Discussion Leading Edge Foundry Node advances (TSMC, Samsung Foundry, Intel)

[Doug S]

Future chips are going to look really weird due to cache density not increasing as well as logic density. They will have a huge (like 80% of die area) blob of SRAM in the middle surrounded by teensy weensy chiplets for CPU, GPU and I/O.

Vertical transistors (i.e. nanosheets etc.) and backside power delivery resolve a lot of the SRAM cell scaling issues, so this is a temporary problem.

Of course this problem will be replaced by some other problem, they always are.

 

[DisEnchantment]

According to Samsung, their 4nm results in 11% perf increase, 24% area reduction, and 16% power reduction compared to their 5nm.

Not sure where you pulled that data from.
You can estimate density from published transistor dimension from Samsung SAFE Forum.
Since most of us don't have SAFE access (I assume) go to the next best thing below

A Look At Samsung's 4LPE Process

A look at Samsung's last leading-edge FinFET process node, 4nm 4LPE.

fuse.wikichip.org

137 MTr/mm2 for standard cells using 2x2 pMOS and nMOS devices.
5LPE --> 4LPE ~1.08x density gain, not 1.3x gain (i.e. 24% area reduction is plain wrong, did you ponder about it?). Also efficiency gain is lesser.

It is obvious that Samsung's 4LPE of last year is much worse than N4 or Intel 4 of the future.
4LPP, Samsung's 4nm node of 2022, improves on these metrics albeit on an even smaller scale.

Also based on yields of previous nodes 3GAE is probably terrible as in 10-20% yield.

I hope you are not believing Digitimes 100%, because in their eyes TSMC is best on every possible metric.

5LPE/4LPE is much behind N5 in efficiency, naturally it is not the first preference for mobile SoCs from QCM et al.
Therefore the 45% efficiency gain from 3GAE is crucial for Samsung, for chips they can fab right now.
For 2023 they will have 3GAP and guaranteed in 2024 they will iterate on it.

 

[IntelUser2000]

Not sure where you pulled that data from.
You can estimate density from published transistor dimension from Samsung SAFE Forum.

https://www.gizchina.com/wp-content/uploads/images/2021/07/Samsung-3nm-process-b-1536x714.jpg

I hope you are not believing Digitimes 100%, because in their eyes TSMC is best on every possible metric.

I didn't believe in DigiTimes articles 10+ years ago. I test rumors/leaks to see if they are right to determine how reliable they are. They are not.

No, they are from various articles that source directly from Korean newspapers and other reputable sources. There's a reason the rumors around mobile RDNA2 had it go from 1.7 to 1.3GHz.

5LPE/4LPE is much behind N5 in efficiency, naturally it is not the first preference for mobile SoCs from QCM et al.
Therefore the 45% efficiency gain from 3GAE is crucial for Samsung, for chips they can fab right now.
For 2023 they will have 3GAP and guaranteed in 2024 they will iterate on it.

Yea it was when they first announced it I criticized them for changing numbers to make it seem they are advancing faster than they really are. 6/5/4 their meanings are significantly diluted even from the already diluted numbers everyone is using. If TSMC/Intel and others resorted to calling slightly more than half nodes as full nodes(post 20nm gen), Samsung basically treats slightly more than half node in the post-20nm meaning as being a full node.

They could have compared to direct predecessors but they did not. Because they knew people would be suspicious if you claimed GAA but with muted gains. You'd need such gains just to catch up considering a major manufacturer decided to switch fab suppliers mid cycle! And that would be needed jumping from "4nm" not 5 as they claim.

 

[DisEnchantment]

https://www.gizchina.com/wp-content/uploads/images/2021/07/Samsung-3nm-process-b-1536x714.jpg

I didn't believe in DigiTimes articles 10+ years ago. I test rumors/leaks to see if they are right to determine how reliable they are. They are not.

No, they are from various articles that source directly from Korean newspapers and other reputable sources. There's a reason the rumors around mobile RDNA2 had it go from 1.7 to 1.3GHz.



