Globalfoundries 7LP 7nm Leading Performance FINFET process and FX-7 ASIC platform

Page 5 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

ajc9988

Senior member
Apr 1, 2015
278
171
116
I'm not really sure about EPYC and TR will get first treatment of 7nm. Because, as usual, a new process means low yield and AMD can't afford such loss. I bet they will use the smaller chip, either small GPU like P11 successor or mainstream version of Zen2.
After all, AMD hold their schedule very close to their chest, even I won't dare to bet when first mobile Raven Ridge will be delivered, let alone which process (LPP or 14+) they will be used.
Keep all your heads down, guys. It's just about prediction and speculation after all.
That is using the reasoning Intel gave for years, although Intel needed it because you wanted the process refined BEFORE trying the big die chips. Here, all chips are the same size and Epyc even goes down to 8 cores, which is 1 core per CCX. Mainstream cannot do that. Mobile cannot do that. The server chips can.

Now, as we have discussed above, 14nm+ may be the 7nm transistor style on 14nm. This would be doing similar to what is done by Intel, like Sandy for Ivy, Haswell for BW, and Coffee for cannon. So, since they can use as low as one working core per CCX, you could fill out the entire Epyc and TR lineups even with lower yields. You then use IF like they are now. Do you see why server side makes the most sense now?

Edit: also, the server line clocks lower than the mainstream and TR lines. This means if you have an imperfect die that qualifies, but at a lower speed, you don't have to eat the cost like the other lines you mentioned. Intel went with mobile because it has lower clocks for dual and quad cores.
 
Last edited:

raghu78

Diamond Member
Aug 23, 2012
4,093
1,475
136
It looks like GF is improving execution at the right time. Hopefully GF and AMD together maintain the momentum gained with Ryzen launch.

https://www.semiwiki.com/forum/content/7047-globalfoundries-hitting-all-cylinders.html

"GF is now demonstrating execution across the board on their roadmaps and the industry is noticing. GF's process portfolio includes competitive leading edge FinFET processes in 14LPP and 7LP and industry leading RFSOI and SiGe solutions for RF. The new 22FDX FDSOI platform appears well positioned for the emerging IOT space with the 12FDX process in development for future extensions.

GF is now a viable leading foundry from a technology and execution standpoint. The next steps for GF will be to achieve profitability and demonstrate a sustainable financial model for the future
."
 

scannall

Golden Member
Jan 1, 2012
1,944
1,638
136
Honestly, I have a hard time believing this is true.

Intel was first to 14nm back in 2014, while GF only reached 14nm in 2016, so GF was 2 years behind. And now they claim 7nm in 2018, while Intel will only be doing 10nm at that time ? So we are supposed to believe that they have suddenly jumped from being 2 years behind to being one node ahead ? 14nm to 7nm in 2 years while Intel has taken 4 years to go from 14nm to 10nm ? That would be quite miraculous if they pull it off, but I have my doubts.
Certainly possible. Intel has had big stumbles at both 14 an 10nm. And nobody does more, or better bleeding edge research than IBM. An IBM POWER9 probably won't play Crysis on your potato monitor. But they'll drive everything else in your life, that you're barely aware of with ease.
 
Mar 10, 2006
11,715
2,012
126
I love how IBM is now being deified by the enthusiast community because of the connection to GloFo/AMD.

IBM was always big on hyping up fancy new technologies in the labs, but always fell way short when it came time to bring those technologies into mass production at reasonable costs.

There's a reason IBM's foundry business failed and that IBM paid GloFo to take the operation off its hands.
 
  • Like
Reactions: Phynaz and NTMBK

krumme

Diamond Member
Oct 9, 2009
5,952
1,585
136
I love how IBM is now being deified by the enthusiast community because of the connection to GloFo/AMD.

IBM was always big on hyping up fancy new technologies in the labs, but always fell way short when it came time to bring those technologies into mass production at reasonable costs.

There's a reason IBM's foundry business failed and that IBM paid GloFo to take the operation off its hands.
What is actually the problem here ? ;)

The partnership with amd excactly solves that. Add synergy to the entire arm ecosystem with samsung and apple whatnot.

Some enthusiast is living in a cave* as of right now. Each day they are hammering on Intels Ceo or other enthusiast.
What a show.

