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

[jpiniero]

As for Turing, I'm sure nVidia didn't want to wait for 7 nm, given how long it has been since the consumer GPUs got updated.

 

[darkswordsman17]

Agreed. And AMD has historically been first to market with a GPU on a new node from TSMC, while nvidia's attempt prior to 16nm was the disastrous Fermi. They haven't even launched 12nm (optimized 16nm), i doubt they'll be taping out anything on 7nm until the first part of next year. Especially with wafer supply ramping and multiple companies ahead of them eating up available wafers.

I might be wrong, but I think they beat AMD to using some new node in that time, by making a low end (the 750?) chip. I thought they might be doing that with 7nm, lead with the lower chips (1060 and below), giving it time to mature (and letting them add some of the ray tracing hardware without it being a huge chip), then go big, and move on down from there. Unless AMD has a wonder in Navi and its like 4000 series all over again, but even then, Pascal would still be fairly competitive and should be cheap to produce still (so they could easily get a lot more competitive on price).

Will be interesting to see how things play out. I'm not sure if this is good or bad for AMD, will depend on how they deal with it.

 

[Muhammed]

If Nvidia knew they were jumping onto 7NM in the near future,they would not have invested so much dosh into developing an optimised 16NM node which can make massive chips,and it hints that Nvidia was of the opinion they would not be shrinking their GPUs anytime soon.

Volta at 12nm is almost a year and half old now, waiting is for morons, you create your own tech to compete. Volta and Turing paid off, as 7nm is still far away now for consumers, and they knew it. NVIDIA will use 7nm for Ampere. The next arc after Volta.

Even AMD made Zen+ using 12nm, again playing the waiting game is for dummies only.

Agreed. And AMD has historically been first to market with a GPU on a new node from TSMC, while nvidia's attempt prior to 16nm was the disastrous Fermi. They haven't even launched 12nm (optimized 16nm), i doubt they'll be taping out anything on 7nm until the first part of next year. Especially with wafer supply ramping and multiple companies ahead of them eating up available wafers.

History doesn't mean much here, NVIDIA is more keen on advanced nodes to subsidize it's IP in AI and HPC. Rest assured they are keeping track of 7nm. NVIDIA is keeping it's card close to the chest, we never knew Volta is on custom 12nm until it launched.

 

[Muhammed]

I thought they might be doing that with 7nm, lead with the lower chips (1060 and below), giving it time to mature (and letting them add some of the ray tracing hardware without it being a huge chip), then go big, and move on down from there

That won't be possible this time because 7nm is too damn expensive at the moment. Even AMD chose AI chips as the candidate for a pipe cleaner.

 

[USER8000]

If Nvidia knew they were jumping onto 7NM in the near future,they would not have invested so much dosh into developing an optimised 16NM node which can make massive chips,and it hints that Nvidia was of the opinion they would not be shrinking their GPUs anytime soon.

Volta at 12nm is almost a year and half old now, waiting is for morons, you create your own tech to compete. Volta and Turing paid off, 7nm is still away now for consumers, and they knew it. NVIDIA will use 7nm for Ampere. The next arc after Volta.

Which does not change what I said - they knew very well 7NM would be congested so that is why 12NM was funded so they knew very well they would not be shrinking GPUs for a while. If they knew 7NM was viable for consumer and professional GPUs,you would not be having a 754MM2 Turing GPU for consumer and professional usage being made on 12NM,or have you ignored the Turing chips are also being used in non-consumer products:

https://nvidianews.nvidia.com/news/nvidia-unveils-quadro-rtx-worlds-first-ray-tracing-gpu

A product with 48GB of RAM is not a consumer product.

Maybe in your world,you think 754MM2 GPUs are small,but by every metric that is the second largest chip ever made.

The Turing launch has caught off guard younger enthusiasts - its a move back to the days of Fermi and earlier. Nvidia only split consumer and professional high end chips with Maxwell and Pascal.

Now they are back to what things used to be.

 

[Muhammed]

If they knew 7NM was viable for consumer and professional GPUs,you would not be having a 754MM2 Turing GPU for consumer and professional usage being made on 12NM

Because it will still be available in VERY limited capacities (duh) and will be expensive as hell. With production being sparse and interrupted. No sense in staking your portfolio on an unproven node, that is yet to materialize.

AMD did the same with their 12nm Zen+, if they knew 7nm is close and plentiful they would have waited for 7nm Zen 2, but they didn't.