Yea it was when they first announced it I criticized them for changing numbers to make it seem they are advancing faster than they really are. 6/5/4 their meanings are significantly diluted even from the already diluted numbers everyone is using. If TSMC/Intel and others resorted to calling slightly more than half nodes as full nodes(post 20nm gen), Samsung basically treats slightly more than half node in the post-20nm meaning as being a full node.

They could have compared to direct predecessors but they did not. Because they knew people would be suspicious if you claimed GAA but with muted gains. You'd need such gains just to catch up considering a major manufacturer decided to switch fab suppliers mid cycle! And that would be needed jumping from "4nm" not 5 as they claim.

Samsung's node names have zero meaning, but the slide you shown would mean for 4LPP. Basically they are doing this regularly.
3nm figures announced by Samsung is actually for 3GAP. 3GAE does not reach the advertised gains vs 5LPE/LPP.
4LPE does not reach the advertised gains vs 5LPE. 4LPP is the one which is actually close to that.

There is no 4LPP device out yet, 4LPP go into production around same time like 3GAE.
Samsung is >15 months behind TSMC on FinFET. But on GAA seems like they got a head start.

 

[moinmoin]

It will be in ALL areas not just all semi because of the shipping delays/forced lockdowns etc. Look how much building material pricing has increased, or the cost of food. Don't forget about housing and gas!

Prices for lumber fell again though, so more houses built using only wood, wooooh! (My granddad did one, they are nice.)

 

[Lodix]

Remember at a product-level TSMC 4nm was about 7% faster and offered 23% higher perf/watt while the GPU was about 10% faster and offered 47% higher perf/watt and we're talking device level power consumption.

I think TSMC N4 is enough to compete with the second-generation 3GAP in all areas.

Also based on yields of previous nodes 3GAE is probably terrible as in 10-20% yield.

Where are you taking your numbers for TSMC being 47% more efficient? Compared to what?

The only current product manufactured at Samsung and TSMC 4nm is the Snapdragon 8 Gen 1. The tweaked version and made in TSMC is UP TO 30% more efficient. And maybe they did other changes apart from the process node.

Also "10-20% yield" LoL wtf. Stop with that BS.

 

[NTMBK]

https://www.tomshardware.com/news/us-wants-china-smic-to-stop-making-14nm-chips

The US continues to put the squeeze on Chinese foundries. That's one way for Intel to be competitive- ask your friends in the government to shut down the emerging Chinese rivals!

 

[SteinFG]

I'm kind of uneducated in this area, but after thinking a bit, isn't gaafet/nanosheets overrated? At first I thought - cool, stacking layers. But then it dawned on me - they still share the same gate, meaning no difference in the number of transistors. So it's all just in the name of efficiency, not density gain. Is there other benefits?

 

[Exist50]

I'm kind of uneducated in this area, but after thinking a bit, isn't gaafet/nanosheets overrated? At first I thought - cool, stacking layers. But then it dawned on me - they still share the same gate, meaning no difference in the number of transistors. So it's all just in the name of efficiency, not density gain. Is there other benefits?

You're correct for first gen GAA, but eventually we'll probably see "Complementary FETs" with NMOS and PMOS stacked on top of each other.

 

[DrMrLordX]

The US continues to put the squeeze on Chinese foundries. That's one way for Intel to be competitive- ask your friends in the government to shut down the emerging Chinese rivals!

It helped Qualcomm and Apple fend off competition from Huawei. Not that SMIC amounts to much of a threat (yet).

 

[Doug S]

It helped Qualcomm and Apple fend off competition from Huawei. Not that SMIC amounts to much of a threat (yet).

Huawei wasn't affecting Apple's market share, they were hurting Samsung.

 

[NTMBK]

Huawei wasn't affecting Apple's market share, they were hurting Samsung.

Inside China they were definitely hurting Apple.

 

[DrMrLordX]

Huawei wasn't affecting Apple's market share, they were hurting Samsung.

Samsung? Not really. Samsung was hurting themselves with custom Exynos SoCs that weren't as performant as reference ARM designs. Huawei's designs (like Kirin 970) were plentiful, cheap, and very competitive with Snapdragon chips.

