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

[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.

 

[Lodix]

I still think TSMC 5nm/4nm is substantially better than Samsung 5nm/4nm in performance and efficiency (even if Samsung fixed their yields) because TSMC uses superior processes and materials (such as cobalt- which iirc Samsung hasn't used)

Yeah, or at least late compared to TSMC. The port of the SD8G1 from Samsung to TSMC's N4 brought up to 30% lower power in CPU and GPU.
We don't know how much the yields impacted the performance but at least the newer versions of Samsung 4nm should be more competitive. From 5LPE to 4LPP there is around 20% lower power from process alone without taking in mind the yield problems at the beginning. From the latest update of info from Samsung Foundry 4LPP got to mass production in 2022 and a new version has o will start mass production this year called 4LPP+ which brings more improvements in efficiency.

 

[DrMrLordX]

At least Intel still out-market caps Globalfoundries.

 

[NostaSeronx]

At least Intel still out-market caps Globalfoundries.

Not that hard when Intel produces product in their fabs and TSMCs.

GlobalFoundries is not really announcing anything. So, no hype to pump up the stock/market cap...

Brief search, what they aren't telling investors to what employees-big heads are saying is happening:
1. ASML ArFi/ArFd/KrF upgrades => Fab1/Fab7/Fab8 (2021-2024)
2. Digital Team dropping FinFETs R&D to focus on 28BLK/22FDX/12FDX (Fully Same Tool FabSync across GF1,7,8; extends to Fab11/Fab12). (2021+)
3. 12FDX to achieve 1times volume at Fab8 by end of 2H'23.
4. Fab12(France, Crolles) team to be trained at Fab8(22FDX/12FDX) and Fab1(22FDX/12FDX).
5. Fab11 is being revived and moved to India;(???), customer/government paid.

{extra edit iteration: 12FDX is split into two just like 22FDX
22FDX LP = 116Cx(28nm-like FEOL::14nm FDSOI spec'd gate)
22FDX HP = 104Cx(22nm-like FEOL::14nm FDSOI spec'd gate)
12FDX LP = 84Cx(14nm/12nm-like FEOL::10nm+ or 7nm FDSOI spec'd gate) <-- more dense than 18FDS
12FDX HP = 64Cx(10nm-like FEOL::10nm+ or 7nm FDSOI spec'd gate) <-- much more dense than 18FDS
Which isn't much given SRAM scaling:
18FDS SRAM; HD = 0.102um^2, HP = 0.124um^2
22FDX SRAM; HD = 0.110um^2, HP = 0.124um^2
12FDX SRAM: HDe = 0.058um^2}

 

[oak8292]

if the ifs nodes are slightly less efficient but cost much less than TSMC's 2 and 3 nodes then intel will sweep up. Last I looked the 3nm wafers by tsmc were 20k minimum. I don't know the break even cost for Intel but if they shot for 10% above that investors will overlook the lack of profit for the customer base. if tsmc has their cake taken from them they won't have the capital to expand and must ask the Taiwanese government for handouts. That's not cheap, and that's on the basis they get it right the first time.

interesting times ahead.

The probability that a node from Intel is less than a node from TSMC is very unlikely. It could happen but it would really take TSMC screwing up almost as dramatically as Intel. TSMC has a lead customer in Apple that is very unlikely to ever switch to Intel. Even if Otellini had offered to manufacture processors for Apple they probably would have switched to TSMC in the same way they left Samsung. I believe they left Samsung because TSMC has a larger eco-system of trailing users that allowed them to keep advancing at a more rapid pace. Intel would have wanted them as a trailing customer and not an equal partner in the most advanced node.

When Intel gets their first NA EUV machines, how much customer capacity do you think they will give up? TSMC doesn’t have any competing products and Apple or any other lead customer will get wafers. TSMC will have a lead customer for EUV processes. If TSMC fails to produce a node like Intel then TSMC will lose customers and Intel will get a higher allocation of EUV machines. I think there is a high probability that TSMC has EUV machines initially allocated to Intel but Intel didn’t have a running process to utilize the machines.

