Anton, does TSMC not only Fab but also have the Assembly/Test piece of the flow as well? E.g. Apple or AMD (ones not dependent on GF I/O dies) gets the complete package out of TSMC?
Also "advanced technology" and "specialty technology" haha that's good.
In the very recent Taiwanese local news I read, the site for 3nm fab in Tainan Science Park is under construction (with pictures). Local government is also helping acquiring more lands for the Park ready for 2nm fab. So who will win the 3-nm battle is yet to be determined. Note not like samsung, TSMC is usually keeping lower profile and won't say too much till they are confidence and with concrete plan for HVM.
I realize the feature size between TSMC and Intel is not really comparable at 14nm, 10nm, etc. but how did Intel fall so far behind. They seem to be struggling (still) with 10nm while TSMC is moving to 5nm? Who would have thought it 10 years ago? Is Intel really as far behind in process technology as it seems to someone not really keen on this business?
Yes, but Intel is still promising 7nm for 2021. Let's see how that turns out. Intel 10nm should be around TSMC's 7nm and Intel's 7nm should be like TSMC's 5nm. Which still means that Intel is behind in any case, but it still has a chance to at least get on more or less equal footing. Though at this point I don't really trust Intel projections.
it's always interesting to read "Intel 10nm should be around TSMC's 7nm and Intel's 7nm should be like TSMC's 5nm" - how can one number be like a different one? what qualifies the difference in measurement, and shouldn't the measurement be the reported #? 10<> 7 and 7 <> 5
The numbers used have lost any relation to any real physical dimensions of the actual circuitry. At this pace, soon (in ~10 years) we will be at subatomic sizes, which obviously will be even more manifestly meaningless. Chalk it up to marketing "genius"...
It’s probably better to look at figures of merit of the transistor structure than any particular gate pitch, ie the frequency/voltage curve. Typically smaller gate pitch led to better FOM’s in the past but now with technologies like FinFET and GAA, two transistors with the same pitch can have widely different performance.
Those kind of comments made sense years ago. Now the real test is how well the node performs in benchmarks and how much it costs e.g. Intel (presumably) never used 10nm for desktop because it's performance/yield was worse than 16nm++++.
There is no standard measurement. Mostly because the transistors went 3D with FinFET and the nm size stopped mattering at that point. Really we as techies should give our own ratings to these based of densities. TSMC's 7nm node only has a very few up front features at 7nm the rest of the chip has larger features which is why Intel's 10nm is more like TSMC 7nm.
I would easily bet half of my money in the bank that Intel are not going to release any 7nm part in 2021, not even a "let's pull the wool over the eyes of our investors" semi-disabled 2-core i3 part, like they did with Cannon Lake. I rather predict a low volume 7nm "launch" (i.e. a little more than a paper launch) in 2022 and a high volume (the one that only matters) release in 2023. By that time TSMC will be at 3nm and GAA, and Samsung should not be far behind. If Intel do not pull an early 2021 high volume release of 7nm parts they will lose the next round of "process node wars" as well. And there's no way in hell they're going to achieve that feat.
That is because you don't have a clue how manufacturing works and how big of a problem you have when you test a process that doesn't work as you expect. There are two aspects: 1. marketing guys are trying to sugar coat the problem as much as possible. They get feedback from fab guys that they try something that might work. They translate it as the process will be up and running in one year. So on and so forth. 2. Fab engineers that - if asked - but no one will let them talk to the general public will say that maybe they don't know yet why it doesn't work or they understand why it doesn't, but they are trying to find a fix. Probably at some point they said, screw it, lets start from scratch with 10nm and build it on the base of 14nm. That is what we have today and this takes a LOT of time to get up and running. So yeah, if you wanna believe the marketing guys talking about tech stuff, do it as you want. But I can tell you that when a process starts to work and the yields stabilize that means the dudes at the fab are starting to get the grasp of things and will work their way from there. And it seems like 10nm stuff is now stable. Ice Lake products are plentiful, prices are decent so I don't see any reason they won't continue executing well with 7nm. From their earnings call, they should have a GPU at 7nm (ponte vechio) in 2021 and a CPU in 2022. Which does fit with TSMC 5nm process.
and most of the public also doesnt understand it too.. go accuse them as well.... whats your point ?? most of the public.. seems to be tired of intels lies and its " sugar coating " as you put it.. IF intel had been open and honest about how much they screwed up worh 10nm.. and admit they were too aggressive with it.. then maybe there would be a bit more understanding.. but nope.. they kept lieing about it.. kept saying its on track.. and it still is broken...
