Will we ever see STMicro or GF offer 7nm-class chips?

[Xebec]

Curious - as 7nm is becoming “older”, is there any chance we’ll see companies like STMicro or Global Foundries resume progress towards denser nodes?

I saw with their newest joint fab they talked about 18nm being ‘relevant in automotive for decades’, but IMO that’s also based on legacy automaker engineering/thinking with hundreds of suppliers providing separate parts, which is becoming less viable over time.

 

[nghanayem]

I don't know if I necessarily agree with the assumption, but as for the question itself, if I had to place a bet, I would bet no. For STM to make that jump would require finFETs and then heavy MP, I just don't see that happening in a world where TI doesn't even have a node with HKMG (their best is still their old 45nm if memory serves). GF is a different story because they had the technology mostly done, but I still don't think so just because GF prob doesn't have the people and know-how anymore to just hit a button and restart their 7LPP line. There is also the sucky reality of having to compete with fully depreciated N7, SF 7LPP, and the close enough 8LPP fabs that also already have entrenched design ecosystems. So unless there is big subsidy money being thrown around for them to enter the "7nm" era AND everybody and their mother works to bootstrap their design ecosystems, I just don't see it happening. I would certainly be cool if it did happen though.

 

[Daniel Nenni]

I don't think GF has the talent/resources to develop 7nm and Samsung does not seem to be interested in licensing their technology past 14nm. I heard there were quite a few problems with the 14nm licensing agreement between Samsung and GF. SMIC said they have 7nm but from what I hear it never made it into manufacturing and UMC never really made the transition to FinFETs,

You have to give TSMC, Intel, and Samsung a lot of credit for 7nm and below. That was not an easy transition.

 

[semiman]

Unless we find a technology to significantly reduce EUV costs + applications which requires tons of sub-7nm chips(+with little IP ecosystem), It's unlikely.

 

[MikeGrimes]

Curious - as 7nm is becoming “older”, is there any chance we’ll see companies like STMicro or Global Foundries resume progress towards denser nodes?

I saw with their newest joint fab they talked about 18nm being ‘relevant in automotive for decades’, but IMO that’s also based on legacy automaker engineering/thinking with hundreds of suppliers providing separate parts, which is becoming less viable over time.

To paraphrase an executive from Ford: "You do not need 5nm for a Windows or Seats". The best industry play is multiple technology levels.

 

[ta152h]

I don't think GF has the talent/resources to develop 7nm and Samsung does not seem to be interested in licensing their technology past 14nm. I heard there were quite a few problems with the 14nm licensing agreement between Samsung and GF. SMIC said they have 7nm but from what I hear it never made it into manufacturing and UMC never really made the transition to FinFETs,

You have to give TSMC, Intel, and Samsung a lot of credit for 7nm and below. That was not an easy transition.

Doesn't IBM deserve some credit too?

 

[Daniel Nenni]

Doesn't IBM deserve some credit too?

Credit for what exactly? Credit for press releases? Credit for developing processes that never make it into production? Credit for patent trolling? All of the above?

 

[ta152h]

Credit for what exactly? Credit for press releases? Credit for developing processes that never make it into production? Credit for patent trolling? All of the above?

I was under the impression their processes are behind the CPUs in their mainframes, and POWER based computers. At least the former. And they reach some crazy clock speeds. And they just licensed their tech to Rapidus. So, clearly they felt it was something worth licensing and using.

But, for some reason you discount this, and no doubt for good reasons. What more do you know? Is there stuff going on behind the scenes, or stuff I'm missing, because it seems like they have some pretty impressive tech from the cheap seats. I mean, their mainframes are technological marvels. So, please clarify why you are pretty unimpressed with them, because I was under the opposite impression, and definitely need to reexamine that if it's not accurate.

 

[blueone]

The 7nm Telum processor for Z-series systems are currently manufactured by Samsung.

