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.

TSMC Radically Boosts CapEx to Expand Production Capacities, To Reach $14B For 2019

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.

Samsung’s Aggressive EUV Plans: 6nm Production in H2, 5nm & 4nm On Track

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.

 

[swilli89]

Intel has released quite a few details about their PowerVia implementation on Intel 4 just before the upcoming VLSI presentation.

Intel Details PowerVia Chipmaking Tech: Backside Power Performing Well, On Schedule For 2024

www.anandtech.com

Yeah so the rumors of this year's core refresh being basically cancelled absolutely reminds me of the first 10nm chips that intel demo'd as ultra lower power mobile chips only. Well look what we are seeing now several years later: Alder Lake-S cancelled

Now why would Intel be cancelling a much needed new uArch on its Intel 4 (formally 7nm) process if it wasn't having problems? Especially when it was on the roadmap for some time? This symposium paints a really nice picture about its Intel 4 process, saying that even with advanced new tech like backside power delivery they were able to tape out chips. So why would the base Intel 4 process not be able to be used for high volume production of high power desktop chips?

I was REALLY hoping, both as a potential INTC investor, and as an industry enthusiast that we would be seeing Intel make a large step towards a comeback this year with their new process, but it seems like we are still seeing them slide sideways/downward in slow motion.

Tldr: Intel's first desktop 10nm chips were cancelled, and now rumors are pointing to its first 7nm desktop chips being cancelled, are we in for further multi-year delays on yet ANOTHER process node?

 

[dullard]

Yeah so the rumors of this year's core refresh being basically cancelled absolutely reminds me of the first 10nm chips that intel demo'd as ultra lower power mobile chips only. Well look what we are seeing now several years later: Alder Lake-S cancelled

Now why would Intel be cancelling a much needed new uArch on its Intel 4 (formally 7nm) process if it wasn't having problems? Especially when it was on the roadmap for some time? This symposium paints a really nice picture about its Intel 4 process, saying that even with advanced new tech like backside power delivery they were able to tape out chips. So why would the base Intel 4 process not be able to be used for high volume production of high power desktop chips?

I was REALLY hoping, both as a potential INTC investor, and as an industry enthusiast that we would be seeing Intel make a large step towards a comeback this year with their new process, but it seems like we are still seeing them slide sideways/downward in slow motion.

Tldr: Intel's first desktop 10nm chips were cancelled, and now rumors are pointing to its first 7nm desktop chips being cancelled, are we in for further multi-year delays on yet ANOTHER process node?

A) Your link is miss-labelled. It shouldn't mention Alder Lake-S.

B) When a chip is cancelled, it usually means one of two things. (1) The chip is not going to be profitable due to yield, delays, market conditions, etc. Or (2) The next chip, such as Arrow Lake, is doing better than expected due to being ahead of schedule, better than expected yield, etc. We have no actual public evidence if the truth is (1) or (2).

Intel burned themselves in the past by putting too many new things into one chip and then not having the production time to test many iterations and work out the bugs. If you read the PowerVia article that you quoted, it spells out their solution: to NOT put all things into one basket. That is why it won't be used on Intel 4. It was quite likely that the PowerVia wouldn't work, would need cooling that wasn't ready, or some other problem. So, Intel didn't put it into the Intel 4 basket. Plus the timing is just way off. Intel 4 is already in production*. PowerVia won't be ready for another ~year.

* https://www.anandtech.com/show/18836/intel-meteor-lake-intel-4-process-now-ramping-for-production

 

[Geddagod]

Yeah so the rumors of this year's core refresh being basically cancelled absolutely reminds me of the first 10nm chips that intel demo'd as ultra lower power mobile chips only. Well look what we are seeing now several years later: Alder Lake-S cancelled

Now why would Intel be cancelling a much needed new uArch on its Intel 4 (formally 7nm) process if it wasn't having problems? Especially when it was on the roadmap for some time? This symposium paints a really nice picture about its Intel 4 process, saying that even with advanced new tech like backside power delivery they were able to tape out chips. So why would the base Intel 4 process not be able to be used for high volume production of high power desktop chips?

I was REALLY hoping, both as a potential INTC investor, and as an industry enthusiast that we would be seeing Intel make a large step towards a comeback this year with their new process, but it seems like we are still seeing them slide sideways/downward in slow motion.

Tldr: Intel's first desktop 10nm chips were cancelled, and now rumors are pointing to its first 7nm desktop chips being cancelled, are we in for further multi-year delays on yet ANOTHER process node?