TSMC has a large EUV wafer capacity and has been in production for three years. Intel is going to start with 4 nm chiplets this fall on EUV. Intel needs to walk before they can run. They have a long way to go in IFS. IFS can be done but it really requires a failure by TSMC. Intel will not take lead customers from TSMC without a process failure.

 

[Doug S]

tsmc can only output from what they get in. if they don't have the clients they don't have the income and if they don't have the income they can't keep the volume up. this is how things work. intel could undersell their own wafers and slowly leach clients as they build volume up.

And Intel can only output from what they have the capacity to output. TSMC will have customers because Intel won't have the capacity to serve more than a fraction of TSMC's leading edge output. What good does it do if TSMC customers all go to Intel and then only get a quarter of the wafers they need because Intel is backlogged and everyone is on allocation? Sure if Intel can sign up a bunch of those customers then they can spend tens of billions building and equipping the capacity necessary to serve them and THEN they can take TSMC clients. But it since takes about four years from the "ok let's do it" to mass production output, Intel can't do that for any of their currently announced nodes (i.e. through 18A)

They will have to beat TSMC on performance or price, prove they can deliver as a foundry business, and give everyone confidence they will continue to do those things before TSMC's big clients would consider signing up for lots of guaranteed capacity for some future Intel node. So maybe Intel could start taking a big bite of TSMC's base by the end of the decade at the soonest.

 

[Doug S]

TSMC has a lead customer in Apple that is very unlikely to ever switch to Intel. Even if Otellini had offered to manufacture processors for Apple they probably would have switched to TSMC in the same way they left Samsung.

Apple wouldn't have considered Intel in the Otellini days because they didn't have the capacity to handle Apple nor did they have the ability to act as a true foundry. They were going into shortage on 14nm in multiple consecutive years every time Apple's demand for modems increased prior to the launch of a new iPhone, after all. Can you imagine if they tried to produce iPhone SoCs? They were also having well publicized issues with their 10nm process, while TSMC was executing on all cylinders. It was a no brainer to go with TSMC.

I disagree strongly with your claim that Apple "is very unlikely" to ever switch to Intel. I think they will if the following happens: 1) Intel can demonstrate they are able to be a proper foundry this time around. 2) Apple throws some lower volume business their way and things go well. 3) Intel is able to commit to building sufficient guaranteed capacity to handle Apple's needs - i.e. basically like the deal Apple has with TSMC.

The big attraction for using Intel is "made in the USA". Or at least "made by the USA" as their customers outside the US don't really care if the chips are made in the USA or not but Intel has factories in multiple places around the world rather than having all their eggs in one basket like today with TSMC in Taiwan. So Apple would reduce its business risk in addition to increasing the US made content in their products.

It is of course more difficult to move now than when Apple moved from Samsung to TSMC in the previous decade, because Apple is relying on TSMC for a lot of their advanced packaging technology for e.g. the Apple Silicon Pro/Max and especially Ultra/"Extreme". Intel has similar technologies but they aren't the same so it would require a lot of engineering to switch. So the first step would be smaller scale stuff like the Watch SoC or perhaps Apple's modem which is likely to be a standalone chip for the first few years before it gets integrated onto the iPhone SoC. I couldn't see Apple shifting a majority of their production to Intel any sooner than 2030, and that would be a pretty aggressive timeline, but I think they have reason to do so assuming they can work out the above 3 points and of course Intel is able to actually execute on their process roadmap promises and beat or at least match TSMC in performance and more importantly for Apple, power.

 

[A///]

Apple wouldn't have considered Intel in the Otellini days because they didn't have the capacity to handle Apple nor did they have the ability to act as a true foundry. They were going into shortage on 14nm in multiple consecutive years every time Apple's demand for modems increased prior to the launch of a new iPhone, after all.