 

[NostaSeronx]

HLMC's leading edge and SMIC's leading edge fabs both share SICRD: "The Shanghai Integrated Circuit R&D Center is located in Zhangjiang Hi-Tech Park, Shanghai, adjacent to Huali Microelectronics and SMIC."

HLMC has delayed 14nm introduction to swap out to a full Chinese EDA/Tools path. Which SMIC is likely to move towards as well. If they get DUV tools blocked for expansion of 14nm/12nm, 10nm/8nm, and 7nm nodes.

28nm/22nm is currently the only nodes fully internal.
14nm is set to have tools out by end of this year, but I'll be safe and say 2023 is when 14nm will be out.
7nm internal tools should be 2025, but 2026 is also safe bet.

SMIC is very much still trying to hold out that they are a global-class fab. Which basically means that they have to share ASML/US-style EDA with everyone else.

 

[Doug S]

Inside China they were definitely hurting Apple.

Other than a one time blip in 2015 due to Apple finally releasing bigger phones (which remember were first popularized in China by a few no name brands before Samsung followed their lead, so there clearly would have been pent up demand) Apple's market share remained pretty steady in the 9% to 11% range during Huawei's ascent in the Chinese market. If Huawei going away led to Apple gaining share, why did Huawei going from 16% to 38% between 2016 and 2019 have almost no effect on Apple's share?

Apple went from 11% to 15% in 2020 to 2021 while Huawei dropped from 38% to 12% (under the Honor brand in 2021) so Chinese OEMs gained almost all of what Huawei lost.

It is even debatable whether Apple's gain in 2021 was due to Huawei's loss, or because Apple introduced 5G iPhones in late 2020. Since China has deployed 5G much more quickly than in the US and EU, there's a good chance 2021 is just another blip like 2015 from pent up demand for a new feature, and Apple will return to their 9-11% range now that the pent up demand has been satisfied.

https://www.statista.com/statistics/528833/smartphone-market-share-in-china-by-vendor/

 

[Frenetic Pony]

Samsung? Not really. Samsung was hurting themselves with custom Exynos SoCs that weren't as performant as reference ARM designs. Huawei's designs (like Kirin 970) were plentiful, cheap, and very competitive with Snapdragon chips.

Never felt like Samsung wanted to shell out for a competitive SOC. Even this last effort, sure the AMD gpu they modified was technically slightly better than Qualcomm's. But they flubbed even that by throttling the performance on the Exynos versions but leaving the Qualcomm versions of the same phone unlocked, leaving lazy headline seeking reviewers to declare Samsung efforts worthless. Not that it mattered anyway because other than the GPU the SOC was all around worse than Qualcomm's so it was L regardless.

It's pretty sad that the non Iphone mobile scene has just descended into a Qualcomm monopoly even while the bigger TDP market is heating up in all kinds of ways. They don't seem to care about anything there other than whatever minimum allows them to maintain that monopoly and sees revenue keep coming in.

 

[DrMrLordX]

It's pretty sad that the non Iphone mobile scene has just descended into a Qualcomm monopoly even while the bigger TDP market is heating up in all kinds of ways.

Mediatek to the rescue . . . maybe.

 

[hemedans]

Other than a one time blip in 2015 due to Apple finally releasing bigger phones (which remember were first popularized in China by a few no name brands before Samsung followed their lead, so there clearly would have been pent up demand) Apple's market share remained pretty steady in the 9% to 11% range during Huawei's ascent in the Chinese market. If Huawei going away led to Apple gaining share, why did Huawei going from 16% to 38% between 2016 and 2019 have almost no effect on Apple's share?

Apple went from 11% to 15% in 2020 to 2021 while Huawei dropped from 38% to 12% (under the Honor brand in 2021) so Chinese OEMs gained almost all of what Huawei lost.

It is even debatable whether Apple's gain in 2021 was due to Huawei's loss, or because Apple introduced 5G iPhones in late 2020. Since China has deployed 5G much more quickly than in the US and EU, there's a good chance 2021 is just another blip like 2015 from pent up demand for a new feature, and Apple will return to their 9-11% range now that the pent up demand has been satisfied.

https://www.statista.com/statistics/528833/smartphone-market-share-in-china-by-vendor/

Samsung had Big phones way before Chinese phones, since Galaxy note 1, 2011 and Dell was first to start this trend with Streak 5 in 2010.

