Discussion Leading Edge Foundry Node advances (TSMC, Samsung Foundry, Intel) - [2020 - 2025]

[DisEnchantment]

TSMC's N7 EUV is now in its second year of production and N5 is contributing to revenue for TSMC this quarter. N3 is scheduled for 2022 and I believe they have a good chance to reach that target.


N7 performance is more or less understood.


This year and next year TSMC is mainly increasing capacity to meet demands.

AnandTech Forums: Technology, Hardware, Software, and Deals

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

www.anandtech.com

For Samsung the nodes are basically the same from 7LPP to 4 LPE, they just add incremental scaling boosters while the bulk of the tech is the same.

Samsung is already shipping 7LPP and will ship 6LPP in H2. Hopefully they fix any issues if at all.
They have two more intermediate nodes in between before going to 3GAE, most likely 5LPE will ship next year but for 4LPE it will probably be back to back with 3GAA since 3GAA is a parallel development with 7LPP enhancements.

AnandTech Forums: Technology, Hardware, Software, and Deals

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

www.anandtech.com

Samsung's 3GAA will go for HVM in 2022 most likely, similar timeframe to TSMC's N3.
There are major differences in how the transistor will be fabricated due to the GAA but density for sure Samsung will be behind N3.
But there might be advantages for Samsung with regards to power and performance, so it may be better suited for some applications.
But for now we don't know how much of this is true and we can only rely on the marketing material.

This year there should be a lot more available wafers due to lack of demand from Smartphone vendors and increased capacity from TSMC and Samsung.
Lots of SoCs which dont need to be top end will be fabbed with N7 or 7LPP/6LPP instead of N5, so there will be lots of wafers around.

Most of the current 7nm designs are far from the advertized density from TSMC and Samsung. There is still potential for density increase compared to currently shipping products.
N5 is going to be the leading foundry node for the next couple of years.

For a lot of fabless companies out there, the processes and capacity available are quite good.

---------------------------------------------------------------------------------------------------------------------------------------------------


FEEL FREE TO CREATE A NEW THREAD FOR 2025+ OUTLOOK, I WILL LINK IT HERE

 

[Markfw]

The weirdest thing I’ve noticed on the tech enthusiast internet is the pro TSMC monopoly stance.

The current situation is arguably horrible for both tech advancement and geopolitics. I imagine some of it is people that dislike Intel and view TSMC as an extension of AMD.

Personally, I don't listen to any of this. I evaluate the products that come out of these foundries and use that to evaluate their performance. I could be wrong, but at the moment TSMC does appear to have better products out right now.

Please correct me if I am wrong (with examples/proof). The future ?? no bets on either of them.

 

[H433x0n]

Personally, I don't listen to any of this. I evaluate the products that come out of these foundries and use that to evaluate their performance. I could be wrong, but at the moment TSMC does appear to have better products out right now.

Please correct me if I am wrong (with examples/proof). The future ?? no bets on either of them.

At the moment TSMC has better process tech. My hot take is that era has already ended and will become apparent in the next 12-18 months.

I believe we conflate the success of individual companies designs with process tech quite a bit. An excellent example of this is Lovelace and RDNA3, both using the same nodes but one is much more efficient. I believe this is probably the case with Zen 4 (N5) & RPL (Intel 7) too.

 

[Doug S]

The weirdest thing I’ve noticed on the tech enthusiast internet is the pro TSMC monopoly stance.

The current situation is arguably horrible for both tech advancement and geopolitics. I imagine some of it is people that dislike Intel and view TSMC as an extension of AMD.

Intel was happily a monopolist for decades, but they got too fat and happy and made tons of stupid decisions like telling Steve Jobs they weren't interested in making chips for his crazy idea of an Apple phone unless it was x86, which was utterly unsuitable for such a low power device. Then they doubled down on their short sightedness by bribing Android OEMs into using terrible x86 chips in tablets, because Intel execs were so narrow minded they wouldn't consider opening up their (then) world leading fabs to anything that dared to not use x86.