*[cave]

- Moores law

- Dx9 like benchmark on 1080p 144Hz screen with 1080ti and 7700k

- "when software cost is so high its better to buy a xxx (insert whatever cpu or gpu or individual component) because it only adds 10% to the total cost"
 

NTMBK

Lifer
Nov 14, 2011
10,208
4,940
136
What is actually the problem here ? ;)

The partnership with amd excactly solves that. Add synergy to the entire arm ecosystem with samsung and apple whatnot.

If IBM's foundries have expertise in producing extremely high performance, extremely low yield process nodes which are only suitable for high margin, low volume parts like Z series or POWER... where is the synergy? AMD need a process they can use to crank out mass market GPUs, CPUs and APUs by the million, not some commercially unviable 5GHz special.
 

moinmoin

Diamond Member
Jun 1, 2017
4,933
7,619
136
If IBM's foundries have expertise in producing extremely high performance, extremely low yield process nodes which are only suitable for high margin, low volume parts like Z series or POWER... where is the synergy? AMD need a process they can use to crank out mass market GPUs, CPUs and APUs by the million, not some commercially unviable 5GHz special.
R&D is commonly much wider reaching than the end product they work toward. That IBM only asked for extremely high performance, extremely low yield process nodes from its foundries tells us nothing about their expertise in other areas as they simply were of no use up to now.
 

raghu78

Diamond Member
Aug 23, 2012
4,093
1,475
136
I love how IBM is now being deified by the enthusiast community because of the connection to GloFo/AMD.

IBM was always big on hyping up fancy new technologies in the labs, but always fell way short when it came time to bring those technologies into mass production at reasonable costs.

There's a reason IBM's foundry business failed and that IBM paid GloFo to take the operation off its hands.

IBM's semi-conductor business failed because they did not have the volumes to sustain leading edge fab development. Plain and simple. No other reason. IBM used to design process nodes for their server processor requirements. These server CPUs needed very high frequency, massive die sizes, lots of EDRAM and lots of metal layers. IBM's server CPUs are much more complex than even what Intel delivers. The process node was developed for a specific customer and a specific use case - IBM high performance server CPUs. High yields were never a prerequisite as these server CPUs sold at massive prices and IBM could afford to have poor yields. What do you think is the yield on Volta GV100 at 12FFN. Jensen Huang said the chances of a single V100 chip which is fully enabled working on a wafer are almost zero . I would guess the cumulative yields on V100 are definitely well below 50%.

For a foundry the case is entirely different. GF have very important customers like AMD and IBM. But to remain a viable foundry Globalfoundries must design a leading edge process to address the needs of a larger set of customers and use cases. GF needs to actively pursue and gain the business of multiple fabless customers across mobile, high performance and RF. The performance , power and density needs to be competitive against other foundries like TSMC. The process needs to yield reasonably well and should be able to ramp into very high volume and must be economically viable for customers and the foundry. The design ecosystem and tools need to be robust. The design ecosystem needs to have confidence that the foundry process is viable, both technically and financially otherwise they would not bother spending time and effort to develop tools. GF's 7LP combines IBM's high performance process node development expertise with GF's volume manufacturing expertise. GF is providing 6T libraries for mobile and 9T for high performance. The design ecosystem is robust and there is strong early customer engagement.

GF has a differentiated process offering in its FD-SOI roadmap. 22FDX is in production with 92% yields. More importantly GF 22FDX is ahead of TSMC 22ULP and Intel 22FL in time to market and is better on performance, power and cost with superior RF perf.

http://www.eetimes.com/document.asp?doc_id=1332328&page_number=2

"“You can see chips from 22FDX in the market,” said chief executive Sanjay Jha, noting an ADAS device from Dreamchip and a 450-MHz dual ARM M4 processor from Ineda Systems. The promise of the technology motivated rivals TSMC and Intel to roll out their own 22-nm alternatives.

“Our competitors woke up, but instead of a ground-up design, they tweaked what they had … they could match us in some areas, but not across the board … and their technologies are in development while ours is qualified with products yielding at 92%,” said Patton, claiming that both rivals require more masks.

