What mistakes? Remember that initial schedules are just estimates. As it gets closer, final dates become clearer. It’s never sooner, always a bit later. Nothing new about that.
What’s amazing is that 3nm can be done at all, much less 2nm, which seems insane.
TSMC's 2nm using GAA is going to be insane especially for SRAM density. When Intel//Samsung/TSMC are all on there GAA nodes densities will allow all sorts of fun like good gaming APU's.
TSMC isn't just "estimating" introducing a node for the Apple order. That's a goal. If they miss it it's a failure. There have been at least one and possibly two delays associated with 3 nm now. First the Apple order was missed, and now probably another, since they are only recognizing revenue in 2023. At least they are giving more clarity on the extent of the delay and it's bigger than many were anticipating. It's looking a bit like a 6 month delay now.
Revenue in 2023 means process start in 4Q2021 since it takes 3+ months get N3 chips finished from start. So the timeline of 2H2021 mass production is met.
I don't know the normal revenue-recognition for TSMC, but I doubt it's 3 months from wafer start. The "timeline is met". In other words, it can be spun.. Mass production was supposed to start in Q2. Then it was 2H 2022. Now it apparently will be Q4. But when in Q4? Since revenue recognition won't take place until 2023 it's likely not early Q4. Who knows, maybe it will be December 20th 2022? That will "meet the timeline". Or, now that they've shifted from talking about 2H 2022 production to Q1 2023 revenue recognition they don't need to actually have volume production start in 2022 at all to "meet the timeline". When they successfully start recognizing the revenue in Q1 2023 they will declare success regardless.
Didn't miss the date if apple just didn't buy enough 3nm to launch with NOT figuring Intel would jump in and buy as much 3nm as possible to screw all involved. Brilliant. :) Apple is dancing because they thought they had 3nm for macs next year, now AMD is dancing because apple took the 4nm they hoped to have. My guess is AMD is least prepared to deal with what has happened in the last 6-9 months. NV/Intel/Apple just launch other parts they have stored up just in case of crap like this. AMD on the other hand does each design as needed pretty much. They don't have the NET INCOME to FART around with test projects or just in case designs. Now NV is coming back big for gpu at TSMC also, so everyone is WAFER-FU fighting...LOL. Good times. Intel/MU probably both crack 100 before next xmas (not this year, DDR5 rollout etc next year causes it, higher margin new junk to pitch).
Tesla 4000 by 2025 end. 10K by 2030. :) Unless the govt figures out how to kill him, or big pharma, food industry, unions, battery makers, car makers, battery pack makers, etc etc etc...Jeez, this guy better have good protection, trying to take out wasteful multi-billion dollar markets can be VERY dangerous. These people at the FDA, FBI, CIA, NSA, DOJ etc etc, all have no problem knocking off anyone that gets in their POWER way. They have all proven themselves to be ENEMIES of the state run probably by vatican/soros etc from DC (that little place that is a vatican business, running the fake USA). hmm...Just saying. 1870act/1871 act....You have to look up the first act to get what they really did in the 2nd. Traitors.
'There is plenty of room at the bottom' but not *that* much room left. The limits of quantum mechanics are either an enemy (for conventional FETs) or a friend (for quantum FETs, or QFETs, which employ quantum tunneling).
QFETs work better, not worse, with cutting edge nodes because their very fine features make quantum tunneling easier and more efficient. Although they employ quantum tunneling (QT) they are not quantum computers; they use QT as an optimization technique at the transistor level, not as a high level operating principle.
Even QFETs, though, have size limits. And before those limits are reached the economics of building fabs with.. subarmstrong process nodes will stop making sense, since they will always have a negative ROI.
It wasn't that many generations ago when Intel appeared to be the untouchable leader, or when Samsung 14nm was beating TSMC 16nm on apple chips. Anything can happen and a single node can make all the difference.
(a) What mistakes? "Many observers, however, expected these chips to ship in late 2022."
TSMC can't change the rate at which ASML produces EUV machines, and that is the primary gating factor right now for everyone. It makes zero sense, under conditions where they can't even meet the demand for 5nm given the number of EUV layers, to promise, and make attempts to rush, a 3nm that uses even more layers.