IBM and Samsung Unveil Semiconductor Breakthrough That Defies Conventional Design

IBM and Samsung Electronics jointly announced a breakthrough in semiconductor design utilizing a new vertical transistor architecture that demonstrates a path to scaling beyond nanosheet, and has the potential to reduce energy usage by 85 percent.

newsroom.ibm.com

Rapidus did license IBM's 2nm process and the joint development was widely publicized; the target date for chip production is 2025. Odd that they would choose Rapidus over TSMC, since TSMC is ahead of Rapidus on 2nm. Or perhaps the Japanese subsidies are going to benefit IBM in some way, even an indirect way (increasing Japanese mainframe sales)?

Japan to hike subsidies to 2nm chip maker Rapidus, an IBM partner

As countries globally look to boost homegrown chip production, Japan will increase support for IBM partner Rapidus, which aims to produce 2nm chips by 2025.

www.networkworld.com

Perhaps Z-series volumes are so low and margins are so high perhaps IBM can afford to take process risks others can't.

 

[ta152h]

The 7nm Telum processor for Z-series systems are currently manufactured by Samsung.

IBM and Samsung Unveil Semiconductor Breakthrough That Defies Conventional Design

IBM and Samsung Electronics jointly announced a breakthrough in semiconductor design utilizing a new vertical transistor architecture that demonstrates a path to scaling beyond nanosheet, and has the potential to reduce energy usage by 85 percent.

newsroom.ibm.com

Rapidus did license IBM's 2nm process and the joint development was widely publicized; the target date for chip production is 2025. Odd that they would choose Rapidus over TSMC, since TSMC is ahead of Rapidus on 2nm. Or perhaps the Japanese subsidies are going to benefit IBM in some way, even an indirect way (increasing Japanese mainframe sales)?

Japan to hike subsidies to 2nm chip maker Rapidus, an IBM partner

As countries globally look to boost homegrown chip production, Japan will increase support for IBM partner Rapidus, which aims to produce 2nm chips by 2025.

www.networkworld.com

Perhaps Z-series volumes are so low and margins are so high perhaps IBM can afford to take process risks others can't.

True about Samsung, but it's an IBM designed process. And it works quite well.

IBM has pretty much never worked with TSMC as far as I know. It could easily be TSMC isn't as interested in those arrangements, since they do pretty well on their own. I'm not sure.

IBM clearly has different criteria with their processes, since the machines are so expensive, and have such exotic cooling available to them. And they surely can't be slow, not at that price, although reliability is probably their strongest card. Obviously, cost isn't a big consideration.

I'm just confused why they get so little respect when their chips are clearly quite good at what they are supposed to be good at. But, I'm probably missing something, so this will be informative.

 

[blueone]

I'm just confused why they get so little respect when their chips are clearly quite good at what they are supposed to be good at. But, I'm probably missing something, so this will be informative.

I respect IBM a lot in chip design. I'm not a fabrication process expert, so I can't judge whether their process technology is awesome or not. I suspect that IBM, like Intel has for decades, tuned their process research to benefit the specific needs of their proprietary chips, and that could be one reason (among perhaps multiple) that IBM still does process research. I've never seen anything published specifically on the topic.

 

[ta152h]

I respect IBM a lot in chip design. I'm not a fabrication process expert, so I can't judge whether their process technology is awesome or not. I suspect that IBM, like Intel has for decades, tuned their process research to benefit the specific needs of their proprietary chips, and that could be one reason (among perhaps multiple) that IBM still does process research. I've never seen anything published specifically on the topic.

Yes, absolutely true. Their process research is for their own chips, both mainframe and POWER.

I read a year or so ago they were working with Intel now, so maybe this will be the end of their relationship with Samsung for the near future.

 

[blueone]

Yes, absolutely true. Their process research is for their own chips, both mainframe and POWER.

I read a year or so ago they were working with Intel now, so maybe this will be the end of their relationship with Samsung for the near future.

The Power10 is my favorite CPU of its generation. I'm looking forward to seeing what the Power11 looks like. IMO, IBM made a mistake by just putting the Power architecture into open source. I think they should have spun out the entire engineering group to compete with the likes of Ampere.