Intel have other paths to win back CPU than node shrinks. If they can pull of a 3d stacked win it will work as well. If they do none of it, they are heading for mass layoffs and a stock price dive.
maybe i dont. but a company like intel.. with all its money and personel.. cant get something to work ?? come on... 10nm must of been screwed up big time...
I'm inclined to disagree - I think they'll perform exactly the same sort of 7nm "release" in 2021 that they did with Cannon Lake, with a slow ramp into 2022.
We'll see, though. At this stage in semiconductor development, I think it's safe to say that all bets are off.
1. Just like 10nm in 2019 from Intel, in reality it isn't any where near the volume you would expect. They barely ship it out in last quarter.
2. When TSMC state 7nm, and HVM that means there will be 50M+ not only fabbed or shipped, but SOLD on the market. That is going like clock work from Apple.
So looking at years shipped doesn't realistically show the gap between Intel and TSMC.
I would point out again, ASML is the only EUV machines supplier. TSMC has been getting as many as they could, you could bet Intel 7nm won't have any volume from the start either.
Rumor has it they fell behind because they thought they could still progress to 10nm and beyond without buying expensive eternal tools (e.g. EUV tools).
Well for 10nm it's not a rumor. They went with quad patterning which worked except scaling the frequencies was severely hampered. I have a feeling they were thinking EUV for 7nm all along as they were already working on it a long time ago.
Not exactly happy with the Amount. $1B Capex increase is tiny in comparison. The demand for 7nm is much higher than expected. From Network Equipment, 5G Modem, Nvidia GPU, AMD CPU, etc... and hundreds of other queueing up.
But knowing TSMC they are very conservative, so an increase of Capex is at least acknowledgement to supply and demand issues.
Poor GloFo, they were so quick in cancelling 7nm right before ramping risk production having two ASML EUV machine already installed in factory... what a smart move.
They'd have to have still researched and developed a 7nm process node, it was the right call for them to fold. Being at the forefront takes a lot of capital for a risky venture, TSMC managed to get it right at a very good timing, only other with as much capital is Samsung and Intel.
Note that it was not even necessary to mention competition from Intel 7nm because Intel is still rehashing 14nm, and has not released any hedt or server 10nm. So Intel is going to go from two to three nodes behind this year.
If they will fix the reported frequence regression versus 7nm this will become true. If data reported by Anandtech in a recent article is true, 5nm will be one of the slower processes available on the market. Density is good, but without raw clock speed you are lost in action badly.
Mind posting what article? That sounds wrong, I don't remember AT stating lower frequencies for 5nm, even this article states +20% performance at same power which should usually be at least a couple of percent higher frequency (no, not linear).
"Density is good, but without raw clock speed you are lost in action badly."
Yes, that's why AMD's 5Ghz Piledriver chip totally dominated the market. Those denser, more complex 14nm Intel chips just couldn't keep up with the raw clock speed of AMD's 28nm beast.
(This is sarcasm. Also, Gondalf has previously claimed that Ice Lake beats other chips due to IPC so consistency is not his strong point)
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
40 Comments
Back to Article
Teckk - Wednesday, January 22, 2020 - link
Anton, does TSMC not only Fab but also have the Assembly/Test piece of the flow as well? E.g. Apple or AMD (ones not dependent on GF I/O dies) gets the complete package out of TSMC?Also "advanced technology" and "specialty technology" haha that's good.