MTL-S doesn't feature a "new" architecture. IPC gain is not the standard 15-20% like GLC, SNC. CNL was an abomination, and it only released like 1 or 2 skus. MTL mobile is going all the way up to 45 watt TDP chips and the top end mobile skus (6+8) so it does look better than CNL, if that's any consolation for you.
Even if MTL-S released, the only major advantage it would have had over RPL-R would have been way better efficiency. I don't think it's that big of a deal.
If I were an Intel investor though, I would panic if SRF doesn't come out 1H 2024. Not only is SRF a data center product, where the high margin products are, but it's also the 'true' test of Intel's node- Intel 3 is a slightly buffed, complete Intel 4, and the tiles themselves are going to be pretty large, meaning you would have to have good yields. Server products usually also demand high perf/watt with low leakage, which is also important for mobile - where you also have high margins and volume. Desktop chips should honestly be Intel's lowest priority IMO.
Or you know, you could also panic if MTL gets cancelled all together, but at this point Intel cancelling MTL would require a extremely well hidden, last second crucial bug or something, considering how late in the development cycle MTL is. Unlikely IMO.

 

[SpudLobby]

There is also a 4LPP+ (SF4P), which is better than 4LPP (SF4). Some twitter leaders believe the E2400 will use S4P and nor SF4, while the Tensor G3 will indeed use SF4View attachment 81360

Yeah Tensor G3 is our real test, assuming yields have improved it will be our first new look at Samsung's 4NM process and with *very* comparable IP configurations to the 8 Gen 2 and Dimensity 9200.

 

[DrMrLordX]

Tldr: Intel's first desktop 10nm chips were cancelled, and now rumors are pointing to its first 7nm desktop chips being cancelled, are we in for further multi-year delays on yet ANOTHER process node?

Intel 4 had most of its major products moved to other processes. Meteor Lake is the "last man standing", and even then it's only the 6p+8e and 2p+8e mobile products. Intel is banking on being able to use N3 where necessary and major gains from 20a and 18a to carry them.

 

[A///]

Intel burned themselves in the past by putting too many new things into one chip and then not having the production time to test many iterations and work out the bugs. If you read the PowerVia article that you quoted, it spells out their solution: to NOT put all things into one basket. That is why it won't be used on Intel 4. It was quite likely that the PowerVia wouldn't work, would need cooling that wasn't ready, or some other problem. So, Intel didn't put it into the Intel 4 basket. Plus the timing is just way off. Intel 4 is already in production*. PowerVia won't be ready for another ~year.

Also biting more than they could chew when producing the layers. They even admitted as much back in 2017 or 2018 after admitting the comp, amd, had a leg up against them. there are many who are quick to poke fun at AMD, criticize, or blame them for profiteering when they approach design with a simple philosophy. With AM5 they introduced new features, some still being worked on due to bugs. I will expect AM5 to get an updated chipset to support pcie6 in future. You needn't use a drive or gpu for that but having excess lanes to play with and not hinder component performance is wonderful. wifi7 and 5gbe ports replacing the disastrous 2.5 gbe ones by intel would be another start. we see asrock doing this with their refreshed and might I say less gaudy z790s.

Agree on points a and b, too. Rumors are as useful as putting out a fire with petrol.

 

[A///]

Intel 4 had most of its major products moved to other processes. Meteor Lake is the "last man standing", and even then it's only the 6p+8e and 2p+8e mobile products. Intel is banking on being able to use N3 where necessary and major gains from 20a and 18a to carry them.

N3 here being tsmc? People are quick to point out Intel being a loser and relying on TSMC when they've used TSMC in the past. But the big brains online won't be able to comprehend that Intel may be using N3 here, after paying quite some sum for the fab space, because n3 may be so power efficient it's a no brainer for intel to use them on power critical portions of their design and not waste their own fab space if the competitor is offering a better deal than doing it internally, and may be more power efficient. that to me is smart.

of course go back a year or two and people assumed this was the beginning of of the end for intel selling off everything. give me a break.

 

[DrMrLordX]

N3 here being tsmc? People are quick to point out Intel being a loser and relying on TSMC when they've used TSMC in the past. But the big brains online won't be able to comprehend that Intel may be using N3 here, after paying quite some sum for the fab space, because n3 may be so power efficient it's a no brainer for intel to use them on power critical portions of their design and not waste their own fab space if the competitor is offering a better deal than doing it internally, and may be more power efficient. that to me is smart.

of course go back a year or two and people assumed this was the beginning of of the end for intel selling off everything. give me a break.

Yes, TSMC N3. I still strongly suspect that Arrow Lake-S will need to be on N3 if Intel expects to launch it anytime in 2024. Feel free to surprise us, Intel, but at their current pace it looks like the best internal node Intel will have available in 2024 will be Intel 3 (which if I recall correctly is only slated for Granite Rapids and Sierra Forest at this point).

