. It would be like signing an agreement with RIM in 2010 that your company would buy employees exclusively Blackberries for the next 10 years and not have any verbiage that the penalties don't apply if they stop making new handsets. You would need terrible governance and lawyers to agree to that.

 

[krumme]

Obviously, we don't have the full copy. However I find it hard to believe that there are not terms in there that would allow AMD an out if GF is unable to produce a competitive node or abandons it completely. AMD was bound to produce with GF (or pay) and thus delays in their production would materially affect AMD's bottom line as is stated in their investor guidance. I just really find it a stretch that AMD wouldn't have identified a non-competitive GF as a massive risk factor and would not have structured in an escape clause of some sort. It would be like signing an agreement with RIM in 2010 that your company would buy employees exclusively Blackberries for the next 10 years and not have any verbiage that the penalties don't apply if they stop making new handsets. You would need terrible governance and lawyers to agree to that.

We dont have the details but we know Mabadala have effectively acted as bank for amd until just 2 years ago or so. Cash reserve.
They could put up demands no matter how irrational they were business wise. That have shaped the wsa i guess. Its a miracle they havnt destroyed both companies. The ownership of amd was a way to avoid that and solve the dilemma.

We have to look at it 6 years back. Mubadala would have political pressure to move tech to united arab emirates. There isnt a place on this earth that is less suited than excactly this spot. All qualified personel is imported and wages is sky high. The ones in power dont want to work and is utterly incompettent. Its a cultural and productive dump. Worse than sahara worse than greenland for a fab.

Gf was at that time already doomed for going down for reasons pretty much out of their hands and have executed badly. With the new owners it went from worse to hillarious. Funky unrealistic ppt to please stupid political owners and no execution until samsung came along with 14nm.

Amd had a basic team of meager 100 man working on the zen core. Those cost could easily have been stripped to put lipstuc on the gf cost. Indirectly via the wsa.

 

[NTMBK]

No where does GlobalFoundries state they aren't going to have 7FDX.

There's also nowhere that GlobalFoundries states that they aren't going to start a patisserie, doesn't mean that's going to happen either.

If as a customer of 28nm bulk from GlobalFoundries. The only option that is safe and secure is thus the 22FDX -> 12FDX route. There being no technical issues with 7LP means that GF can safely pursue the 7FDX node. Since, FDSOI is two years behind but still leading edge in performance(FD > FF), power(FD > FF), yield(FD > FF) and cost(FD > FF). It is only natural for the 7FDX node not to be in roadmap as of yet. CEA-Leti/Fab 1 isn't done prototyping the 7nm FDSOI node.

So let me get this straight... the only safe and secure option is to port my entire product line from bulk to SOI, and rely on the existence of a future node which is still in early development and that there is no evidence will ever go into mass production? (And given that they just bailed on 7nm FinFET, that seems very unlikely.) Or alternatively you could go to the foundry with a proven track record of delivering, TSMC.

Again, GlobalFoundries has been sandbagging their FDSOI benchmarks. As not to anger the 14LPP/7LP crowd. Except, the 7LP crowd left and half of GlobalFoudnries never went 14LPP. With majority of the new customers and old customers going 22FDX.

Sure, GloFo- the foundry with a history of wildly optimistic Powerpoints followed by complete failure- have been underselling a product. That's totally in keeping with their behaviour, and definitely isn't entirely at odds with their aim of selling a product to customers. Nope.

 

[Gideon]

Sure, GloFo- the foundry with a history of wildly optimistic Powerpoints followed by complete failure- have been underselling a product. That's totally in keeping with their behaviour, and definitely isn't entirely at odds with their aim of selling a product to customers. Nope.

Not only that. Despite putting all it's money and dev-resources to Zen family, AMD has also been somehow secretly developing Bulldozer derivatives (you know, with all that cash laying around) on the SOI process, that are so-much better than Zen!