 

[NostaSeronx]

You're missing or ignoring that was paired with "low power and relatively low cost" as in its products that meet all of those things. Plus you're ignoring that they're talking about completely different types of chips with regards to that as well (the "RF, analog, or mixed-signal"). They're trying to become the cutting edge place for chips that used to go on nodes generation or more back (where cost was the biggest factor). They're getting out of bleeding edge high performance processor production entirely. I don't think AMD is going to make designs targeting GF's processes beyond the 14/12nm nodes they're already using..

Low power is a given and low cost is also a given. FDSOI follows Moore's Law, where FinFETs do not.

Per transistor FDSOI is cheaper than FinFETs on a given node. Given the same R&D FDSOI can have three shrinks for every one FinFET shrink.

12FDX is the only successor to 12LP/14LPP now:
https://images.anandtech.com/doci/12438/fab8_media_day_patton_final-page-005.jpg
Someone cut and paste 12FDX into the middle of this roadmap and wipe 7LP off it. Have everything go into 12FDX as it is the only known leading edge High Performance node at GlobalFoundries.

With the 7LP gone, 12FDX must fit into premium now.

 

[Jimzz]

I wonder where this leaves IBMs P10? They were designing around 10nm at GF. So either higher power/lower speed at 12nm or work with Samsung/TSMC?

 

[scannall]

It's really too bad. They have the talent, from the IBM foundry "purchase". Prediciting a severe brain drain from GF shortly.

 

[JDG1980]

Because it will still be available in VERY limited capacities (duh) and will be expensive as hell.

Production capacities can't be that limited if this year's iPhones will have TSMC 7nm chips.

 

[darkswordsman17]

Low power is a given and low cost is also a given. FDSOI follows Moore's Law, where FinFETs do not.

Per transistor FDSOI is cheaper than FinFETs on a given node. Given the same R&D FDSOI can have three shrinks for every one FinFET shrink.

12FDX is the only successor to 12LP/14LPP now:
https://images.anandtech.com/doci/12438/fab8_media_day_patton_final-page-005.jpg
Someone cut and paste 12FDX into the middle of this roadmap and wipe 7LP off it. Have everything go into 12FDX as it is the only known leading edge High Performance node at GlobalFoundries.

With the 7LP gone, 12FDX must fit into premium now.

No its not, you're refusing to accept that statement was not about being relative thing to FinFET, but relative to the types of chips they're going to produce. They explicitly say they're targeting different chips. I know you so desperately want that magical wonder Bulldozer architecture CPU using FDSOI that you rave about so much, but you need to stop trying to project your nonsense and claim it as fact.

No it doesn't because they can simply just not make those chips any more. That's deliberately their point even and they outright say they're targeting that process for a completely different type/class of chip. That's what you're refusing to accept. AMD is not transitioning to that, they're moving to TSMC 7nm. So cut and paste GF 7nm with just TSMC 7nm.

 

[darkswordsman17]

Production capacities can't be that limited if this year's iPhones will have TSMC 7nm chips.

I think your logic is backwards. Apple is likely their biggest customers and gets preferential treatment. That the iPhone SoC is being fabbed there is why people are talking about potential production constraints, because Apple will get priority.

That doesn't mean there necessarily will be any actual constraints (although TSMC hopefully has put some extra emphasis on their security protocol after their little issue just a bit ago), but now AMD is going to be looking at making all of their stuff there (unless Samsung gets up and going and offers a compelling reason - mostly just competitive production for good price), Nvidia is, many ARM designs are likely to be there (although with Qualcomm and Samsung buddied up that's a big one that isn't needing to get some of TSMC's capacity). I think even Intel has some modem fabbed with them (although it doesn't use the bleeding edge so it doesn't matter). I'm not sure who all else GF was fabbing for (IBM and not sure who else) but some of them might be looking at moving over.

I haven't heard much about Samsung's 7nm, so I don't know if its supposed to be good or when its even supposed to be available. And Intel is struggling with 10nm so even if they wanted to they're probably not going to be picking up any of the fab work for companies that were looking to do 7nm at GF or TSMC.

 

[maddie]

No its not, you're refusing to accept that statement was not about being relative thing to FinFET, but relative to the types of chips they're going to produce. They explicitly say they're targeting different chips. I know you so desperately want that magical wonder Bulldozer architecture CPU using FDSOI that you rave about so much, but you need to stop trying to project your nonsense and claim it as fact.