Now that they've had their comeuppance we're supposed to cheer for them because they're a US company? Or because they break TSMC's monopoly? Which I agree is bad from a "what if there's a huge earthquake or China attacks" black swan risk but has had no issues from a "tech advancement" standpoint. You have to be "king" for longer than TSMC has been (but obviously not as long as Intel was) before the evil side of monopoly starts to show through and you stop trying so hard because you know you don't have to.

I have said here many times I think it would be great if Intel was competitive with TSMC - as a TRUE foundry not just an x86 chip company that toys around with certain clients but has no real separation between the x86 business and the fab business. I have also said here many times I remain skeptical Intel is suddenly going to be back on track just because they announced an aggressive roadmap. My skepticism only increases since they're already behind on that roadmap, given that they won't (finally!) be shipping Intel 4 chips (their very first EUV) until the third to last week of 2023. I don't care that they claimed "HVM" a year ago, that claim was meaningless. The ONLY thing that matters is when chips are delivered into customer hands in mass quantities, not when you claim a milestone or "ship" something it is almost impossible for a normal person to buy (like Intel's 10nm fairy story in 2018, which was the point when I stopped believing anything they said)

I don't think Intel can ever get to where the industry - and the United States - needs them to be without spinning off the fab business. They can't be selling x86 chips and acting as a foundry out of the same company. No chance AMD would consider them as a foundry unless the foundry is a separate business. Probably a lot of the really big companies they would want a shot to win (Apple, Qualcomm, Nvidia) would have some deep reservations that if there was a shortage and it came down to deciding who gets more of a limited number of wafers, the "internal customer" would somehow always win. Hopefully they are paving the way towards that, because even if they catch up to TSMC, even if they take back the lead, that's a hollow victory for the people who say "oh we need a strong Intel for competition" if they think they can go back to the days when their fabs were a competitive advantage in the x86 market and have a few token clients so they can claim "we're a foundry". They played that game once, we won't get fooled again.

 

[DrMrLordX]

Nonsense. If you call an upcoming Intel node vaporware you have to call an upcoming TSMC node vaporware as well unless you are biased. You don't like this right because TSMC is the good guy and you hate Intel right? I got it.

No, because 18a isn't going to be available even to internal products in the same timeframe that N3P will be available to foundry customers.

 

[Ajay]

I don't think Intel can ever get to where the industry - and the United States - needs them to be without spinning off the fab business. They can't be selling x86 chips and acting as a foundry out of the same company. No chance AMD would consider them as a foundry unless the foundry is a separate business. Probably a lot of the really big companies they would want a shot to win (Apple, Qualcomm, Nvidia) would have some deep reservations that if there was a shortage and it came down to deciding who gets more of a limited number of wafers, the "internal customer" would somehow always win. Hopefully they are paving the way towards that, because even if they catch up to TSMC, even if they take back the lead, that's a hollow victory for the people who say "oh we need a strong Intel for competition" if they think they can go back to the days when their fabs were a competitive advantage in the x86 market and have a few token clients so they can claim "we're a foundry". They played that game once, we won't get fooled again.

I agree, and I wouldn’t have two years ago. You, and others, are well ahead of me. The people who knew this business well, saw this reality five years ago. The only advantage left for Intel is that they don’t compete for wafer starts, as you mentioned. Every other presumed advantage has become a liability. In hindsight, the failure of Intel's 10nm process completely wrecked their entire lineup and the processor designs too tightly tied to the specific 'expected' characteristics of that node - making a back port problematic. It also exposed both their hubris and broken risk management system. Undoubtedly, Intel would not have gone as aggressively with 10nm if they risked customer revenues that were greater than their own sales of products on that same node. Unless they were even more arrogant than I can imagine. Their R&D for process node development is spread across too few wafer starts as compared to TSMC. A serious competitive disadvantage that if left uncorrected will cause their business model to collapse.