(b) TSMC missing iPhones and A16 is hardly a catastrophe. We have been through this before -- look at A10 (on 16nm) and A10X (on 10nm). Apple will ship A16 on N5PP or N4 or whatever makes the most sense, and they will also schedule some chip (maybe M2/M1X/whatever the Mac Pro chip will be called) for N3. TSMC have a wide range of clients -- which is how they get value from processes all the way from leading edge and leading edge-1 back to 90nm and older. And APPLE have a range of SoCs -- which is how they can always find something to schedule on the very latest leading edge.
(c) Apart from lithography, there are two interesting new developments coming up. One is GAA, one is BPR. It's basically a business decision as to which you implement first. SS has decided for GAA first, TSMC has decided for BPR first. You can fault this ordering; what you can't do is both complain that SRAM density is not increasing very much node to node AND complain that TSMC is introducing BPR first, given that a primary consequence of BOR should be to free up some space in the metal layers and thus help SRAM shrink a little...
Kinda sad how we really are reaching the limits of tech and struggling to achieve better processing power. But I guess 2nm will be enough for 4k/120fps and I can settle with that.
Agree.. ram on CPU chip will allow for performance jumps in bandwidth restricted processing like GPU. Not to mention PC motherboard shrinking to tiny sizes in laptops. May be we will have Nano-itx boards and cases rise in demand.
That was not a very serious comment but considering that Apple starts selling iPhones in September, having them shipping & available by the end of March would be good I guess. Also, they have a long cycle time due to complexity so, maybe they don't wanna risk it? There's always AMD, Intel also maybe - TSMC won't have trouble selling those for sure!
The US government's fiscal year starts in October. The point of a "fiscal year" is that every company can have its own. TSMC's fiscal 2020 ended December 31, 2020. So TSMC's fiscal year is apparently very well aligned with the calendar year.
So if Intel stays on their Node roadmap (big if, I know), this means they are behind TSM by 1-2 quarters for 3nm and should be ahead of TSM by 2024 (20ang) ?
It's not a big if. All indications are that Intel has confidence in its roadmap. I doubt Pat Gelsinger signed on to be a fall guy.
Of course unforeseen complications can happen. But that's different from what was going on with Intel before. The unforeseen complications happened once or twice but Intel tried to obscure it while they were hoping they'd find a fix. That could very well be something going on with TSMC now, in fact. Notice that it's now slipped from 4 months delay to 6 months delay. Probably that's where it will stay, but TSMC management would likely be just as dodgy as Intel's was if there were some major issues to work out.
The narrative that Intel has suddenly forgot how to do semiconductor engineering and TSMC has suddenly become an infallible leader is nonsense. It was only 2014 when TSMC last ran into major problems. But Samsung and GlobalFoundries did as well. Only Intel navigated the planar to finfet transition well and since Intel did not have a big foundry business TSMC was not burned by their poor execution.
TSMC's FinFET nodes are usually one step ahead of it's competitors in performance, so they can afford to be conservative and take it slow with GAA. They could lose some space in ultra-low-power applications where GAA will shine, but sticking to FinFET means both cheaper wafers and more wafers. That in turn means they'll make healthy profits, while Samsung and Intel will be stuck with yield issues and an expensive and slow manufacturing process. By then TSMC will have to transverse a path already explored by the industry.
Even if Samsung and Intel catch up in terms of node density and performance, TSMC will have more money to invest in R&D, which they could use to offer something neither of the other foundries have. And that's important: as Moore's Law decays in it's coma, the question of technology leadership will get more and more complex. It might soon be decided not necessarily by the silicon node, but by the maturity and cost/effectiveness of packaging technologies, or through the integration of new materials, equipment or techniques which make production easier (E.g. TSMC's EUV masks)
TSMC never actually said N2 would be released in 2025. Wei only said (implicitly) that N2 would be TSMC's most advanced node in 2025. It does make sense that N2 would be released in 2025 because it doesn't make sense to miss the Apple order in 2024, assuming they can continue to win it. But neither the schedule of N3E nor the fact that Wei implied N2 would be TSMC's more advanced node in 2025 should be used to infer conclusively that N2 will only appear in 2025.
Admittedly, TSMC won't need 2nm before 2025 since AMD will continue with Zen3 while Apple will sharpen 3nm next year,,, AMD hasn't released Zen4 yet so with that, enhanced 7nm and 5nm (it was said that AMD had their own process), that will last for 2 years or more...