Apple couldn't have considered it because Intel was not shipping 14nm at that point. Intel announced 14nm in 2011 with a plan to begin shipping in 2014. Ottelini announced his retirement in 2012 and retired in 2013, long before 14nm's woes. Which would come years after their initial release. Intel gave Apple the middle finger for manufacturing their phone SoC because at the time they and a lot of people didn't think they would sell enough that it would be worth the business investment. Instead Apple and in line with that, Steve Jobs's bravado spurred the development of other smart phones on the Android operating system for mobiles. If the iPhone never came into existence we would be living in a very different world today, IMO.

 

[A///]

The probability that a node from Intel is less than a node from TSMC is very unlikely. It could happen but it would really take TSMC screwing up almost as dramatically as Intel. TSMC has a lead customer in Apple that is very unlikely to ever switch to Intel. Even if Otellini had offered to manufacture processors for Apple they probably would have switched to TSMC in the same way they left Samsung. I believe they left Samsung because TSMC has a larger eco-system of trailing users that allowed them to keep advancing at a more rapid pace. Intel would have wanted them as a trailing customer and not an equal partner in the most advanced node.

When Intel gets their first NA EUV machines, how much customer capacity do you think they will give up? TSMC doesn’t have any competing products and Apple or any other lead customer will get wafers. TSMC will have a lead customer for EUV processes. If TSMC fails to produce a node like Intel then TSMC will lose customers and Intel will get a higher allocation of EUV machines. I think there is a high probability that TSMC has EUV machines initially allocated to Intel but Intel didn’t have a running process to utilize the machines.

TSMC has a large EUV wafer capacity and has been in production for three years. Intel is going to start with 4 nm chiplets this fall on EUV. Intel needs to walk before they can run. They have a long way to go in IFS. IFS can be done but it really requires a failure by TSMC. Intel will not take lead customers from TSMC without a process failure.

All it requires of TSMC is to lose their mark, their node performance not to be as good or in par with intel or the yield for tsmc to be terrible. No one here nor online has any idea of tsmc's true yields unless a tsmc engineer familiar with the node were to leak it. This is why I often ignore any media reports of a node performing badly because over the years I've seen so many claims only for them to turn out to be bullshit. No one stays at the top forever. If you want to see what happens when you have little mishaps that add up over time glance over at Samsung's way or look at Intel's history. Either most on here were not old enough to see the damage Intel's 90nm woes had on the company or they were not yet born. That wasn't the first time and obviously not the last time Intel had a multi year mishap with a leading node.

 

[A///]

And Intel can only output from what they have the capacity to output. TSMC will have customers because Intel won't have the capacity to serve more than a fraction of TSMC's leading edge output. What good does it do if TSMC customers all go to Intel and then only get a quarter of the wafers they need because Intel is backlogged and everyone is on allocation? Sure if Intel can sign up a bunch of those customers then they can spend tens of billions building and equipping the capacity necessary to serve them and THEN they can take TSMC clients. But it since takes about four years from the "ok let's do it" to mass production output, Intel can't do that for any of their currently announced nodes (i.e. through 18A)

They will have to beat TSMC on performance or price, prove they can deliver as a foundry business, and give everyone confidence they will continue to do those things before TSMC's big clients would consider signing up for lots of guaranteed capacity for some future Intel node. So maybe Intel could start taking a big bite of TSMC's base by the end of the decade at the soonest.

I read halfway into this before realizing you're repeating yourself like a parrot. Sit and wait. Let's circle back in 4 years provided I'm still alive and we'll discuss how bad or well Intel did and what TSMC's position is.

 

[oak8292]

Apple wouldn't have considered Intel in the Otellini days because they didn't have the capacity to handle Apple nor did they have the ability to act as a true foundry. They were going into shortage on 14nm in multiple consecutive years every time Apple's demand for modems increased prior to the launch of a new iPhone, after all. Can you imagine if they tried to produce iPhone SoCs? They were also having well publicized issues with their 10nm process, while TSMC was executing on all cylinders. It was a no brainer to go with TSMC.