s.yu - Wednesday, January 22, 2020 - link
IIRC 2 years ago TSMC was expected to maintain the lead until 5nm, after which Samsung might gain an advantage with 3nm and GAA.Curiousland - Wednesday, January 22, 2020 - link
In the very recent Taiwanese local news I read, the site for 3nm fab in Tainan Science Park is under construction (with pictures). Local government is also helping acquiring more lands for the Park ready for 2nm fab. So who will win the 3-nm battle is yet to be determined. Note not like samsung, TSMC is usually keeping lower profile and won't say too much till they are confidence and with concrete plan for HVM.ksec - Thursday, January 23, 2020 - link
And we are now pretty sure GAA wont be ready for 3nm in volume. So it looks like Samsung made the wrong bet yet again.s.yu - Thursday, January 23, 2020 - link
Yeah, I said might, I don't have that much confidence in Samsung either.Hulk - Wednesday, January 22, 2020 - link
I realize the feature size between TSMC and Intel is not really comparable at 14nm, 10nm, etc. but how did Intel fall so far behind. They seem to be struggling (still) with 10nm while TSMC is moving to 5nm? Who would have thought it 10 years ago? Is Intel really as far behind in process technology as it seems to someone not really keen on this business?ET - Wednesday, January 22, 2020 - link
Yes, but Intel is still promising 7nm for 2021. Let's see how that turns out. Intel 10nm should be around TSMC's 7nm and Intel's 7nm should be like TSMC's 5nm. Which still means that Intel is behind in any case, but it still has a chance to at least get on more or less equal footing. Though at this point I don't really trust Intel projections.Toadster - Wednesday, January 22, 2020 - link
it's always interesting to read "Intel 10nm should be around TSMC's 7nm and Intel's 7nm should be like TSMC's 5nm" - how can one number be like a different one? what qualifies the difference in measurement, and shouldn't the measurement be the reported #? 10<> 7 and 7 <> 5boeush - Wednesday, January 22, 2020 - link
The numbers used have lost any relation to any real physical dimensions of the actual circuitry. At this pace, soon (in ~10 years) we will be at subatomic sizes, which obviously will be even more manifestly meaningless. Chalk it up to marketing "genius"...flgt - Wednesday, January 22, 2020 - link
It’s probably better to look at figures of merit of the transistor structure than any particular gate pitch, ie the frequency/voltage curve. Typically smaller gate pitch led to better FOM’s in the past but now with technologies like FinFET and GAA, two transistors with the same pitch can have widely different performance.Amandtec - Thursday, January 23, 2020 - link
Those kind of comments made sense years ago. Now the real test is how well the node performs in benchmarks and how much it costs e.g. Intel (presumably) never used 10nm for desktop because it's performance/yield was worse than 16nm++++.FreckledTrout - Thursday, January 23, 2020 - link
There is no standard measurement. Mostly because the transistors went 3D with FinFET and the nm size stopped mattering at that point. Really we as techies should give our own ratings to these based of densities. TSMC's 7nm node only has a very few up front features at 7nm the rest of the chip has larger features which is why Intel's 10nm is more like TSMC 7nm.yeeeeman - Friday, January 24, 2020 - link
It is marketing not real physical sizes they talk about.levizx - Saturday, January 25, 2020 - link
So you'll also find it interesting Cortex A76@3GHz is around Intel@1GHz?Santoval - Wednesday, January 22, 2020 - link
I would easily bet half of my money in the bank that Intel are not going to release any 7nm part in 2021, not even a "let's pull the wool over the eyes of our investors" semi-disabled 2-core i3 part, like they did with Cannon Lake. I rather predict a low volume 7nm "launch" (i.e. a little more than a paper launch) in 2022 and a high volume (the one that only matters) release in 2023. By that time TSMC will be at 3nm and GAA, and Samsung should not be far behind.If Intel do not pull an early 2021 high volume release of 7nm parts they will lose the next round of "process node wars" as well. And there's no way in hell they're going to achieve that feat.
Hulk - Wednesday, January 22, 2020 - link
Actually I think that's a good prediction based on past history.FreckledTrout - Thursday, January 23, 2020 - link
Not really. How many times has Intel goofed two generations? Never. Its his money but I wouldn't bet against Intel faltering two times in a row.Korguz - Thursday, January 23, 2020 - link
maybe.. but seems intels words.. are worth nothing now.. most are " ill believe it when it happens " view with intel now...yeeeeman - Friday, January 24, 2020 - link
That is because you don't have a clue how manufacturing works and how big of a problem you have when you test a process that doesn't work as you expect.There are two aspects:
1. marketing guys are trying to sugar coat the problem as much as possible. They get feedback from fab guys that they try something that might work. They translate it as the process will be up and running in one year. So on and so forth.
2. Fab engineers that - if asked - but no one will let them talk to the general public will say that maybe they don't know yet why it doesn't work or they understand why it doesn't, but they are trying to find a fix. Probably at some point they said, screw it, lets start from scratch with 10nm and build it on the base of 14nm. That is what we have today and this takes a LOT of time to get up and running.
So yeah, if you wanna believe the marketing guys talking about tech stuff, do it as you want. But I can tell you that when a process starts to work and the yields stabilize that means the dudes at the fab are starting to get the grasp of things and will work their way from there. And it seems like 10nm stuff is now stable. Ice Lake products are plentiful, prices are decent so I don't see any reason they won't continue executing well with 7nm. From their earnings call, they should have a GPU at 7nm (ponte vechio) in 2021 and a CPU in 2022. Which does fit with TSMC 5nm process.