With RDNA2 APUs AMD can eliminate certain low end GPUs... And 5nm should make an MI400+ since compute doesn't require additional processing for frames... They should be able to get 30x25 with 3nm in 2025 with chiplets and HBM (created by AMD and Hynix)..
And Apple probably won't need a full node by 2023 so the roadmaqp seems to be more a design time issue with smaller geometries and EUV... AMD and Apple have multiple choices...Apple took the perf/w crown but 5nm should be s good for AMD with ~85% and last until maybe 2nm...
We don't know if Nvidia will fully move back to TSMC but the 5/4/3nm processes will allow for more dies/wafer and/or more logic...
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
32 Comments
Back to Article
liahos1 - Monday, October 18, 2021 - link
So it seems even the mighty TSM is not without their mistakes...melgross - Monday, October 18, 2021 - link
What mistakes? Remember that initial schedules are just estimates. As it gets closer, final dates become clearer. It’s never sooner, always a bit later. Nothing new about that.What’s amazing is that 3nm can be done at all, much less 2nm, which seems insane.
FreckledTrout - Monday, October 18, 2021 - link
TSMC's 2nm using GAA is going to be insane especially for SRAM density. When Intel//Samsung/TSMC are all on there GAA nodes densities will allow all sorts of fun like good gaming APU's.Yojimbo - Monday, October 18, 2021 - link
TSMC isn't just "estimating" introducing a node for the Apple order. That's a goal. If they miss it it's a failure. There have been at least one and possibly two delays associated with 3 nm now. First the Apple order was missed, and now probably another, since they are only recognizing revenue in 2023. At least they are giving more clarity on the extent of the delay and it's bigger than many were anticipating. It's looking a bit like a 6 month delay now.Curiousland - Monday, October 18, 2021 - link
Revenue in 2023 means process start in 4Q2021 since it takes 3+ months get N3 chips finished from start. So the timeline of 2H2021 mass production is met.Yojimbo - Monday, October 18, 2021 - link
I don't know the normal revenue-recognition for TSMC, but I doubt it's 3 months from wafer start. The "timeline is met". In other words, it can be spun.. Mass production was supposed to start in Q2. Then it was 2H 2022. Now it apparently will be Q4. But when in Q4? Since revenue recognition won't take place until 2023 it's likely not early Q4. Who knows, maybe it will be December 20th 2022? That will "meet the timeline". Or, now that they've shifted from talking about 2H 2022 production to Q1 2023 revenue recognition they don't need to actually have volume production start in 2022 at all to "meet the timeline". When they successfully start recognizing the revenue in Q1 2023 they will declare success regardless.whatthe123 - Tuesday, October 19, 2021 - link
They named 2H 2022 as the shipment date because that's Apple's cadence. they have missed the date, there's no dancing around it.TheJian - Wednesday, November 3, 2021 - link
Didn't miss the date if apple just didn't buy enough 3nm to launch with NOT figuring Intel would jump in and buy as much 3nm as possible to screw all involved. Brilliant. :) Apple is dancing because they thought they had 3nm for macs next year, now AMD is dancing because apple took the 4nm they hoped to have. My guess is AMD is least prepared to deal with what has happened in the last 6-9 months. NV/Intel/Apple just launch other parts they have stored up just in case of crap like this. AMD on the other hand does each design as needed pretty much. They don't have the NET INCOME to FART around with test projects or just in case designs. Now NV is coming back big for gpu at TSMC also, so everyone is WAFER-FU fighting...LOL. Good times. Intel/MU probably both crack 100 before next xmas (not this year, DDR5 rollout etc next year causes it, higher margin new junk to pitch).Tesla 4000 by 2025 end. 10K by 2030. :) Unless the govt figures out how to kill him, or big pharma, food industry, unions, battery makers, car makers, battery pack makers, etc etc etc...Jeez, this guy better have good protection, trying to take out wasteful multi-billion dollar markets can be VERY dangerous. These people at the FDA, FBI, CIA, NSA, DOJ etc etc, all have no problem knocking off anyone that gets in their POWER way. They have all proven themselves to be ENEMIES of the state run probably by vatican/soros etc from DC (that little place that is a vatican business, running the fake USA). hmm...Just saying. 1870act/1871 act....You have to look up the first act to get what they really did in the 2nd. Traitors.