I disagree strongly with your claim that Apple "is very unlikely" to ever switch to Intel. I think they will if the following happens: 1) Intel can demonstrate they are able to be a proper foundry this time around. 2) Apple throws some lower volume business their way and things go well. 3) Intel is able to commit to building sufficient guaranteed capacity to handle Apple's needs - i.e. basically like the deal Apple has with TSMC.

The big attraction for using Intel is "made in the USA". Or at least "made by the USA" as their customers outside the US don't really care if the chips are made in the USA or not but Intel has factories in multiple places around the world rather than having all their eggs in one basket like today with TSMC in Taiwan. So Apple would reduce its business risk in addition to increasing the US made content in their products.

It is of course more difficult to move now than when Apple moved from Samsung to TSMC in the previous decade, because Apple is relying on TSMC for a lot of their advanced packaging technology for e.g. the Apple Silicon Pro/Max and especially Ultra/"Extreme". Intel has similar technologies but they aren't the same so it would require a lot of engineering to switch. So the first step would be smaller scale stuff like the Watch SoC or perhaps Apple's modem which is likely to be a standalone chip for the first few years before it gets integrated onto the iPhone SoC. I couldn't see Apple shifting a majority of their production to Intel any sooner than 2030, and that would be a pretty aggressive timeline, but I think they have reason to do so assuming they can work out the above 3 points and of course Intel is able to actually execute on their process roadmap promises and beat or at least match TSMC in performance and more importantly for Apple, power.

The Apple A4 was built on the 45 nm node at Samsung, at what I believe was Samsung’s Austin, Tx facility. Apple stuck with Samsung in Austin until their wafer demand was equal or starting to exceed the 40K wpm capacity of the Austin facility on the 28 nm node. (still larger than the 22 nm that Intel was using). The transition to TSMC occurred with Apple and Qualcomm swapping fab on the 20 nm node. Qualcomm had been the lead customer at TSMC. The Qualcomm modems that Apple used limited the wafer capacity available at TSMC on the 16 nm FinFET node and they split an order for processors between TSMC and Samsung for one year.

Intel contracted to build Intel modems for Apple on 14 nm when Intel thought it would be selling capacity on a trailing node and most of their Core products would be on 10 nm. The delay in the 10 nm transition meant Intel contracted out wafers that weren’t available. The Intel transition from 14 to 10 nm will go down as one of the biggest business train wrecks in history.

 

[Doug S]

Apple couldn't have considered it because Intel was not shipping 14nm at that point. Intel announced 14nm in 2011 with a plan to begin shipping in 2014. Ottelini announced his retirement in 2012 and retired in 2013, long before 14nm's woes. Which would come years after their initial release. Intel gave Apple the middle finger for manufacturing their phone SoC because at the time they and a lot of people didn't think they would sell enough that it would be worth the business investment. Instead Apple and in line with that, Steve Jobs's bravado spurred the development of other smart phones on the Android operating system for mobiles. If the iPhone never came into existence we would be living in a very different world today, IMO.

OK I don't pay much attention to who is running Intel at what time, I thought he was the CEO at the time Apple switched from Samsung to TSMC. That's the timeframe I was talking about, not in 2012 when Apple was launching their first fully custom core.

I agree Intel screwed up not accepting Apple's business when Jobs approached them in 2005 or so. My personal belief is that Apple was interested in StrongARM, which was the best performing ARM core at the time which Intel owned (they sold it in 2006) and I'm guessing Intel was telling them they'll do it if Apple uses x86 instead of ARM. But that's just a guess knowing what Jobs was looking for and how Intel's corporate culture at the time was "x86 everywhere" (though they didn't come up with that slogan until a few years later)

 

[SpudLobby]

I still think TSMC 5nm/4nm is substantially better than Samsung 5nm/4nm in performance and efficiency (even if Samsung fixed their yields) because TSMC uses superior processes and materials (such as cobalt- which iirc Samsung hasn't used)

I think it really was partially about yields and being a genuinely inferior ”mid” process for SS 5NM which was kind of like N6 in terms of density, but I am starting to believe that TSMC has some extra sauce, it’s definitely possible. Their + or P nodes for ex are impressive I think.