A follow-on 12-nm FD-SOI node should deliver 25% more performance or 47% less power compared to 22FDX and support both forward and reverse body bias capabilities. Prototyping will start next year with full production targeting early 2019. It is expected to use about 50 masks compared to about 80 for 7-nm FinFET."


https://www.semiwiki.com/forum/content/7029-different-globalfoundries.html

"The third keynote was Gary Patton which covered the GF CMOS and FD-SOI roadmaps which included slides on Photonics, EUV, and of course Machine Learning. A couple of Gary's FD-SOI slides caught my eye. 22FDX has a nice ecosystem developing with all of the top names. Currently, 22FDX has 15 confirmed tape-outs happening this year and next. I know several of those first hand so that number is easy to believe. Scott Jones attended the conference as well and spent time with Gary so expect much more detail from him next week."

https://www.semiwiki.com/forum/content/7047-globalfoundries-hitting-all-cylinders.html

"GF has the most aggressive FDSOI roadmap in the industry with their 22FDX process ramping now and 12FDX in development. 22FDX and 12FDX are targeted at IOT, mobile and RF applications and offer FinFET like performance at lower cost. The thin silicon device layer and thin buried oxide layer of FDSOI enable body biasing to be used to tune performance and power consumption. GF picked 22nm and 12nm as their technology nodes to limit multi-patterning and provide lower costs. 22FDX has 40% less masks than 14nm FinFET and 12FDX has 40% less masks than 10nm/7nm FinFET. For large chips FDSOI can't compete with FinFETs but for smaller chips FDX offers lower cost, lower power and better analog and RF performance. 22FDX was qualified in March of 2017 and is approaching 28nm mature yields. PDK 1.2 has been delivered along with validated design flows and IP. GF is building up 1.5 million wafers per year of capacity between Dresden and the new Chengdu China fab. Figure 2 illustrates the current status of 22FDX.

I often get asked whether I think FDSOI will really take off and my answer is that I do think it will take off for IOT and other mobile applications. GF reported they now have 135 early engagements, 102 PDK downloads, 72 Invecas IP downloads and there will be 20 test designs on multi product wafers by the end of 2017 with 15 product tape-outs by the end of 2018.

12FDX is in development and expected to provide 26% better performance and 47% lower power than 22FDX. 12FDX will have a minimum energy point of <0.4 volts well below what FinFETs can offer. Samsung is the only other foundry pursuing FDSOI and their process nodes are 28nm currently with 18nm in development, both less dense than GF's competing 22nm and 12nm offerings
."


Your problem is always viewing anything related to AMD and GF as negative and viewing everything related to Intel as positive.
 
Last edited:
  • Like
Reactions: moinmoin
Mar 10, 2006
11,715
2,012
126
IBM's semi-conductor business failed because they did not have the volumes to sustain leading edge fab development.

IBM used to invest very heavily in leading edge process technology development, and time and again they always had issues ramping its technologies into mass production at high yields. This isn't a question of not having the volumes to invest (IBM makes so much money that it can invest all it cares to), but a question of making bad technology bets time and time again.

This is why a lot of the IBM Alliance members would diverge their own processes from the technologies that IBM developed. Samsung was one of the companies that used to build parallel processes -- one based on the IBM alliance technology and its own home-grown recipes, and it was ultimately Samsung's own home-grown technology that succeeded in the market.

The reason IBM didn't have the volumes was because they couldn't win the business away from more capable foundries like TSMC and Samsung. Given IBM's chest-beating about its technological superiority, first to new techniques, etc. was worth absolutely zero. Plain and simple.

Saying that IBM's semi biz only failed because they didn't have the volumes really is a tautology. Of course a business that doesn't manage to sell things fails. What I'm telling you is why they couldn't sell things.
 
Last edited:

Phynaz

Lifer
Mar 13, 2006
10,140
819
126
GF has a differentiated process offering in its FD-SOI roadmap. 22FDX is in production with 92% yields. More importantly GF 22FDX is ahead of TSMC 22ULP and Intel 22FL in time to market and is better on performance, power and cost with superior RF perf.

No need to get all ADF about GloFo. You're even making page size posts about 22FDX and 12FDX in the 7LP thread.
They are two different things, 12FDX is planer FD-SOI, 7LP is bulk FinFet. They serve two different markets.

You are also comparing 22FDX to Intel 22FFL. Again, you're comparing apples to oranges. You're falling for a marketing number without even knowing it.