Santoval - Monday, October 25, 2021 - link
'There is plenty of room at the bottom' but not *that* much room left. The limits of quantum mechanics are either an enemy (for conventional FETs) or a friend (for quantum FETs, or QFETs, which employ quantum tunneling).QFETs work better, not worse, with cutting edge nodes because their very fine features make quantum tunneling easier and more efficient. Although they employ quantum tunneling (QT) they are not quantum computers; they use QT as an optimization technique at the transistor level, not as a high level operating principle.
Even QFETs, though, have size limits. And before those limits are reached the economics of building fabs with.. subarmstrong process nodes will stop making sense, since they will always have a negative ROI.
geoxile - Monday, October 18, 2021 - link
It wasn't that many generations ago when Intel appeared to be the untouchable leader, or when Samsung 14nm was beating TSMC 16nm on apple chips. Anything can happen and a single node can make all the difference.Curiousland - Monday, October 18, 2021 - link
Not true. It was well known that the performance of TSMC 16nm beat Samsung 14nm, so after that Apple went TSMC exclusively.MetaCube - Friday, October 22, 2021 - link
"when Samsung 14nm was beating TSMC 16nm on apple chips."Never ?
name99 - Monday, October 18, 2021 - link
(a) What mistakes? "Many observers, however, expected these chips to ship in late 2022."TSMC can't change the rate at which ASML produces EUV machines, and that is the primary gating factor right now for everyone. It makes zero sense, under conditions where they can't even meet the demand for 5nm given the number of EUV layers, to promise, and make attempts to rush, a 3nm that uses even more layers.
(b) TSMC missing iPhones and A16 is hardly a catastrophe. We have been through this before -- look at A10 (on 16nm) and A10X (on 10nm). Apple will ship A16 on N5PP or N4 or whatever makes the most sense, and they will also schedule some chip (maybe M2/M1X/whatever the Mac Pro chip will be called) for N3.
TSMC have a wide range of clients -- which is how they get value from processes all the way from leading edge and leading edge-1 back to 90nm and older. And APPLE have a range of SoCs -- which is how they can always find something to schedule on the very latest leading edge.
(c) Apart from lithography, there are two interesting new developments coming up. One is GAA, one is BPR. It's basically a business decision as to which you implement first. SS has decided for GAA first, TSMC has decided for BPR first. You can fault this ordering; what you can't do is both complain that SRAM density is not increasing very much node to node AND complain that TSMC is introducing BPR first, given that a primary consequence of BOR should be to free up some space in the metal layers and thus help SRAM shrink a little...
Vitor - Monday, October 18, 2021 - link
Kinda sad how we really are reaching the limits of tech and struggling to achieve better processing power. But I guess 2nm will be enough for 4k/120fps and I can settle with that.FreckledTrout - Monday, October 18, 2021 - link
Probably 2nm will allow that in an APU and no I'm not joking. The densities on 2nm are going to be insane.sharath.naik - Monday, October 18, 2021 - link
Agree.. ram on CPU chip will allow for performance jumps in bandwidth restricted processing like GPU. Not to mention PC motherboard shrinking to tiny sizes in laptops. May be we will have Nano-itx boards and cases rise in demand.nandnandnand - Monday, October 18, 2021 - link
What Trout said. At least 4K60 in APUs (RDNA 5?). Highest end will be doing better than 8K120.FreckledTrout - Monday, October 18, 2021 - link
Don't forget Intel. Intel and AMD will have exceptional gaming APU's by 2026. Mark my words.Wereweeb - Monday, October 18, 2021 - link
No, 2025 APU's will not play 2025 games at 4K.Teckk - Monday, October 18, 2021 - link
2023 will be the year for Intel to claw back on the process, if they can manage to.Also, "undisclosed client" is Apple for iPhone 15? lol
shabby - Monday, October 18, 2021 - link
Apple usually ships iPhones in 4th quarter, not first. But they usually get first dibs so who knows.Teckk - Monday, October 18, 2021 - link
That was not a very serious comment but considering that Apple starts selling iPhones in September, having them shipping & available by the end of March would be good I guess. Also, they have a long cycle time due to complexity so, maybe they don't wanna risk it?There's always AMD, Intel also maybe - TSMC won't have trouble selling those for sure!