At any rate, with similar or even just “good” parametric yields what I want to see is SF4. The Exynos 2200 might have been on this but yields were horrible, and the 8 Gen 1 was on 5NM LPE/LPP, “4NM LPX” was just branding. And since Samsung’s 4NM is a genuinely different and notably (not just 5-6%) denser process unlike N5 -> N4, and yields may rise, it would be encouraging to see if they can get it to mostly match where N5 or N4 are in terms of power/performance.

If they need 3NM GAAFET’s 20-30% density improvement over 4NM and the GAAFET transistors to match N4/N5 on power/performance though, that’s honestly just pathetic assuming that’s with good yields. Means TSMC will probably wax them with N3E/N3S, and that’s not even using a new transistor architecture at least intrinsically (they do have FinFlex though).

 

[SpudLobby]

OK I don't pay much attention to who is running Intel at what time, I thought he was the CEO at the time Apple switched from Samsung to TSMC. That's the timeframe I was talking about, not in 2012 when Apple was launching their first fully custom core.

I agree Intel screwed up not accepting Apple's business when Jobs approached them in 2005 or so. My personal belief is that Apple was interested in StrongARM, which was the best performing ARM core at the time which Intel owned (they sold it in 2006) and I'm guessing Intel was telling them they'll do it if Apple uses x86 instead of ARM. But that's just a guess knowing what Jobs was looking for and how Intel's corporate culture at the time was "x86 everywhere" (though they didn't come up with that slogan until a few years later)

I agree with this. I think Intel had an opportunity to sell them a custom chip but were wedded to their X86 team IP and possibly might’ve had some insane branding demands. But I think further Intel thought it wouldn’t pan out, e.g. they really didn’t think it would be worth the time which might’ve changed negotiation.

 

[SpudLobby]

Yes SD888 and SD8 Gen 1 use Samsung 5nm. But it is not correct to call it just a derivative, because it has some meaningful changes and improvements to density compared to their 7nm. And In density Samsung 5/4nm is comparable to TSMC's 5/4nm. In terms of efficiency is difficult to say, because as you mention when this SoC were made Samsung had terrible yields that are now corrected.
Also to mention Samsung 3nm should be comparable to TSMC 3nm in density and efficiency.

This is their latest information. But we don't know how a good yielding 4LPP fairs.

Thanks. Really, really want to see 4NM LPP though, I am pretty sure it’s actually meaningfully different in a directional magnitude that N5 -> N4 is not. You see this also with how Samsung compares the power/perf from 3NM: I will have to grab it later but they’ve listed significantly higher gains over 5NM than 4NM (which was after the recent conference) in a way suggesting significantly improved performance and power in 4NM.

I say this also because it’s a good clue to how Samsung’s really doing. You want to see them with a real FinFET-based node using EUV that can roughly achieve parity with TSMC on power, performance. From the 8 Gen 1 to 8 Gen 1+, Qualcomm got up to 35% lower power iso-arch, frequency (35% seen with the A710), and in general about 20-30% lower iso-perf. And that was a quicker port as far as we know, in reality Ming Kuo among others implied a further process would have optimized for TSMC’s design rule a bit better which supposedly gave the 8 Gen 2 some extra gains.


Frankly the Tensor G3 and maybe Exynos 2400 will be the best way to see this. Apparently the former is based on Exynos 2300 and 4NM LPP, and the latter is obviously the 2400 and also on 4NM LPP per rumors.

Tensor G3: soon announcement, this fall full release.
Exynos 2400: probably January.