So, the only reasonable option for Intel is for IDF to be completely spun off as an independent company after a hopefully short transition period. A competitive US headquartered pure foundry would be very welcome in a time when reshoring is accelerating in the face of growing geopolitical instability. I wish them the best and hope the Intel board understands the stakes and remain committed to the reinvention of the company.

Just sharing my 2 cents in the wake of finally grasping some of the bigger picture.

 

[capricornsunvoyager2]

The X3 is a huge jump from the X1 though, and the four A715s if scheduled for background tasks properly will be a significant jump on performance/watt and energy efficiency vs the two A78’s and the two X1’s + four A55’s.

Remember for Google they used two X1’s and those X1’s are probably less energy efficient than newer A715’s on whole SoC power for some modestly clocked stuff. They’re going from 2 X1, 2 A78 to 1 X3, 4 A715, I do think we should see a jump in energy efficiency across the board from IP alone — background tasks will no longer need to be on an A55 or X core.

But it seems that, even within the TSMC 5nm, Both X1 and the A78 where the most energy-efficient core designs yet on their markets. This can change with the X4 and a720 but will remain to be seen.

And I got that's why they kept using them for the Tensor.

What they lose is in the ARM9.1 instruction set, and features and GPU compatibility.

 

[repoman27]

When was the last time anyone shipped a ~400 mm² SoC/CPU/GPU within 10 months of a new process node (not an inter-node optimization) reaching volume production/HVM? I'm sure there's precedent for what Apple and TSMC just did, but nothing in recent history is coming to mind...

 

[Hitman928]

When was the last time anyone shipped a ~400 mm² SoC/CPU/GPU within 10 months of a new process node (not an inter-node optimization) reaching volume production/HVM? I'm sure there's precedent for what Apple and TSMC just did, but nothing in recent history is coming to mind.

Ok, sorry, edited this post twice now but I believe Radeon VII from AMD which was 331 mm2. Not quite 400 mm2 but pretty large. Someone can double check my timeline though as I confused myself when I went to double check on it at first, lol.

 

[Geddagod]

When was the last time anyone shipped a ~400 mm² SoC/CPU/GPU within 10 months of a new process node (not an inter-node optimization) reaching volume production/HVM? I'm sure there's precedent for what Apple and TSMC just did, but nothing in recent history is coming to mind...

To be fair, TSMC 3nm was also delayed lol

 

[repoman27]

Ok, sorry, edited this post twice now but I believe Radeon VII from AMD which was 331 mm2. Not quite 400 mm2 but pretty large. Someone can double check my timeline though as I confused myself when I went to double check on it at first, lol.

That's pretty darn close, because TSMC N7 entered volume production in Q2'18, and Vega 20 actually launched in November 2018 with the Radeon Instinct MI60. Slightly smaller chip but an even quicker ramp, and a slightly different compare because the initial product wasn't consumer focused.

To be fair, TSMC 3nm was also delayed lol

Was it? TSMC always said H2'22, and it entered volume production in December 2022. And since it turned out to be an Apple only node, they didn't start volume production until Apple had fully qualified designs ready to go. Volume production can only happen if both the process and the customer are good to go, otherwise you're simply "manufacturing ready", like Intel 3.

 

[Doug S]

When was the last time anyone shipped a ~400 mm² SoC/CPU/GPU within 10 months of a new process node (not an inter-node optimization) reaching volume production/HVM? I'm sure there's precedent for what Apple and TSMC just did, but nothing in recent history is coming to mind...

Intel has shipped chips that size a year or less after a new node went online, so I don't think it is THAT remarkable. What difference does it make if it is a CPU only versus a CPU/GPU/SoC?

The only thing that surprises me about the M3 Max is that they're doing it with TSMC's problem plagued N3B process. With its reported yields of 50-60% (AFAIK they typically report such yield numbers based on a 100 mm^2 chip) the yield would be truly awful for a chip the size of M3 Max.