About sales Apple did benefit from Huawei loss, Apple only sell highend phones so its natural for them not to take all Huawei marketshare, they just replace Those P and Mate series, while Other Chinese brands replace mid and low end.

As of now no Chinese brand can sell flagship as many as Huawei did.

 

[Doug S]

Samsung had Big phones way before Chinese phones, since Galaxy note 1, 2011 and Dell was first to start this trend with Streak 5 in 2010.

About sales Apple did benefit from Huawei loss, Apple only sell highend phones so its natural for them not to take all Huawei marketshare, they just replace Those P and Mate series, while Other Chinese brands replace mid and low end.

As of now no Chinese brand can sell flagship as many as Huawei did.

That's absolutely not true, I remember seeing an article about the big phones appearing in China BEFORE the Galaxy Note.

 

[Ajay]

Freaking politics.

https://www.tomshardware.com/news/intel-ohio-fab-europe

Intel has made it clear that its expansion plans in the US partially hinge on government subsidies. Speaking live with the Washington Post earlier this week, Intel CEO Pat Gelsinger made it clear that if the US Innovation and Competition Act (or USICA) isn't approved on the floor of Congress, the company will be seeking greener pastures across the Atlantic - namely in Europe, where the EU has been aggressively moving to increase local leading-edge chip manufacturing.
"We've made super clear to McConnell, to the Democrats, to the Republicans, that if this doesn't pass, I will change my plans," Gelsinger said. "The Europeans have moved forward very aggressively, and they're ready to give us the incentives that allow us to move forward," he said, likely referring to the Silicon Junction initiative, which saw the EU open up its strings to a cool $43 billion in support for local manufacturing capabilities.
The measure is currently languishing on the Congress floor due to differences in opinion between Republicans and Democrats. The USICA bill would allocate up to $52 billion in incentives for semiconductor initiatives on US soil, and saw the creation of the Semiconductors in America Coalition (SIAC) lobbying group.
The point of contention mostly lies in that Democrats could attempt to approve it as part of a broader reconciliation package that would allow Medicare to negotiate prices with drug manufacturers, which would help control prescription drug prices while putting a cap on insulin prices {..snip..}

NOT MAKING THIS A P&N THREAD. Just posting an update on the hold up, including the relevant politic sticking point which may not be known to our international posters. Please limit comments to semiconductor competitiveness.

 

[jpiniero]

IMO they should call Intel's bluff. Not that they won't build without subsidies, but that Europe will pay up.

 

[CakeMonster]

I wouldn't mind some money thrown around, as long as it results in more advanced fabs in more allied countries spread around for good measure (and for creating several science hubs in the future). New physical locations. It would suck if this ends up doing nothing for research in the in the long run, and if the advanced process nodes in the generation after the next doesn't benefit from it either.

 

[Saylick]

IMO they should call Intel's bluff. Not that they won't build without subsidies, but that Europe will pay up.

Agreed. Pat is just trying to use Europe as leverage over the US and it is in his best interest to have subsidy offers in both regions since it gives him the most financial assistance. It's not like he has to choose one region over the other either, because Intel can totally build in both, so there's little reason why the US should get into a bidding war with Europe. If Pat says that he is only building one megafab and the build site is in the region which provides him with the most subsidies, he'd be lying because building in one region is counter to the whole idea of geographic diversification.

 

[jpiniero]

I wouldn't mind some money thrown around, as long as it results in more advanced fabs in more allied countries spread around for good measure (and for creating several science hubs in the future). New physical locations. It would suck if this ends up doing nothing for research in the in the long run, and if the advanced process nodes in the generation after the next doesn't benefit from it either.

These fabs are (I'm assuming) not R&D fabs. These are just adding more capacity.

 

[Markfw]

I don't care where they build, so long as its not communist or dictator ruled. Diversification will give EVERYONE more security and capacity.