 

[nghanayem]

My observations on IBM's semi manufacturing arm:
The common platform alliance has to my knowledge only ever had one success (Samsung's 14LPP). Every other node has failed so spectacularly that it's members had to leave the leading edge. UMC/IBM/GF's finFET's failed to materialize on time (UMC/IBM) or at all (GF). For HKMG they went for a scheme that didn't work. SOI was too expensive for how little the benefits were, and to this day only has niche use cases. Since 95% of the team went to GF after the acquisition, all that remains at IBM today is basically a REALLY BIG university research org that also helps Samsung with component's research/pre-pathfinding work. As for GF: by my count, they have never once developed and ramped a high volume finFET technology.

So as a member of academia I would say IBM is great. I would even go so far to say they are almost IMEC levels of good. However, as a manufacturer IBM is subpar. As for the IBM influence at intel we will have to see. Considering they aren't using a BPR like IBM and are using 4 nano sheets rather than IBM's 3, it doesn't seem like it is anything major. But there are certainly tells we should be able to see once we can get 20A teardowns, and we compare them to IBM's 2nm integration flow paper.

 

[Barnsley]

Was at a Meeting with STmicro on Friday and their expanisons at Crolles /Agrate and Catania sites look very impressive. Done modular style parts already up and running no need to wait for last screw to be tightened in order for production to start.
It look very different to the UMC expansion here in Singapore.

WIth respect to the GF collaboration as stated in intial post here they done appear to be looking beyond 18nm tech at the minute with automotive being primary driver.

They are also pushing Silicon Carbide heavily a lot.

All in all they look to be doing quite well at this point in time.

 

[ta152h]

My observations on IBM's semi manufacturing arm:
The common platform alliance has to my knowledge only ever had one success (Samsung's 14LPP). Every other node has failed so spectacularly that it's members had to leave the leading edge. UMC/IBM/GF's finFET's failed to materialize on time (UMC/IBM) or at all (GF). For HKMG they went for a scheme that didn't work. SOI was too expensive for how little the benefits were, and to this day only has niche use cases. Since 95% of the team went to GF after the acquisition, all that remains at IBM today is basically a REALLY BIG university research org that also helps Samsung with component's research/pre-pathfinding work. As for GF: by my count, they have never once developed and ramped a high volume finFET technology.

So as a member of academia I would say IBM is great. I would even go so far to say they are almost IMEC levels of good. However, as a manufacturer IBM is subpar. As for the IBM influence at intel we will have to see. Considering they aren't using a BPR like IBM and are using 4 nano sheets rather than IBM's 3, it doesn't seem like it is anything major. But there are certainly tells we should be able to see once we can get 20A teardowns, and we compare them to IBM's 2nm integration flow paper.

IBM left manufacturing years ago, as you know.

When you talk about little benefit, you have to keep in mind, mainframes aren't really cost sensitive, and yes, IBM does not do research for mainstream nodes or processors. They are high-end servers, and particularly with mainframes, extremely expensive. They don't have to make the same compromises other companies do, given the products they sell.

With that in mind, and how successful and powerful mainframes are, I have a difficult time not seeing them as successful at what they do.

Now, if you want to tell me the incremental gains they benefit from aren't worth the cost, I would not be able to say you're wrong; I don't know. I think possibly with Intel's fairly high performance nodes, maybe that argument is stronger, whereas with TSMC and Samsung, they probably feel they need a custom node to meet their customer's expectations. And likely TSMC wouldn't be too interested anyway, because their more "mainstream" nodes are far better for their business.

I think Intel and IBM are in the same boat though. They both have to (well, IBM maybe doesn't) pay for the development costs for a new node, and their own products don't necessarily validate the costs associated with them. Since Intel has to make a lot of parts for their own purposes, it makes sense they split the costs by making products for other companies as well. So they share their expense that way. It looks like IBM wants to share their burden by licensing their technology. But, I have to wonder, how much additional work/cost will be needed to make more mainstream parts, rather than super high-performance/high-reliability processors that expect exotic cooling. It's not my field of expertise, so maybe it isn't that big of a deal, but it's difficult to see the same process used to make mainframe parts, used to make lower power ones, for phones, cars, appliances, etc... without significant modification.

 

[Nokokara]

Curious - as 7nm is becoming “older”, is there any chance we’ll see companies like STMicro or Global Foundries resume progress towards denser nodes?