The yield would be so bad I have to wonder whether that yield rate is defect based. What if it isn't, what if defects aren't the problem for N3B but power is? If the pass/fail "known good die" binning was mostly to do with power then larger dies wouldn't be as much of an issue as they would be if N3B's yield problem has to do with defects.

 

[repoman27]

Intel has shipped chips that size a year or less after a new node went online, so I don't think it is THAT remarkable.

Certainly not in the last decade, which was sort of what I was driving at. When was the last time Intel actually did something like that?

What difference does it make if it is a CPU only versus a CPU/GPU/SoC?

None, I intended the slashes to be read as "or". Basically any chip with some degree of complexity, not an FPGA.

The only thing that surprises me about the M3 Max is that they're doing it with TSMC's problem plagued N3B process. With its reported yields of 50-60% (AFAIK they typically report such yield numbers based on a 100 mm^2 chip) the yield would be truly awful for a chip the size of M3 Max.

The yield would be so bad I have to wonder whether that yield rate is defect based. What if it isn't, what if defects aren't the problem for N3B but power is? If the pass/fail "known good die" binning was mostly to do with power then larger dies wouldn't be as much of an issue as they would be if N3B's yield problem has to do with defects.

Perhaps the problem is with where these "reported yields" are coming from. Why would anybody other than TSMC employees or customers know what the actual defect densities are? And seeing as Apple is the only customer for N3, what would motivate either stakeholder to share this kind of information with the public/media at this point?

FWIW, people who claim to have "seen the curves" say that defect densities are fine, but there's not much margin, so minor deviations from nominal can cause yields to take a dive.

 

[repoman27]

Someone reminded me about NVIDIA's Pascal lineup. Those were produced using TSMC's 16nm FinFET, which entered volume production in Q2'15.

The GP102 in the TITAN X was 471 mm² and shipped Aug 2, 2016, so just over a year out.

The GP100, a 610 mm² behemoth used in the various Tesla P100 products, was actually released first, in Q2'16. Those weren't even remotely consumer oriented products though.

 

[Doug S]

Certainly not in the last decade, which was sort of what I was driving at. When was the last time Intel actually did something like that?

Well 14nm came out almost a decade ago and that was when the first cracks in Intel's previous bulletproof process rollouts became apparent. Intel hasn't done that for a long time because they've been dysfunctional for a long time. TSMC has been able to do that (for Apple as their biggest customer who gets first crack at all the new nodes) because they've been executing well. Until N3B at least - if they had maintained their previous cadence that Apple had become used to then N3 would have come out a year ago and been used for A16.

 

[repoman27]

Well 14nm came out almost a decade ago and that was when the first cracks in Intel's previous bulletproof process rollouts became apparent. Intel hasn't done that for a long time because they've been dysfunctional for a long time. TSMC has been able to do that (for Apple as their biggest customer who gets first crack at all the new nodes) because they've been executing well. Until N3B at least - if they had maintained their previous cadence that Apple had become used to then N3 would have come out a year ago and been used for A16.

As far as I can tell, the answer is Intel never shipped anything that big in the first year on a new process. Because they didn't for 22nm, 32nm, or 45nm either, and I'm not sure what they would have been producing prior to that that would have approached that size, aside from Itanium. The Knights Corner Xeon Phi gets an honorable mention because it was 720 mm² and launched just over a year after 22nm reached HVM, but the initial production was also driven by a supercomputer order.

Apple and TSMC were in full agreement that N3 would be the 2023 node at least as far back as 2019. Would they have taken N3 sooner if it was ready? Sure. But they knew it wasn't and had N4 ready to go instead.

But yeah, hand-wave away.

 

[dr1337]

The weirdest thing I’ve noticed on the tech enthusiast internet is the pro TSMC monopoly stance.