 

[FlameTail]

TSMC N4 is a rebranded N5?

https://twitter.com/x/status/1562111168684048386

 

[FlameTail]

I wonder what this means. Is N5P > N4 ?

 

[FlameTail]

And also I am curious how Samsung's 5nm/4nm nodes compare with TSMC's in terms of performance/efficiency.

Here is my guess:

5LPE < 4LPE < N5 < 4LPP < N5P < 4LPP+ < N4P

 

[Geddagod]

TSMC N4 is a rebranded N5?

https://twitter.com/x/status/1562111168684048386

Ye I don't think N4 isn't as good as it was supposed to be, I think TSMC just let Qualcomm get away with lying about what node they use. Even the bottom section of the tweet you linked suggests that, since N4P is just supposed to be optimized N4 no?

 

[A///]

Yes, TSMC N3. I still strongly suspect that Arrow Lake-S will need to be on N3 if Intel expects to launch it anytime in 2024. Feel free to surprise us, Intel, but at their current pace it looks like the best internal node Intel will have available in 2024 will be Intel 3 (which if I recall correctly is only slated for Granite Rapids and Sierra Forest at this point).

I want to say this was always Intel's goal because of what I said earlier. There were reports of Intel dropping those plans but I don't remember Intel comfirming either action. In any case it makes a lot of sense for Intel to direct the heavy lifting through TSMC either for the reasons I listed, either, or or both. I suspect it's both, it'll cost Intel less and takes strain off them from producing the parts. This is a massive guess on my part here but if you were to realise how TSMC operates they may have given a quote intel to intel that was below what intel was expecing and played it off as a high cost. in reality its likely much less than intel's own overhead for the time being as they ramp up and don't want to deal with costs plus yield issues. it was either on this forum or in a recent rumor tweet someone broke down what may be on TSMC 3nm. in any case it doesn't matter does it mate? all matters is intel deliver the end product in retail or trays to partners. saying parts were tsmc is like saying gunslinger blesses each palet of product being the jesus freak he is.

 

[A///]

TSMC N4 is a rebranded N5?

https://twitter.com/x/status/1562111168684048386

Sounds like a lot of flowing bs. it's now considered a lie to launch a product early. this reminds me of the accusations some dummies flung tsmc's way about a year back about the n4 vs 4n node where the ladder is a modified 5nm based node specifically for nvidia. i suspect there's been some miscommunication and this may be a special node variation for qualcomm, which is no small customer. Or there's something more sinister going on by way of lies as suggested. to look into this myself or dig into a nice magnum sized bottle of red.... so many choices and evil satans on me shoulders pointing towards the wine. Stay tuned and see if my posting gets sloppy.

 

[Geddagod]

And also I am curious how Samsung's 5nm/4nm nodes compare with TSMC's in terms of performance/efficiency.

Here is my guess:

5LPE < 4LPE < N5 < 4LPP < N5P < 4LPP+ < N4P

Samsung changes and/or has different performance claims for the same node, which makes estimating perf/watt very, very annoying. I remember, not too long ago, I tried estimating perf and pwr numbers (it was my idea of fun, no judgement lol) based on Samsung's and TSMC's own self reported numbers, and the fact that Apple A9 used both TSMC 16nm and Samsung 12nm. I quickly realized, however, the futility of my calculations when I got to some very... weird numbers to say the least... about what Samsung's perf and power would be on their more recent nodes in comparison to TSMC's.
It is also possible (much more likely knowing how bad I am at math haha) that I made mistakes, and through no fault of Samsung's and TSMC's was it that I deemed my attempt a waste of time. Here is the Samsung data if you want to give it a go.

A couple minor disclaimers, as I mentioned above, some of Samsung's own claims dispute the figures reported by wikichip here. This chart was released in Mid 2022, some figures could also have changed. Lastly, since the point of similarity we are using to estimate perf is Samsung 14nm (LPE? IIRC), estimating the perf of nodes further and further into the future are more likely to have larger and larger margins of error, especially considering Samsung hasn't been the most clear on what exact perf gains they are getting on each node, sometimes contradicting older claims.

 

[Geddagod]

Sounds like a lot of flowing bs. it's now considered a lie to launch a product early. this reminds me of the accusations some dummies flung tsmc's way about a year back about the n4 vs 4n node where the ladder is a modified 5nm based node specifically for nvidia. i suspect there's been some miscommunication and this may be a special node variation for qualcomm, which is no small customer. Or there's something more sinister going on by way of lies as suggested. to look into this myself or dig into a nice magnum sized bottle of red.... so many choices and evil satans on me shoulders pointing towards the wine. Stay tuned and see if my posting gets sloppy.