I saw with their newest joint fab they talked about 18nm being ‘relevant in automotive for decades’, but IMO that’s also based on legacy automaker engineering/thinking with hundreds of suppliers providing separate parts, which is becoming less viable over time.

Currently no « chance » (perspective…) for STMicroelectronics to go into 7nm or denser nodes…

STMicroelectronics was initially an historical pure IDM but now, since many years, they also become a fabless company (like Infineon and some other ones…) for part of their products in a bit (not a lot…) more advanced technologies than the ones which they produce internally…
Fabless was initially concerning technologies below 65nm and now STMicroelectronics included in 2012 in their Crolles fab the FD-SOI 28nm but, even in their fabless activity for the markets/applications in which they are focusing, they don’t need/use yet any current most advanced technologies below 22 ~ 14/10 nm…

 

[nghanayem]

IBM left manufacturing years ago, as you know.

When you talk about little benefit, you have to keep in mind, mainframes aren't really cost sensitive, and yes, IBM does not do research for mainstream nodes or processors.

No SOI was always an inferior tech to peddle (at least as a GP or HPC technology). Real shame that was too, because I think it is neat. Reading through some old EE times articles and seeing when IBM started their SOI work. It was always IBM's solution to extend planar transistors, and it always costed more and offered worse performance characteristics than the technologies that the industry actually adopted to extend the life planar devices and eventually leave them behind. Additionally IBM does not even use SOI anymore, as they moved to the cheaper and higher performing finFETs just like everybody else. Additionally ever since 7LPP IBM's mainframes have been using the same mainstream Samsung nodes that everyone else uses.

They are high-end servers, and particularly with mainframes, extremely expensive. They don't have to make the same compromises other companies do, given the products they sell.

With that in mind, and how successful and powerful mainframes are, I have a difficult time not seeing them as successful at what they do.

They tried to be the industry's northstar, and when the last thing they did that was truly successful was copper damascene's to the tune of 30 years ago; I would say that is a failure to hit that goal. Mind you this is not a failure in chip design or software (IBM is without a doubt good at those things). But as I said their manufacturing arm is to me less impressive (in no small part because it has literally ceased to exist and all that is left is the university lab they share with the state of NY).

Now, if you want to tell me the incremental gains they benefit from aren't worth the cost, I would not be able to say you're wrong; I don't know. I think possibly with Intel's fairly high performance nodes, maybe that argument is stronger, whereas with TSMC and Samsung, they probably feel they need a custom node to meet their customer's expectations. And likely TSMC wouldn't be too interested anyway, because their more "mainstream" nodes are far better for their business.

For 90nm intel claimed that strained silicon was a 2% cost adder and gave like 10-20% performance. Off the top of my head at 22nm intel said that if they wanted to do a planar SOI node it would have cost like 15% extra, meanwhile finFET was only an 8% cost adder. And I don't think anyone in their right mind would say 22nm (or even 16FF and 14LPP) had worse performance than GF 22FDX (which also seems to use more power, came out much later, used FDSOI rahter than SOI, and had worse density).

I think Intel and IBM are in the same boat though. They both have to (well, IBM maybe doesn't) pay for the development costs for a new node, and their own products don't necessarily validate the costs associated with them. Since Intel has to make a lot of parts for their own purposes, it makes sense they split the costs by making products for other companies as well. So they share their expense that way. It looks like IBM wants to share their burden by licensing their technology. But, I have to wonder, how much additional work/cost will be needed to make more mainstream parts, rather than super high-performance/high-reliability processors that expect exotic cooling. It's not my field of expertise, so maybe it isn't that big of a deal, but it's difficult to see the same process used to make mainframe parts, used to make lower power ones, for phones, cars, appliances, etc... without significant modification.