NickConrad - Monday, October 18, 2021 - link
"in the first quarter of 2023. Many observers, however, expected these chips to ship in late 2022."Those are the same thing. FYQ1 begins in October 2022.
Yojimbo - Monday, October 18, 2021 - link
The US government's fiscal year starts in October. The point of a "fiscal year" is that every company can have its own. TSMC's fiscal 2020 ended December 31, 2020. So TSMC's fiscal year is apparently very well aligned with the calendar year.Wereweeb - Monday, October 18, 2021 - link
Big insular seppo energyliahos1 - Monday, October 18, 2021 - link
So if Intel stays on their Node roadmap (big if, I know), this means they are behind TSM by 1-2 quarters for 3nm and should be ahead of TSM by 2024 (20ang) ?shabby - Monday, October 18, 2021 - link
Lol if, when they stay on their roadmap for a few nodes then you say if, currently intels if has no value.Yojimbo - Monday, October 18, 2021 - link
It's not a big if. All indications are that Intel has confidence in its roadmap. I doubt Pat Gelsinger signed on to be a fall guy.Of course unforeseen complications can happen. But that's different from what was going on with Intel before. The unforeseen complications happened once or twice but Intel tried to obscure it while they were hoping they'd find a fix. That could very well be something going on with TSMC now, in fact. Notice that it's now slipped from 4 months delay to 6 months delay. Probably that's where it will stay, but TSMC management would likely be just as dodgy as Intel's was if there were some major issues to work out.
The narrative that Intel has suddenly forgot how to do semiconductor engineering and TSMC has suddenly become an infallible leader is nonsense. It was only 2014 when TSMC last ran into major problems. But Samsung and GlobalFoundries did as well. Only Intel navigated the planar to finfet transition well and since Intel did not have a big foundry business TSMC was not burned by their poor execution.
Wereweeb - Monday, October 18, 2021 - link
TSMC's FinFET nodes are usually one step ahead of it's competitors in performance, so they can afford to be conservative and take it slow with GAA. They could lose some space in ultra-low-power applications where GAA will shine, but sticking to FinFET means both cheaper wafers and more wafers. That in turn means they'll make healthy profits, while Samsung and Intel will be stuck with yield issues and an expensive and slow manufacturing process. By then TSMC will have to transverse a path already explored by the industry.Even if Samsung and Intel catch up in terms of node density and performance, TSMC will have more money to invest in R&D, which they could use to offer something neither of the other foundries have. And that's important: as Moore's Law decays in it's coma, the question of technology leadership will get more and more complex. It might soon be decided not necessarily by the silicon node, but by the maturity and cost/effectiveness of packaging technologies, or through the integration of new materials, equipment or techniques which make production easier (E.g. TSMC's EUV masks)
Yojimbo - Monday, October 18, 2021 - link
TSMC never actually said N2 would be released in 2025. Wei only said (implicitly) that N2 would be TSMC's most advanced node in 2025. It does make sense that N2 would be released in 2025 because it doesn't make sense to miss the Apple order in 2024, assuming they can continue to win it. But neither the schedule of N3E nor the fact that Wei implied N2 would be TSMC's more advanced node in 2025 should be used to infer conclusively that N2 will only appear in 2025.BaronMatrix - Wednesday, October 20, 2021 - link
BaronMatrix is back...Admittedly, TSMC won't need 2nm before 2025 since AMD will continue with Zen3 while Apple will sharpen 3nm next year,,, AMD hasn't released Zen4 yet so with that, enhanced 7nm and 5nm (it was said that AMD had their own process), that will last for 2 years or more...
With RDNA2 APUs AMD can eliminate certain low end GPUs... And 5nm should make an MI400+ since compute doesn't require additional processing for frames... They should be able to get 30x25 with 3nm in 2025 with chiplets and HBM (created by AMD and Hynix)..
And Apple probably won't need a full node by 2023 so the roadmaqp seems to be more a design time issue with smaller geometries and EUV... AMD and Apple have multiple choices...Apple took the perf/w crown but 5nm should be s good for AMD with ~85% and last until maybe 2nm...
We don't know if Nvidia will fully move back to TSMC but the 5/4/3nm processes will allow for more dies/wafer and/or more logic...
bfonnes43 - Tuesday, October 26, 2021 - link
So, can we look forward to seeing this in the RTX 6090? :-)