Slow down, read, understand what you are reading and perhaps comment on it. Currently your posts of pasting in a wall of bolded (shouting) text taken from another forum just appear to be ranting.
 

maddie

Diamond Member
Jul 18, 2010
4,723
4,628
136
IBM used to invest very heavily in leading edge process technology development, and time and again they always had issues ramping its technologies into mass production at high yields. This isn't a question of not having the volumes to invest (IBM makes so much money that it can invest all it cares to), but a question of making bad technology bets time and time again.

This is why a lot of the IBM Alliance members would diverge their own processes from the technologies that IBM developed. Samsung was one of the companies that used to build parallel processes -- one based on the IBM alliance technology and its own home-grown recipes, and it was ultimately Samsung's own home-grown technology that succeeded in the market.

The reason IBM didn't have the volumes was because they couldn't win the business away from more capable foundries like TSMC and Samsung. Given IBM's chest-beating about its technological superiority, first to new techniques, etc. was worth absolutely zero. Plain and simple.

Saying that IBM's semi biz only failed because they didn't have the volumes really is a tautology. Of course a business that doesn't manage to sell things fails. What I'm telling you is why they couldn't sell things.
I thought I had a reasonable idea of the computer industry but it appears not.

Can you please expand on this? What IBM products using their custom process tech ever had a mass production status?
 

raghu78

Diamond Member
Aug 23, 2012
4,093
1,475
136
IBM used to invest very heavily in leading edge process technology development, and time and again they always had issues ramping its technologies into mass production at high yields. This isn't a question of not having the volumes to invest (IBM makes so much money that it can invest all it cares to), but a question of making bad technology bets time and time again.

This is why a lot of the IBM Alliance members would diverge their own processes from the technologies that IBM developed. Samsung was one of the companies that used to build parallel processes -- one based on the IBM alliance technology and its own home-grown recipes, and it was ultimately Samsung's own home-grown technology that succeeded in the market.

The reason IBM didn't have the volumes was because they couldn't win the business away from more capable foundries like TSMC and Samsung. Given IBM's chest-beating about its technological superiority, first to new techniques, etc. was worth absolutely zero. Plain and simple.

Saying that IBM's semi biz only failed because they didn't have the volumes really is a tautology. Of course a business that doesn't manage to sell things fails. What I'm telling you is why they couldn't sell things.

IBM was a vertically integrated company from the early days of computing. IBM used to invest in leading edge process development as a result of that philosophy. Moreover it allowed them to build very high performance processors. IBM's semiconductor manufacturing was to support IBM product development. Plain and simple. IBM semiconductor never had the culture or business model or financial discipline of a pure play foundry like TSMC. IBM was able to develop process nodes as long as it was not exorbitantly expensive to build and maintain fabs. Starting from 32nm/28nm node those fab costs have gotten to a level where if you do not have the product volume you cannot afford to have fabs at the leading edge. Plain and simple. Comparing IBM semiconductor with TSMC is ridiculous. They are poles apart. IBM semiconductor was never financially accountable like TSMC was to it shareholders. IBM maintained the fabs even if they were not hugely profitable as it served their product needs. It cannot be done now due to the costs of building/maintaining a leading edge fab. Today Intel is the only company in the world which fabs its own products. Samsung Foundry has been spun off as an independent business unit within Samsung group.

btw just as you mention Samsung builds from the efforts of the research alliance which includes IBM, GF and Samsung the same is true for GF. GF 7LP is the first node after TSMC to enter 7nm risk production (roughly 9-12 months behind TSMC). So that should give you an idea of the pragmatism and execution discipline required as GF wisely did not bet everything on EUV as Samsung did. If GF fails to hit risk production milestone or fails to ramp volume production on time then you can say that the problems of the pre IBM semi acquisition era remain. But until then we will have to wait and see.
 

raghu78

Diamond Member
Aug 23, 2012
4,093
1,475
136
No need to get all ADF about GloFo. You're even making page size posts about 22FDX and 12FDX in the 7LP thread.
They are two different things, 12FDX is planer FD-SOI, 7LP is bulk FinFet. They serve two different markets.

You are also comparing 22FDX to Intel 22FFL. Again, you're comparing apples to oranges. You're falling for a marketing number without even knowing it.