All those useless millions developing a new arch from ground-up. Never mind the 7nm TSMC masks, when that will be "bulldozed" on a FD-SOI process

Let's say what the CTO of GlobalFoundries had to say about the subject:

You can get performance when you need it, so on our 22FDX you can easily get 14nm FinFET performance. You can’t get it all the time, but when you need it. With our 12FDX, you can get close to 7nm performance when you need it.

When I say that, I should qualify: you are not going to get that from a big chip. If you are making a big chip, you got tons of wire capacitance, so you are not going to beat a FinFET design. FinFET has huge drive current, so if you're dealing with smaller chips, wiring capacitance isn't as a big, but gate capacitance of the transistor is more important. FD-SOI has less gate capacitance than FinFET because you don't have the gate wrapping around the whole device. So it is ideally suited for smaller chips, I say less than 150 square millimeters.

But what does he know ... I'm sure Nostra will reply with a unreadable wall of text, claiming there is another super-excavator coming that would be faster than a server-room full of Zen 3's on EUV 7nm

 

[darkswordsman17]

Actually you do or you just wilfully ignore the very strong cyclical nature of Apple sales that frees up production capacity at the start of every year, for say, over ten years now.

I'm not sure how you can possibly be arguing that Apple's demand doesn't affect others just because it happens at similar time every year. That you somehow think that not only do I not see that but you choosing to ignore that's exactly what I'm arguing is just baffling. I have to outright laugh that you're saying that I'm willfully ignoring something, when for some reason you're choosing to ignore that Apple had been using Samsung extensively before settling in with TSMC as of late. Of course that doesn't tell the whole story either as TSMC expanded in part to win Apple's production

That's assuming that Apple holds at how they've been operating too. It also assumes no problems arise. What happens if TSMC has an issue and production is delayed for 6 months or more? What happens if Apple has a bit of a setback? What about if the other companies do, so they aren't ready to start production when Apple's tapers off? I think that played a role in Qualcomm's screwup with the 810.

There are finite resources. Wasn't there an issue with sourcing wafers several years back? And there's been a fair amount of talk about some of the new machines needed for the more advanced EUV production when they start moving to that. And while TSMC has expanded expecting significant demand, I doubt they were taking into account that GF would completely abandon 7nm.

Again, that doesn't mean it absolutely will cause problems, but that Apple will always get preferential treatment means that other companies have to work around that. And when it comes to pricing, Apple can more easily handle it, whereas companies like AMD that were heavily invested in GF are now more limited on their ability to negotiate production.

 

[PeterScott]

No where does GlobalFoundries state they aren't going to have 7FDX.

GF 7FDX was never on the roadmap, never will be on the roadmap.

As noted somewhere in your wall of stuff. SDOI lags FinFet by ~2years.

Since GF 7nm FinFet is arriving never. GF 7FDX is arriving 2 years after never.

The whole point of GF FDX designs is cost containment. 22FDX size was chosen, so they would only need single patterning. 12FDX is what they can do with double patterning. To do 7FDX you need Quad Patterning (or EUV) which is exactly the crazy high cost that stopped GF in their tracks at 7nm FinFet.

 

[NostaSeronx]

GF 7FDX was never on the roadmap, never will be on the roadmap.

It's on the Dresden Fab 1 and CEA Tech R&D roadmap. Which the VLSI Research group based it on.

As noted somewhere in your wall of stuff. SDOI lags FinFet by ~2years.

Since GF 7nm FinFet is arriving never. GF 7FDX is arriving 2 years after never.

Originally GlobalFoundries had a FinFET complements FDSOI outlook. Now, since >75% of their base will be 22FDX by 2020. It makes sense for them to speed up 7FDX for the 7nm operational plan.

-> GF is realigning its leading-edge FinFET roadmap to serve the next wave of clients that will adopt the technology in the coming years.
<GlobalFoundries is not realigning its leading-edge FDSOI roadmap. (So far it has implications of speeding up FDX)

-> To support this transition, GF is putting its 7nm FinFET program on hold indefinitely and restructuring its research and development teams to support its enhanced portfolio initiatives.
<GlobalFoundries is retiring FinFET and keeping some of the top FinFET TDs to maintain 14LPP/12LP. While, most will be redistributed to different than everyone else niches.