The current situation is arguably horrible for both tech advancement and geopolitics. I imagine some of it is people that dislike Intel and view TSMC as an extension of AMD.

Yeah wouldn't it be great if all the fabs came together and joined into one? But seriously tho there isn't a true monopoly in semi-fab and TSMC only has cutting edge nodes for so long. Before their 7nm became mainstream, which was barely 4 years ago, they were considered on the same level as Samsung and thought to be behind Intel.

Calling TSMC as an extension of AMD and not Apple would say to me that you don't actually mean what you are saying, or that you are completely ignorant of reality.

 

[H433x0n]

Yeah wouldn't it be great if all the fabs came together and joined into one? But seriously tho there isn't a true monopoly in semi-fab and TSMC only has cutting edge nodes for so long. Before their 7nm became mainstream, which was barely 4 years ago, they were considered on the same level as Samsung and thought to be behind Intel.

Calling TSMC as an extension of AMD and not Apple would say to me that you don't actually mean what you are saying, or that you are completely ignorant of reality.

I actually agree with everything you wrote.

 

[mikk]

Was it? TSMC always said H2'22, and it entered volume production in December 2022. And since it turned out to be an Apple only node, they didn't start volume production until Apple had fully qualified designs ready to go. Volume production can only happen if both the process and the customer are good to go, otherwise you're simply "manufacturing ready", like Intel 3.

3nm entered revenue for the first time in Q3 2023 3 years after 5nm. TSMC struggled with N3B and developed N3E which is design incompatible. It makes sense that most are waiting for N3E. You think it all went according to plan? I don't think so.

10nm Q2-Q3 2017
7nm Q3 2018
5nm Q3 2020
3nm Q3 2023

 

[moinmoin]

The weirdest thing I’ve noticed on the tech enthusiast internet is the pro TSMC monopoly stance.

Are you sure it's a pro TSMC and not actually an anti Intel and anti Samsung stance out of disappointment?

A couple of years ago it looked like an exciting three way battle, Intel being the leader that was slipping up but intend to fight back, Samsung being the foundry which 14nm node at GloFo made AMD start its revival, and TSMC being that workhorse foundry for Apple, Huawei and other mobile phone giants.

A couple of years later Intel and Samsung so far delivered constant disappointments whereas TSMC just executed, with AMD's switch to it bringing it also into the view of people who ignored the mobile phone market before.

This can change now again with TSMC's botched N3, if Intel and Samsung can deliver now to fill the void. But that will again take time to change the general impression of the competitive scene.

 

[H433x0n]

A couple of years later Intel and Samsung so far delivered constant disappointments whereas TSMC just executed, with AMD's switch to it bringing it also into the view of people who ignored the mobile phone market before.

Going to contend this point. Intel hasn’t delivered constant disappointments - until recently it just didn’t deliver. They’ve introduced 10nm/Intel 7 in ~2020 and nothing else. It’s been years of the same process tech.

Samsung delivered multiple generations of disappointments and hype and that ultimately turned out to be paper launches of process nodes with defect densities that weren’t suitable for production. My hope is that they stop the press releases where they declare they’ve got the equivalent to TSMC’s N-1 node in SK Press and then launch vaporware that nobody buys. They need to slow down and fully bake a competitive process.

This can change now again with TSMC's botched N3, if Intel and Samsung can deliver now to fill the void. But that will again take time to change the general impression of the competitive scene.

So far N3B seems eerily similar to Samsung’s 4LPP where on paper it’s incredibly impressive / dense but real world performance of N3B is similar to N4P due to suboptimal defect density. I don’t expect this to be a continuing issue though since TSMC has a superior track record and isn’t plagued with corruption like Samsung.

I personally don’t think people have soured on Intel and Samsung solely due to disappointments because at this point TSMC launched N3 a year late and it’s underperforming. We haven’t really moved forward since 2020, we’re almost in 2024. There really isn’t anything on the horizon that will move the needle much either - N3E is supposed to improve performance over N3B by ~5%. That’s all we’re getting from TSMC for the next few years based on their roadmaps.