No it doesn't because they can simply just not make those chips any more. That's deliberately their point even and they outright say they're targeting that process for a completely different type/class of chip. That's what you're refusing to accept. AMD is not transitioning to that, they're moving to TSMC 7nm. So cut and paste GF 7nm with just TSMC 7nm.

I just checked the GloFlo info and they're claiming the following. My thinking is that this would make an excellent process for mobile designs. Raven Ridge or follow-on would be unbeatable.

"If you look at performance with back-bias 22FDX is the same or better than 16/14nm FinFET process. With 12FDX with back bias you get better than 10nm FinFET processes"
In short 12FDX offers performance of 10nm FinFET with better power consumption and lower cost than 16nm FinFET.

 

[NostaSeronx]

No its not, you're refusing to accept that statement was not about being relative thing to FinFET, but relative to the types of chips they're going to produce.

It is relative to FinFETs built at GlobalFoundries only. GlobalFoundries by 2020 will have only one FinFET foundry, while two for FDSOI. (Officially)
-
-> The CTO stressed that the decision was made not based on technical issues that the company faced, but on a careful consideration of business opportunities the company had with its 7LP platform as well as financial concerns.
-> GlobalFoundries scuttles 7nm chip plans claiming no demand.
-> In announcing the move, Caulfield said companies don’t seem to have much interest in the planned 7nm architecture. Rather, they are planning to stay with the current-gen architectures and squeeze performance out by other means.
-> Gary Patton admits that GlobalFoundries never planned to be a leading producer of 7-nm chips in terms of volume. Furthermore, the company has been seeing increasing adoption of its 14LPP/12LP technologies by designers of various emerging devices, keeping Fab 8 busy and leaving fewer step-and-scan systems for 7LP products.

-> What is next:
1. Scaling Out the 14LPP/12LP
2. Investing in FD-SOI;
(quote)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. GF is uniquely positioned to serve this burgeoning market for “connected intelligence,” with strong demand in new areas such as autonomous driving, IoT and the global transition to 5G.(quote)
-
It does not say they are dropping leading edge on FDX. Since, the only comparison is Samsung currently. It is severely lacking in any competition, GlobalFoundries won before it even started.

FDSOI at GlobalFoundries before this decision was the sandbag child. GlobalFoundries didn't want FDSOI to compete with FinFETs. So, they bombed the benchmarks of FDSOI and removed any mention of high performance.

Only one foundry at GlobalFoundries is getting the GigaFab treatment:
https://www.anandtech.com/show/12534/change-of-strategy-globalfoundries-3-0/3

The most expensive UTBB FDSOI is still cheaper than the least expensive FinFET. When GlobalFoundries does not have their own product on the benchmark:
http://www.chinaisgood.com/wn/14/zcixagxa.html
22FDX obliterates the 14nm-based 22FFL node from Intel.

 

[ThatBuzzkiller]

Matters not much in the grand scheme of things anyways ...

Just means that if an Arab government backed investment fund (Mubadala is a UAE state owned enterprise) couldn't keep up, it means that TSMC's days are numbered as well even if they are the market leader of all pure play foundries. If a not too dissimilar sized economy like UAE called it quits before even making the transition to 7nm then 3nm will surely be TSMC's Taiwan backed breaking point considering their economy won't be able to absorb the cost of both developing and running such a facility especially in the face of heated competition from both Intel and Samsung who's looking to break into their share of the pie as well ...

While GF may have went down there's still another serious and growing player out there who's also backed by the ambitions of a government running the 2nd largest economy ... (not a good time that this player is growing seeing how cramped the space is)

In the near future it all will come down to Intel, Samsung, TSMC as well as this supposed growing player. In the forseeable future it is uncertain if either Intel or the unnamed competitor will come out on top ...

GF along with UMC will be stepping down from the race ...

 

[JDG1980]

I think your logic is backwards. Apple is likely their biggest customers and gets preferential treatment. That the iPhone SoC is being fabbed there is why people are talking about potential production constraints, because Apple will get priority.

No doubt, but that's a prioritization issue, not a process maturity issue. And given the demand for 7nm, TSMC has an incentive to maximize production on this lucrative process as much as they can.

 

[slashy16]

GF will ether be Bankrupt or bought out within a year. Any company that stops investing in R&D is finished, especially if they are technology orientated.

 

[ThatBuzzkiller]

GF will ether be Bankrupt or bought out within a year. Any company that stops investing in R&D is finished, especially if they are technology orientated.