Slow down, read, understand what you are reading and perhaps comment on it. Currently your posts of pasting in a wall of bolded (shouting) text taken from another forum just appear to be ranting.

I don't care about what an IDF member thinks. btw do your homework. GF 22FDX vs TSMC 22ULP vs Intel 22FL is very much a valid comparison. These are competing processes for low power and low cost mobile.

https://m.eet.com/content/images/eetimes/22nm nodes x 800_1505974563.jpg
 
Mar 10, 2006
11,715
2,012
126
But until then we will have to wait and see.

That's all I'm trying to say. You will notice in the GloFo 12nm thread, I gave GloFo and AMD praise because it appears that their execution on 12nm -- at least in terms of time-to-market --- is quite solid. If GloFo ramps 7LP and yet again delivers on its promises, then I would happily praise them (and maybe even buy products based on the technology).
 
  • Like
Reactions: CatMerc

raghu78

Diamond Member
Aug 23, 2012
4,093
1,475
136
That's all I'm trying to say. You will notice in the GloFo 12nm thread, I gave GloFo and AMD praise because it appears that their execution on 12nm -- at least in terms of time-to-market --- is quite solid. If GloFo ramps 7LP and yet again delivers on its promises, then I would happily praise them (and maybe even buy products based on the technology).

I agree. GF has to execute their risk and volume production ramp milestones with 7LP on time to prove they can deliver at the leading edge. I think they are on track. Moreover sometimes you get good hints on execution from the design ecosystem and from people who follow the industry from the inside. Thats where I think the article from semiwiki fits in.

https://www.semiwiki.com/forum/content/7047-globalfoundries-hitting-all-cylinders.html

I have guarded optimism wrt GF and their chances in the long run. I am hoping they execute and turn into a strong and financially viable No.2 foundry at the leading edge. That would be great for the foundry and fabless industry and especially AMD.
 

Abwx

Lifer
Apr 2, 2011
10,847
3,297
136
I agree. GF has to execute their risk and volume production ramp milestones with 7LP on time to prove they can deliver at the leading edge. I think they are on track. Moreover sometimes you get good hints on execution from the design ecosystem and from people who follow the industry from the inside. Thats where I think the article from semiwiki fits in.

https://www.semiwiki.com/forum/content/7047-globalfoundries-hitting-all-cylinders.html

I have guarded optimism wrt GF and their chances in the long run. I am hoping they execute and turn into a strong and financially viable No.2 foundry at the leading edge. That would be great for the foundry and fabless industry and especially AMD.

There are some impressive numbers claimed, besides it look like AMD is currently using the 7LP PDKs, or they already used them since they are apparently in a close partnership with GF.

20497d1506621717-2017-sc-gtc_gary-patton_page_26-jpg
 
  • Like
Reactions: ajc9988 and Drazick

maddie

Diamond Member
Jul 18, 2010
4,723
4,628
136
There are some impressive numbers claimed, besides it look like AMD is currently using the 7LP PDKs, or they already used them since they are apparently in a close partnership with GF.

20497d1506621717-2017-sc-gtc_gary-patton_page_26-jpg
If accurate, this suggests better process performance than Intel's 10nm & possibly 10nm+. Hell of ride coming for the next few years.
 
  • Like
Reactions: ajc9988 and Drazick

Excessi0n

Member
Jul 25, 2014
140
36
101
I love how IBM is now being deified by the enthusiast community because of the connection to GloFo/AMD.

People are just excited since it will be the first time a non-Intel company will have consumer products on a real high-performance node since... what, Bulldozer? And that hardly even counts since the architecture was trash anyways.

GloFo's 7LP might very well have problems, but with the IBM design heritage we can be fairly confident that performance won't be one of them. And that's exciting for us enthusiasts since we tend to care about performance first and foremost.
 
  • Like
Reactions: raghu78

raghu78

Diamond Member
Aug 23, 2012
4,093
1,475
136
If accurate, this suggests better process performance than Intel's 10nm & possibly 10nm+. Hell of ride coming for the next few years.

We will have to wait for Zen 2 and Icelake to draw conclusions about the respective process nodes. As you said we are in for one hell of a ride. :)
 
  • Like
Reactions: ajc9988 and Drazick