 

[Thibsie]

Going to contend this point. Intel hasn’t delivered constant disappointments - until recently it just didn’t deliver. They’ve introduced 10nm/Intel 7 in ~2020 and nothing else. It’s been years of the same process tech.

WHAT?
*Fell off my chair*

 

[Doug S]

This can change now again with TSMC's botched N3, if Intel and Samsung can deliver now to fill the void. But that will again take time to change the general impression of the competitive scene.

There is limited ability for TSMC's big customers to move. Apple's needs are far too large for Intel to handle, and they prefer not to spend money with Samsung unless they have no other alternative. AMD will never consider Intel unless Intel's foundry business is spun off.

Qualcomm's needs are likewise too big for Intel, but they have a lot more market segments than Apple does for their Android SoCs so they could move older ones that are less performance sensitive to a foundry that was cheaper or their higher end more performance sensitive ones to a foundry that can't handle their full volume but could offer them better performance.

As much as I'm skeptical of Intel's claims of being back on track until they truly prove it I'm far more skeptical of Samsung. If Intel operated like Samsung they'd have issued press releases talking about how great 20A was and how it was already in mass production, then maybe a year later you'd see some chips using it that would be comparable to TSMC N5. Intel might mislead and overpromise, but they don't outright lie like Samsung does.

 

[repoman27]

3nm entered revenue for the first time in Q3 2023 3 years after 5nm. TSMC struggled with N3B and developed N3E which is design incompatible. It makes sense that most are waiting for N3E. You think it all went according to plan? I don't think so.

10nm Q2-Q3 2017
7nm Q3 2018
5nm Q3 2020
3nm Q3 2023

Respectfully, I've already argued this into the ground in other threads. Every public statement TSMC ever made about N3 timelines was that volume production was targeted for H2'22. They delivered within that timeframe. Therefore it was not late as far as TSMC, their customers, or their investors were concerned. Yes, they shifted from a 2 year cadence to 2.5 years with N3 and didn't recognize revenue until Q3'23 because they only had a single customer and cycle times are very long.

Do I think it all went according to plan? That really depends on whose plan you're talking about and at what point it was their plan of record.

Did TSMC's N3 plans work out? Not entirely, for sure. N3B was too aggressive and they ended up with only one customer for it. But they pulled off a heck of a save, delivered 4 SoCs for their most important customer, and are on track to recognize 3.5B USD this year from the node. Delivering a 400 mm² SoC on a totally new process family in under 11 months represents one of the fastest ramps in history. The knowledge they gained this year from bringing up the most EUV intensive node in history is going to be incredibly valuable going forward. Meanwhile, how many 3nm wafer starts do Intel and Samsung have under their belts?

Apple made TSMC stick with N3B this year despite the fact that it would soon be scrapped for the N3E-based family. They got the denser node they wanted, on the timeline they dictated, and ended up with it all to themselves. Apple was able to launch key products containing 4 new SoCs right on schedule. Strategically, this was a solid kick in Intel's groin, while also being a reminder to Apple's former colleagues at Qualcomm of the power of their most favored nation status with TSMC.

Intel's plans to utilize TSMC N3 this year certainly didn't work out, but that's all on them.

And it obviously didn't go according to plan for the Chinese OEMs that were barred access to the node by US trade restrictions.

 

[jpiniero]

Speaking of TSMC, it seems that N6 is really slumping. They have to be cutting prices. That's good news for AMD and probably why they feel like they can finally move on from Cezanne.

 

[Joe NYC]

Speaking of TSMC, it seems that N6 is really slumping. They have to be cutting prices. That's good news for AMD and probably why they feel like they can finally move on from Cezanne.

AMD has a lot of products on N6, that should benefit... IO dies, Rembrandt, Navi33, not to mention Zen 3 still going strong.