G3 will have an X3 core, probably 4 A715’s, and 4 A510’s on a real 4nm. Perfect opportunity this fall to compare between it and the D9200, 8 Gen 2 with their X3’s, A7x class cores.

Same for the Exynos 2400 next year and presumably X4’s.

 

[A///]

OK I don't pay much attention to who is running Intel at what time, I thought he was the CEO at the time Apple switched from Samsung to TSMC. That's the timeframe I was talking about, not in 2012 when Apple was launching their first fully custom core.

I agree Intel screwed up not accepting Apple's business when Jobs approached them in 2005 or so. My personal belief is that Apple was interested in StrongARM, which was the best performing ARM core at the time which Intel owned (they sold it in 2006) and I'm guessing Intel was telling them they'll do it if Apple uses x86 instead of ARM. But that's just a guess knowing what Jobs was looking for and how Intel's corporate culture at the time was "x86 everywhere" (though they didn't come up with that slogan until a few years later)

Apple never made an immediate switch. They used both Samsung and TSMC and ultimately phased out Samsung as a partner for their soc's. It's been a hell of a long time since I read it but I want to say Intel in their stupid mode a few years later once realising that smart phones were gonna be the next big thing short of a phallic space rocket tried to sell them on an x86 soc. Or the year following when jobs first met with the then intel great leader of moronic foresight. my best guess here is that intel was alread y in talks to sell strongarm and jobs got to them real late in the process. but even then if you told me was a smart phone was and how it would overtake the world's flips or palms, I'd have called you a crack head then too. Apple had been running x86 pcs both prebuilts and customs in their secret lab, forget the name of it, for years leading to the adoption of x86. there's some old articles kicking around online pointing toward 97 to 99 when they began their experiments. I forget the guy's name but john shineburg??????? he's still alive but retired from apple almost 20 years ago. he ran project maklar or marklar which was kind of like the project startrek from several generations with older macs. apple kept feature parity with an identical osx to run on x86 and went all in when they had the idea that maybe they would need to eventually jump ship, which is what happened. the time between 2000-2001 to 2005 is fuzzy and there's not a lot of public info on that apart from steve getting up on stae in 2005 with the dev transition kit announcement which was a pentium 4 running in a mac pro case. I don't remember what mac first made the transition but we got my then wife one of the first intel based mac pros for her work. that machine was incredible for that time. apple had a big leg up for years over pc in performance and still had that leg up for a few years until both reached parity. now it's kind of the same leg up with the m processors but x86 is and will surpass it eventually because that is what consumers are demanding. how soon we get there is anyone's guess. will apple ever come back to x86? dunno. don't care tbh. I love apple but I'm a consumer wanting the best hardware available for an affordable price.

if apple releases a mac pro on monday i want the pricing to be sane. not intel mac pro pricing and not the insane pricing they had back in the day. apple carries a premium of course but going beyond premium pricing into new german car territory is stupid.

 

[igor_kavinski]

if apple releases a mac pro on monday i want the pricing to be sane.

Don't see why the starting price would be less than $4999. They know it will still sell out at that high price. A medium build would be $9999 and fully specced out model could be $20K. For professionals, paying that much for something they will use for the next 5 or 6 years isn't extravagance. They see it as a necessity.

 

[A///]

Don't see why the starting price would be less than $4999. They know it will still sell out at that high price. A medium build would be $9999 and fully specced out model could be $20K. For professionals, paying that much for something they will use for the next 5 or 6 years isn't extravagance. They see it as a necessity.

It'll be lower entry imo. unless apple can work around the problem of their current setup with everything being close to the soc then it would be a huge financial headache when repairs need to be done and that needs to be replaced under warranty. dunno how they'll do it tbh. it'd be nice to see add in cards of additional memory and storage using some fancy comms method they came up with or how the old mac pros were with tray insets of the cpu. imagine inserting 4 aftermarket trays of additional socs if you need some wall busting performance.

tdp is gonna be high on these for sure.