Korguz - Friday, January 24, 2020 - link
and most of the public also doesnt understand it too.. go accuse them as well.... whats your point ?? most of the public.. seems to be tired of intels lies and its " sugar coating " as you put it.. IF intel had been open and honest about how much they screwed up worh 10nm.. and admit they were too aggressive with it.. then maybe there would be a bit more understanding.. but nope.. they kept lieing about it.. kept saying its on track.. and it still is broken...RSAUser - Sunday, January 26, 2020 - link
GPU is a lot more tolerant of faults, easier to fuse off and generally runs at a lower clock.Amandtec - Thursday, January 23, 2020 - link
Intel have other paths to win back CPU than node shrinks. If they can pull of a 3d stacked win it will work as well. If they do none of it, they are heading for mass layoffs and a stock price dive.FreckledTrout - Thursday, January 23, 2020 - link
I wouldn't bet against Intel. History has shown when Intel messes up they come back strong.Korguz - Thursday, January 23, 2020 - link
intel has been messing up for like 5 years now, when will they come back ??? they have also been lieing for about that long too...yeeeeman - Friday, January 24, 2020 - link
You say that because you don't understand a thing about manufacturing.Korguz - Friday, January 24, 2020 - link
maybe i dont. but a company like intel.. with all its money and personel.. cant get something to work ?? come on... 10nm must of been screwed up big time...Spunjji - Monday, January 27, 2020 - link
I'm inclined to disagree - I think they'll perform exactly the same sort of 7nm "release" in 2021 that they did with Cannon Lake, with a slow ramp into 2022.We'll see, though. At this stage in semiconductor development, I think it's safe to say that all bets are off.
ksec - Thursday, January 23, 2020 - link
1. Just like 10nm in 2019 from Intel, in reality it isn't any where near the volume you would expect. They barely ship it out in last quarter.2. When TSMC state 7nm, and HVM that means there will be 50M+ not only fabbed or shipped, but SOLD on the market. That is going like clock work from Apple.
So looking at years shipped doesn't realistically show the gap between Intel and TSMC.
I would point out again, ASML is the only EUV machines supplier. TSMC has been getting as many as they could, you could bet Intel 7nm won't have any volume from the start either.
Amandtec - Thursday, January 23, 2020 - link
Rumor has it they fell behind because they thought they could still progress to 10nm and beyond without buying expensive eternal tools (e.g. EUV tools).FreckledTrout - Thursday, January 23, 2020 - link
Well for 10nm it's not a rumor. They went with quad patterning which worked except scaling the frequencies was severely hampered. I have a feeling they were thinking EUV for 7nm all along as they were already working on it a long time ago.ksec - Thursday, January 23, 2020 - link
Not exactly happy with the Amount. $1B Capex increase is tiny in comparison. The demand for 7nm is much higher than expected. From Network Equipment, 5G Modem, Nvidia GPU, AMD CPU, etc... and hundreds of other queueing up.But knowing TSMC they are very conservative, so an increase of Capex is at least acknowledgement to supply and demand issues.
del42sa - Thursday, January 23, 2020 - link
Poor GloFo, they were so quick in cancelling 7nm right before ramping risk production having two ASML EUV machine already installed in factory... what a smart move.RSAUser - Sunday, January 26, 2020 - link
They'd have to have still researched and developed a 7nm process node, it was the right call for them to fold. Being at the forefront takes a lot of capital for a risky venture, TSMC managed to get it right at a very good timing, only other with as much capital is Samsung and Intel.Spunjji - Monday, January 27, 2020 - link
It really was a smart move. Chasing losses is a loser's strategy when you don't have the capital to stay in the game.Machinus - Thursday, January 23, 2020 - link
Note that it was not even necessary to mention competition from Intel 7nm because Intel is still rehashing 14nm, and has not released any hedt or server 10nm. So Intel is going to go from two to three nodes behind this year.Gondalf - Thursday, January 23, 2020 - link
If they will fix the reported frequence regression versus 7nm this will become true. If data reported by Anandtech in a recent article is true, 5nm will be one of the slower processes available on the market. Density is good, but without raw clock speed you are lost in action badly.RSAUser - Sunday, January 26, 2020 - link
Mind posting what article? That sounds wrong, I don't remember AT stating lower frequencies for 5nm, even this article states +20% performance at same power which should usually be at least a couple of percent higher frequency (no, not linear).Spunjji - Monday, January 27, 2020 - link
"Density is good, but without raw clock speed you are lost in action badly."Yes, that's why AMD's 5Ghz Piledriver chip totally dominated the market. Those denser, more complex 14nm Intel chips just couldn't keep up with the raw clock speed of AMD's 28nm beast.
(This is sarcasm. Also, Gondalf has previously claimed that Ice Lake beats other chips due to IPC so consistency is not his strong point)
Korguz - Tuesday, January 28, 2020 - link
thats where you are wrong Spunjji... gondalf IS consistent, he is consistent spreading fud, BS, misinformation, etc :-)Foeketijn - Wednesday, January 29, 2020 - link
It would be interesting if somehow TSMC takes GoFlo along. In the form of a takeover or Partnership.