-> GF is intensifying investment in areas where it has clear differentiation and adds true value for clients, with an emphasis on delivering feature-rich offerings across its portfolio. This includes continued focus on its FDXTM platform, leading RF offerings (including RF SOI and high-performance SiGe), analog/mixed signal, and other technologies designed for a growing number of applications that require low power, real-time connectivity, and on-board intelligence.
<FDX platform is one of the bigger ones out of all of them. So, it is getting majority of the 7LP/5LP/3LP pathfinding budget. Straight up FinFET isn't in this list of intensifying investment for true value. Meaning that in this new GlobalFoundries outlook FinFET has no value for the modern customer.

The whole point of GF FDX designs is cost containment. 22FDX size was chosen, so they would only need single patterning. 12FDX is what they can do with double patterning. To do 7FDX you need Quad Patterning (or EUV) which is exactly the crazy high cost that stopped GF in their tracks at 7nm FinFet.

(Redacted part in regards to NIL and GlobalFoundries)
7FDX can avoid EUV with DUV-optimized shrink or avoid DUV with J-FIL. The more simple the process, the better J-FIL will perform. CNTs J-FIL machines scale back so 22FDX/12FDX could also be shifted for the lower process steps and lower BEOL masks.

I recommend waiting for Canon Nanotechnologies 2018-2019 stuff to be revealed. Which should be soon in regard to the calender.

 

[iBoMbY]

It is still possible GloFo is going to buy some more advanced tech at a later point, like they did with 14LPP. Actually I'm surprised they didn't already do that again, if they had so much trouble. Maybe Samsung was asking too much?

 

[dark zero]

Sadly GloFo is on risk to be broken... and is because they are not getting more customers besides AMD.

There was supposed to be Mediatek there too, what happened to them?

Their current chips are made by TSMC.

 

[Dayman1225]

Maybe Samsung was asking too much?

I think the problem is that GF does not have the EUV machines to put SS 7nm process into any form of HVM that is acceptable, hell, SS barely does.

 

[jpiniero]

It is still possible GloFo is going to buy some more advanced tech at a later point, like they did with 14LPP. Actually I'm surprised they didn't already do that again, if they had so much trouble. Maybe Samsung was asking too much?

Highly unlikely Samsung licenses out another node to GloFo. Hell, they barely wanted to license out 14 nm in the first place but only did it because of Apple.

 

[Hitman928]

It's on the Dresden Fab 1 and CEA Tech R&D roadmap.

Do you have a link or is this one of those things that since it hasn't been explicitly shown that it doesn't exist, then it must actually exist

Which the VLSI Research group based it on.

VLSI Research group just based it on a trend. They looked at GF process roadmap history and noticed a trend and projected from there. The problem is, GF just broke that trend, so their projection is now worthless.

Could 7nm FD-SOI exist at some point? Yes. But it's not likely and based upon your earlier quotes from those involved with R&D at GF, if it does come out, it won't be for several years yet in any kind of volume. TSMC will probably be ready with 3 nm by the time that happens (if it ever does).

 

[Guru]

AMD actually negotiated to get all of the initial 7nm products from TSMC, so if anything its Nvidia that are going to loose out as they will be really late to market with 7nm.

There is still an issue with desktop GPU's with AMD as their initial 7nm products are Vega 20 for the professional market and then their EPYC processors. Its going to be Q2 2019 until there is enough 7nm for AMD to start planning their desktop graphics.

Good move for glo-fo as they have a solid 16nm and 12nm process, so they will stay really strong in the market for the next two years. I mean at some point they will have to reduce size as well, but they have good two years to make solid profits to make a push towards probably 5nm in 2021.

 

[darkswordsman17]

It is still possible GloFo is going to buy some more advanced tech at a later point, like they did with 14LPP. Actually I'm surprised they didn't already do that again, if they had so much trouble. Maybe Samsung was asking too much?