The testing was done by Tech Insights, who I am pretty sure is a legit company with the machinery and expertise to test these chips. While most of their stuff is paywalled, due to the fact that their company sells this info as part of consulting services IIRC, there are still some public data the company as voluntarily released. An example of this are these amazing detailed pictures of SRAM and transistors of TSMC 28nm.
Tech insights has since confirmed Qualcom's version of N4 (not N4E) was a true optical shrink , even if Mediatek's wasn't.

 

[A///]

Samsung changes and/or has different performance claims for the same node, which makes estimating perf/watt very, very annoying. I remember, not too long ago, I tried estimating perf and pwr numbers (it was my idea of fun, no judgement lol) based on Samsung's and TSMC's own self reported numbers, and the fact that Apple A9 used both TSMC 16nm and Samsung 12nm. I quickly realized, however, the futility of my calculations when I got to some very... weird numbers to say the least... about what Samsung's perf and power would be on their more recent nodes in comparison to TSMC's.
It is also possible (much more likely knowing how bad I am at math haha) that I made mistakes, and through no fault of Samsung's and TSMC's was it that I deemed my attempt a waste of time. Here is the Samsung data if you want to give it a go.

A couple minor disclaimers, as I mentioned above, some of Samsung's own claims dispute the figures reported by wikichip here. This chart was released in Mid 2022, some figures could also have changed. Lastly, since the point of similarity we are using to estimate perf is Samsung 14nm (LPE? IIRC), estimating the perf of nodes further and further into the future are more likely to have larger and larger margins of error, especially considering Samsung hasn't been the most clear on what exact perf gains they are getting on each node, sometimes contradicting older claims.

On the one hand I'm impressed you put time into this and on the other i'm confused why you wasted your precious time on earth trying to make sense of samsung's messes. samsung is a non concern for me mentally because they haven't been a reliable foundery in several years now.

 

[A///]

The testing was done by Tech Insights, who I am pretty sure is a legit company with the machinery and expertise to test these chips. While most of their stuff is paywalled, due to the fact that their company sells this info as part of consulting services IIRC, there are still some public data the company as voluntarily released. An example of this are these amazing detailed pictures of SRAM and transistors of TSMC 28nm.
Tech insights has since confirmed Qualcom's version of N4 (not N4E) was a true optical shrink , even if Mediatek's wasn't.

yes they are, quality research firm them. never mind me then. Not sure why mediatek/tsmc saw the need to lie here. But I don't keep up with their news. I don't keep up with Taiwanese investing but you'll often read weird translated reports about TSMC or their semicond companies that will often artificially boost investment figures. they've learned to play our own game. this is why I have some faith in intel. people will ride on tsmc's high wave til their dying breath as if they're unstoppable. cheers gedda, will dedicate this first glass to you for being a co believer.

 

[Geddagod]

On the one hand I'm impressed you put time into this and on the other i'm confused why you wasted your precious time on earth trying to make sense of samsung's messes. samsung is a non issue for me mentally because they haven't been a reliable foundery in several years now.

Eh it wasn't a total waste, I had fun while I was making it, and had nothing better to do hahah. Had a couple other, 'waste of time' moments, remember my EMR mock up, which turned out to be totally wrong because EMR turned up being 2 tiles rather than 4? I remember Exist heavily implying I was wrong about that, but I couldn't figure it out. It's unfortunate though, if I were paying more attention to other servers on discord, I could have found out about EMR being 2 tiles earlier, it was leaked on the chips and cheese server there a short bit before EMR reveal.
Anyway getting back on topic, if anyone were interested in the specific examples of Samsung being contradictory/vague, here they are:
Anandtech

Officially, 3GAE can offer 45% reduced power consumption or 23% improved performance compared to Samsung’s 5nm process (the company doesn’t state which flavor)

These figures are notably different from Samsung’s previous (2019) figures, which compared the tech to Samsung’s 7LPP node.

SemiWiki

LPP (performance) – offers a 14% performance boost over LPE.

Contradicting SemiWiki's own chart which showed LPE offering 10% perf

Anandtech

(Graph) showing 20% performance between 7 LPP vs 10LPE

SemiWiki's chart shows what should be a 32% perf difference between the two

There's bound to be more cases too from what I remember. While some of this might be attributed to further refining from predictions of future nodes versus final results, in many cases the difference between the different data are way too large, and in other cases both data points were found after the node was well into production.