IBM's tech isn't designed for that anymore (as the mainframes get built on Samsung's GP tech). IBM nodes are now just technology demonstration vehicles. Rapidus is the real test of that theory though. How much does the design have to change from IBM's prototype? How many issues do they run into to make the tech HVM ready? What work (if any) needs to be done to adapt this node for LP? As for process tuning

I think TSMC and intel are two good angles to look at. Back when intel was making mobile chips the CPU portions didn't seem that bad compared to the ARM chips at the time. However based on how folks on the internet talk about them it sounds like the SOCs as a whole were more lackluster. Now when you look at the process tech used Intel phone chips were about 1 years ahead on density and like 2-3 years up on technology over Samsung, and about 0 years on density and one on tech vs TSMC. Considering how intel seemed to have designed everything as a discrete component rather than an integrated SOC, and without having a huge process lead to me says that intel 45/32nm could punch down to mobile phone power envelops. While hard to know what the exact impact of all these factors are, it seems it couldn't punch down quite as well as 40/28nm. TSMC on the other hand is now punching up with the X nodes, and these by all accounts seem to be going pretty well. Looking at N3 HPC, N3X, and intel 3 server parts will probably give us more insight into the X nodes and the smaller tweaks TSMC needs to do to enable HPC products.

 

[Daniel Nenni]

I was under the impression their processes are behind the CPUs in their mainframes, and POWER based computers. At least the former. And they reach some crazy clock speeds. And they just licensed their tech to Rapidus. So, clearly they felt it was something worth licensing and using.

IBM has announced many semiconductor technology partnerships that have failed to bear fruit. That has been my experience. I may be wrong about this whole thing but what I can tell you is that the amount of wafers IBM needs for the servers is miniscule in comparison to mainstream semiconductor companies. It just would not be feasible to have an IBM specific process like TSMC does for Apple.

The Samsung/IBM collaboration is 20+ years in the making so it has had many incarnations. True, the Common Platform collaboration was a spectacular failure. IBM and GF also collaborated on a 7nm process that never made it into production and resulted in GF's pivot away from leading edge process technologies.

I will have to do more digging but I do know that IBM gets wafers from Samsung and it is not a specialized process. It is Samsung 7nm HP like everyone else uses. I have not heard about IBM using Samsung 5nm or 3nm.

I did read about an Intel/IBM relationship but know little about it. It would make much more sense for IBM to get wafers from Intel due to the stumbling of Samsung at 5nm, and 3nm and Intel's high performance expertise, but I do not see Intel licensing IBM process technology. It would be a clash of company cultures, a made for TV movie type of thing or maybe more like a Real Housewives of xxx realty TV type of thing.

And here is a rumor we can start right now: IBM is responsible for Samsung's problems at 5nm and 3nm.

 

[ta152h]

IBM has announced many semiconductor technology partnerships that have failed to bear fruit. That has been my experience. I may be wrong about this whole thing but what I can tell you is that the amount of wafers IBM needs for the servers is miniscule in comparison to mainstream semiconductor companies. It just would not be feasible to have an IBM specific process like TSMC does for Apple.

The Samsung/IBM collaboration is 20+ years in the making so it has had many incarnations. True, the Common Platform collaboration was a spectacular failure. IBM and GF also collaborated on a 7nm process that never made it into production and resulted in GF's pivot away from leading edge process technologies.

I will have to do more digging but I do know that IBM gets wafers from Samsung and it is not a specialized process. It is Samsung 7nm HP like everyone else uses. I have not heard about IBM using Samsung 5nm or 3nm.

I did read about an Intel/IBM relationship but know little about it. It would make much more sense for IBM to get wafers from Intel due to the stumbling of Samsung at 5nm, and 3nm and Intel's high performance expertise, but I do not see Intel licensing IBM process technology. It would be a clash of company cultures, a made for TV movie type of thing or maybe more like a Real Housewives of xxx realty TV type of thing.

And here is a rumor we can start right now: IBM is responsible for Samsung's problems at 5nm and 3nm.

So why are they spending so much money on advancing process technology, with their own research if they aren't using it for their own products? There's got to be something we're missing, because it doesn't make sense why they'd continue on this path. It's not cheap, as you know.

Samsung has problems regardless of what or who is involved. That's a company I never have quite understood. It doesn't seem they're ever really good at anything; their CPU designs are shockingly poor for what they are as well. They still can't figure out how to make an even half-competent GPU, and have to license technology. I don't really get them, at all. Even mediocrity would be a welcome step up for them. And now they're going to compete against not just TSMC, and Intel. They better step up their game, and by that I don't mean press releases.