Yeah, I kinda wondered if they wouldn't just wait for the industry to get 7nm up and rolling and then they finish it. But without a major customer and being behind, it likely was not going to produce much.

Still just seems so odd to develop it as far along as they did and then just abandon it entirely. I have to imagine they've been in talks with Samsung about leasing or selling their 7nm fabs. Or someone. I don't know if Apple would be interested in buying their own fabs (thinking beyond 7nm, or even 7nm for stuff that usually lags behind, like Apple Watch, but then GF's FDX processes seem like it'd be a better fit there). But maybe Apple works to get 5nm up and running for them to bring out their first non-Intel/x86 Mac chips or something (maybe Apple follows AMD's Zen design, where they can utilize one die for consumer stuff, 2 for Pro, and then more for their own server chips).

Wait, I know, let's have RIM and Imagination Technologies merge and then buy GF. Maybe HTC would like to get in on it as well. And they could make a new Sega console. See its all coming together, things happen for a reason!

I think the problem is that GF does not have the EUV machines to put SS 7nm process into any form of HVM that is acceptable, hell, SS barely does.

I think the bigger problem is that because of their past screwups, GF had only two customers (AMD and IBM) for their 7nm, and they knew AMD was already working with TSMC's 7nm, so that means AMD's demand would be less than typical. They in a way didn't really have a choice as 7nm was likely only going to lose them a lot of money, and they were likely going to be abandoning 5nm and beyond (maybe not fully, but its definitely going to be pushed well back of when they would have).

 

[moonbogg]

OK so I'm a total laymen here, but I have to ask, if Intel is having so much trouble with 10nm and global foundries are having so much trouble with 7nm, what happens to everyone going forward after 7 and 10nm? 5nm is going to be a piece of cake all of a sudden? A node only lasts a few years, right? We looking at a giant brick wall about to hit us pretty soon?

 

[Xpage]

OK so I'm a total laymen here, but I have to ask, if Intel is having so much trouble with 10nm and global foundries are having so much trouble with 7nm, what happens to everyone going forward after 7 and 10nm? 5nm is going to be a piece of cake all of a sudden? A node only lasts a few years, right? We looking at a giant brick wall about to hit us pretty soon?

Node nomenclature doesn't exactly equal the actual size of transistors. However, I think if at each node drop half of the designs don't move to that process, overall revenues drop raising the cost for everybody that does move to the node. Hopefully the costs to produce the node drop significantly over time, so more people migrate but I think what will happen is that node times will increase significantly for the foundries to recoup their investment then they may decide to make a next node.

I htink a 22FDSOI and 12FDSOI using single / dual patterning respectively, would be a good choice for GF to do if they are dropping out of 7nm forever or at least the next few years. That will give them time to reverse engineer somebody else's 7nm process if they choose.

I thought for the longest time we'd reach physical limits, but I think economics might dominate if we ever reach a 3nm node. I think the 7nm node 7nm, 7nm+, ++, +++ will dominate for at least 5 years before we migrate to a 5nm gate all around, or III-V materials around 2022. Of course we just might make 3nm via marketing but it wouldn't be what was originally planned technically as a 3nm node from what was considered a 3nm node 4 years ago.

 

[PeterScott]

OK so I'm a total laymen here, but I have to ask, if Intel is having so much trouble with 10nm and global foundries are having so much trouble with 7nm, what happens to everyone going forward after 7 and 10nm? 5nm is going to be a piece of cake all of a sudden? A node only lasts a few years, right? We looking at a giant brick wall about to hit us pretty soon?

7nm(10nm Intel) is at the limit of what can be done without EUV. So most of these are starting without EUV, or very limited EUV usage.

TSMC is planning to transition much more EUV during the the 7nm lifecycle, so when EUV is fully debugged at 7nm, it will be more ready to take on 5nm, where it will be completely necessary.

But it will likely still be another difficult transition.
Some info on EUV in 7nm and 5nm implications.
https://www.eetimes.com/document.asp?doc_id=1332860