 

[A///]

Eh it wasn't a total waste, I had fun while I was making it, and had nothing better to do hahah. Had a couple other, 'waste of time' moments, remember my EMR mock up, which turned out to be totally wrong because EMR turned up being 2 tiles rather than 4? I remember Exist heavily implying I was wrong about that, but I couldn't figure it out. It's unfortunate though, if I were paying more attention to other servers on discord, I could have found out about EMR being 2 tiles earlier, it was leaked on the chips and cheese server there a short bit before EMR reveal.
Anyway getting back on topic, if anyone were interested in the specific examples of Samsung being contradictory/vague, here they are:
Anandtech



SemiWiki

Contradicting SemiWiki's own chart which showed LPE offering 10% perf

Anandtech

SemiWiki's chart shows what should be a 32% perf difference between the two

There's bound to be more cases too from what I remember. While some of this might be attributed to further refining from predictions of future nodes versus final results, in many cases the difference between the different data are way too large, and in other cases both data points were found after the node was well into production.

say that first part in easy to understand english. who are these servers, are they some kind of messenger from a higher spirit or an organization involved in corporate espionage? someone it may have been markw told me about this discord but it was too complicated to figure out.

I'll take a look at the rest of your links later when I haven't consumed 3 glasses of wine.

 

[Geddagod]

say that first part in easy to understand english. who are these servers, are they some kind of messenger from a higher spirit or an organization involved in corporate espionage? someone it may have been markw told me about this discord but it was too complicated to figure out.

I'll take a look at the rest of your links later when I haven't consumed 3 glasses of wine.

Discord is a 'social media app' essentially. I use it to keep in touch with friends, and use voice chat with my friends when playing video games. A lot of companies/websites have their own discord servers, which are essentially giant group message forums. Chips and Cheese, the website, created their own server, which you can join. Intel has their own server, called Intel Insiders IIRC, where you can talk about their products. The general PC building subreddit has their own discord server as well. Often there are the website founders/officials in those servers as well. The authors of many Chips and Cheese articles are in the Chips and Cheese server, several Intel engineers are in the Intel Insiders server, etc etc.
A lot of useful information are dropped in those servers. For example, rumors are talked about in the Chips and Cheese server, as well as other technical information such as discussions about CPU architecture, nodes, GPUs, and other information that flies over my head. A very good place to learn and discuss opinions. An example of info from the Intel Insiders server was that one Intel engineer who confirmed that MLID's ARC 'A780' leak was infact, complete BS, like Ryan Shrout mentioned on twitter as well. Another great place to talk about technology.

 

[A///]

Discord is a 'social media app' essentially. I use it to keep in touch with friends, and use voice chat with my friends when playing video games. A lot of companies/websites have their own discord servers, which are essentially giant group message forums. Chips and Cheese, the website, created their own server, which you can join. Intel has their own server, called Intel Insiders IIRC, where you can talk about their products. The general PC building subreddit has their own discord server as well. Often there are the website founders/officials in those servers as well. The authors of many Chips and Cheese articles are in the Chips and Cheese server, several Intel engineers are in the Intel Insiders server, etc etc.
A lot of useful information are dropped in those servers. For example, rumors are talked about in the Chips and Cheese server, as well as other technical information such as discussions about CPU architecture, nodes, GPUs, and other information that flies over my head. A very good place to learn and discuss opinions. An example of info from the Intel Insiders server was that one Intel engineer who confirmed that MLID's ARC 'A780' leak was infact, complete BS, like Ryan Shrout mentioned on twitter as well. Another great place to talk about technology.

we call that an msn messenger room.

 

[FlameTail]

The testing was done by Tech Insights, who I am pretty sure is a legit company with the machinery and expertise to test these chips. While most of their stuff is paywalled, due to the fact that their company sells this info as part of consulting services IIRC, there are still some public data the company as voluntarily released. An example of this are these amazing detailed pictures of SRAM and transistors of TSMC 28nm.
Tech insights has since confirmed Qualcom's version of N4 (not N4E) was a true optical shrink , even if Mediatek's wasn't.

So we have N4, N4P and an N4E?

 

[Geddagod]

So we have N4, N4P and an N4E?

Tech Insights called the N4 Mediatek used N4E, since it was the 'early' version of N4 Mediatek used, with no optical shrinkage. Whether N4E is officially what TSMC calls it, is unknown because I have not bothered to dig, but I suspect it is not the case, and Tech Insights just called it N4E to indicate it being early N4. I could be wrong.
N4 and N4P are real though, from what I see.

 

[DrMrLordX]

There is also an N4X node:

TSMC Unveils N4X Node: Extreme High-Performance at High Voltages

www.anandtech.com