 

[FlameTail]

Thanks. Really, really want to see 4NM LPP though, I am pretty sure it’s actually meaningfully different in a directional magnitude that N5 -> N4 is not. You see this also with how Samsung compares the power/perf from 3NM: I will have to grab it later but they’ve listed significantly higher gains over 5NM than 4NM (which was after the recent conference) in a way suggesting significantly improved performance and power in 4NM.

I say this also because it’s a good clue to how Samsung’s really doing. You want to see them with a real FinFET-based node using EUV that can roughly achieve parity with TSMC on power, performance. From the 8 Gen 1 to 8 Gen 1+, Qualcomm got up to 35% lower power iso-arch, frequency (35% seen with the A710), and in general about 20-30% lower iso-perf. And that was a quicker port as far as we know, in reality Ming Kuo among others implied a further process would have optimized for TSMC’s design rule a bit better which supposedly gave the 8 Gen 2 some extra gains.


Frankly the Tensor G3 and maybe Exynos 2400 will be the best way to see this. Apparently the former is based on Exynos 2300 and 4NM LPP, and the latter is obviously the 2400 and also on 4NM LPP per rumors.

Tensor G3: soon announcement, this fall full release.
Exynos 2400: probably January.

G3 will have an X3 core, probably 4 A715’s, and 4 A510’s on a real 4nm. Perfect opportunity this fall to compare between it and the D9200, 8 Gen 2 with their X3’s, A7x class cores.

Same for the Exynos 2400 next year and presumably X4’s.

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 SF4

 

[Ajay]

Does Samsung Logic have any customers, aside from themselves, on their leading edge nodes?

 

[Doug S]

Apple never made an immediate switch. They used both Samsung and TSMC and ultimately phased out Samsung as a partner for their soc's.

There was only one year where they used both, so it was a pretty quick transition. Whether that was to "test the waters" before switching or because TSMC couldn't commit to the volumes they wanted without more lead time who knows.

 

[Doug S]

Don't see why the starting price would be less than $4999. They know it will still sell out at that high price. A medium build would be $9999 and fully specced out model could be $20K. For professionals, paying that much for something they will use for the next 5 or 6 years isn't extravagance. They see it as a necessity.

The current one starts at $5999 and I'm skeptical it would start lower than that. It can reach above $50K if you max it out, so the new could easily go much higher than $20K. It all depends on how high they'll go on DRAM since that's where half that $50K comes from in the current model.

They would be limited to 384 GB with M2 using the same LPDDR5 stacks the M2 Pro/Max does but could reach up to 3 TB if they offered higher density stacks. If they wait for M3 to do the Mac Pro (which I expect, so I don't think we'll see one tomorrow but maybe I'm wrong) then the base limit would be 576 GB and the max over 6 TB if they've added 50% more LPDDR controllers as appears may be the case based on the '36 GB' Mac benchmark leak (assuming it isn't a fake and the software isn't being confused about the amount of installed DRAM)

 

[jpiniero]

There was only one year where they used both, so it was a pretty quick transition. Whether that was to "test the waters" before switching or because TSMC couldn't commit to the volumes they wanted without more lead time who knows.

The dual source was because of the GloFo thing.

 

[A///]

There was only one year where they used both, so it was a pretty quick transition. Whether that was to "test the waters" before switching or because TSMC couldn't commit to the volumes they wanted without more lead time who knows.

They used TSMC a few times with earlier processors instead of Samsung for whatever reason.

 

[Dayman1225]

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

 

[A///]

It should be interesting to see it in a finalised product. Next nodes will see near density parity with tsmc and other tricks intel will employ before tsmc. Whether it'll pay off or not is anyone's guess now. intel 7 aka 10nm superfin sits @ 100.76MTr/mm2 with TSMC 5nm at around 140. tsmc also has the node advantage. meeting the node later but still lacking in desnity.