Not really, GF will most likely slowly fade into irrelevance rather than face an imminent collapse just like the others such as TowerJazz or United Microelectronics who are still very much in the industry today. They will just cease to compete in high end manufacturing since there's very few corporations or countries/unions in some cases who can make it into the top of the chain ...

 

[tamz_msc]

AMD did the same with their 12nm Zen+, if they knew 7nm is close and plentiful they would have waited for 7nm Zen 2, but they didn't.

Zen 2 *IS* on 7nm. 7nm is close and plentiful enough for AMD. That's why Rome is ready to be sampled and Vega Instinct 7nm is launching in Q4 this year.

 

[Muhammed]

Zen 2 *IS* on 7nm. 7nm is close and plentiful enough for AMD. That's why Rome is ready to be sampled and Vega Instinct 7nm is launching in Q4 this year.

Oh? so why no consumer 7nm CPUs or GPUs by Q4? Why did Zen+ launch on 12nm few months ago, instead of 7nm?
That's the point we are discussing ..

 

[tamz_msc]

Oh? so why no consumer 7nm CPUs or GPUs by Q4? Why did Zen+ launch on 12nm few months ago, instead of 7nm?
That's the point we are discussing ..

Zen+ is what Zen was supposed to be - tighter tolerances on latencies, fixing a few errata and a properly functioning boost algorithm. These things aren't anything major architecturally such that it warrants Zen+ existing on 7nm. Testing and qualification on a completely new node would have taken a longer time to market, hence 12nm. The distinction between consumer/data centric products is irrelevant because every company initially markets a low-yielding/ expensive to fab product at those who can pay for them. That's the reason why Volta was made before Turing on 12nm FFN, and Intel's fully functional/revised 10nm is going to be used for Icelake-SP first.

 

[french toast]

I am am pretty shocked by this, this is terrible for AMD it really is.
What is zen 2 going to be fabbed on?? Is there going to be a delay? Capacity shortages?..aweful.

What I can say is well done to Lisa su for having some contingency plan, wonder if AMD will hit Samsung up for their 7nm process?..they lost Qualcomm to TSMC so room for AMD to step in?...how long would that take to move over all everything?

Without capacity there is no chance AMD can shift the units required to claw back market share..even if they produce an awesome product, this can't be a good situation for AMD.

Hopefully glofo was honest with AMD all the way through and gave them a long heads up before this, and play fair with the agreement, let's be honest Glofo have royally screwed AMD over down the years with their incompetence!.
Pretty shocking, i was looking forward to this process, had potential to be highest clocking.

Intel sure are laughing right now, they are back in the game.

Edit; AMD have confirmed to Forbes that it is using TSMC for all of it's upcoming processor's and multiple tapeouts have taken place with TSMC, no setbacks are predicted.
https://www.forbes.com/sites/marcoc...sors-to-be-manufactured-at-tsmc/#5407070b4642 Looks like AMD had a heads up or Lisa Su is a fortune teller.

 

[exquisitechar]

Intel messed up, Samsung is well behind TSMC, GF is out of the game...and it is very likely that TSMC's process is underperforming. Nice. If TSMC seriously flounders in the future, things won't be looking good.

Not expecting high clocks out of Zen 2. Hope they make up for it in other ways.

 

[Gideon]

Hopefully glofo was honest with AMD all the way through and gave them a long heads up before this, and play fair with the agreement, let's be honest Glofo have royally screwed AMD over down the years with their incompetence!.
Pretty shocking, i was looking forward to this process, had potential to be highest clocking.

Yes, it's shocking how bad their execution has been. Especially given their starting point and funds initially:

32nm and 28nm were both delayed at least 3-6 months and used the stupid "gate-first" approach, limiting performance
20nm was canned as a failure
14nm was canned as a failure, replaced by a licenced samsung process

Just when it looked that 7nm with IBMs help was coming along, they decide to just throw in the towel. With an AMD chip about to be taped out in Q4.

 

[Gideon]

Intel messed up, Samsung is well behind TSMC, GF is out of the game...and it is very likely that TSMC's process is underperforming. Nice. If TSMC seriously flounders in the future, things won't be looking good.

Not expecting high clocks out of Zen 2. Hope they make up for it in other ways.

And this too. So much for my hope for a ~5Ghz Single Core Boost Zen 2 :'(
It will get there eventually with refined EUV processes, but it will take quite some time