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

Page 12 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

DisEnchantment

Golden Member
Mar 3, 2017
1,590
5,722
136
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.

1587737990547.png
N7 performance is more or less understood.
1587739093721.png

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

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.


1587739615344.png

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.
 

LightningZ71

Golden Member
Mar 10, 2017
1,627
1,898
136
There we comments attributed to TSMC that Intel would have to pay to play. Not sure how that worked out. Very curious about what was discussed in Gelsinger's recent meeting with TSMC execs.



Yes. I waffled on this at first, but now believe that, in the US, we need to allocation some of our Semicon stimulus money to companies from other counties - so long as they start building plants in the US within a couple of years. Samsung and TSMC would certainly qualify and have already committed to new building new plant(s). Given the national security implications, the US military has been pushing for this. We also need to get more of the supply chain pipeline on US soil.
I don't want to stray too political on this forum, but, essentially, due to how uncompetitive the US is with respect to manufacturing labor costs (in other words, we actually have laws requiring somewhat humane working conditions and terms, not that they are always well enforced) until the U.S. Government decides that IC production is a national security level infrastructure and strategic asset and creates a federal foundry, there is no hope that the US will ever be a major contributor as a LEADING EDGE foundry source. Its just not going to happen. Why would a Samsung or a TSMC want to invest the hundreds of billions of dollars into facilities and infrastructure to have that much production capacity in the US when they already have most all that they need in Taiwan and South Korea and nearby asia? It would either sit heavily underutilized, or it would so heavily outproduce the market's needs that it would shove ASPs down to a point where there is hardly any money to be made in leading edge development. Sure, we might get some "supplemental" production facilities at what will be, at the time of their opening, recently relevant trailing nodes (seriously, N5 in two to three years won't be a leading edge node), but that's about it outside of Intel, who has a lot of their own manufacturing overseas as well.
 

jpiniero

Lifer
Oct 1, 2010
14,509
5,159
136
Labor costs really isn't a problem for the fabs themselves. Now I could see advantages being close to China though, given that everything is assembled there.
 

Doug S

Platinum Member
Feb 8, 2020
2,201
3,405
136
I had read about Forksheets, but didn't understand the implication for added scaling then. The clear presence here of a high-K dielectric, vs what I thought was poly, changes the narrative. Now, add CFET's to the mix (lower drive currents, it would seem, unless some high mobility exotic materials enter the mix). I'm not sure how CFETs work yet - I'd like to see a 3-D cell. Some of the slides are too small, sadly.


I suspect the reason for the dearth of really good diagrams and technical explanations for CFETs is 1) no one has finalized exactly how theirs will be made and/or 2) if they have, they don't want to tip their hand to their rivals until it is too late to be of any benefit to them.
 
  • Like
Reactions: Tlh97 and Ajay

Khato

Golden Member
Jul 15, 2001
1,202
240
106
There is an important distinction on labor costs between fab production and assembly/test. Using Intel as an example here as they have a simple page documenting such - https://www.intel.com/content/www/us/en/support/articles/000015142/programs.html Notice that all the fabs (save for the NAND-only fab in China) are in 'high' labor cost areas while assembly and test are not? I believe there are two primary factors to this. One is cost of infrastructure - with how much a fab costs, how much difference does saving a bit on labor make? Second factor is quantity of workers required, which I'd be interested to hear actual figures if anyone has them? My understanding has always been that assembly and test requires more workers than a fab by a fair amount. There may well be a third factor of playing games with taxes.
 

maddie

Diamond Member
Jul 18, 2010
4,722
4,627
136
You know the Spanish Flu was in the 20th century, right?
I guess I made the common mistake of assuming a shared experience. I live in the Caribbean, not the US or Europe. Many are getting impoverished. Stimulus checks? What's that? The big "C" doesn't worry me to any great degree, except for the huge response damage being done to local economies.

Really for P&N, but wanted to clear the misunderstanding.
 

Ajay

Lifer
Jan 8, 2001
15,332
7,792
136
I don't want to stray too political on this forum, but, essentially, due to how uncompetitive the US is with respect to manufacturing labor costs (in other words, we actually have laws requiring somewhat humane working conditions and terms, not that they are always well enforced) until the U.S. Government decides that IC production is a national security level infrastructure and strategic asset and creates a federal foundry, there is no hope that the US will ever be a major contributor as a LEADING EDGE foundry source. Its just not going to happen. Why would a Samsung or a TSMC want to invest the hundreds of billions of dollars into facilities and infrastructure to have that much production capacity in the US when they already have most all that they need in Taiwan and South Korea and nearby asia? It would either sit heavily underutilized, or it would so heavily outproduce the market's needs that it would shove ASPs down to a point where there is hardly any money to be made in leading edge development. Sure, we might get some "supplemental" production facilities at what will be, at the time of their opening, recently relevant trailing nodes (seriously, N5 in two to three years won't be a leading edge node), but that's about it outside of Intel, who has a lot of their own manufacturing overseas as well.
Uhm, okay.
 

Doug S

Platinum Member
Feb 8, 2020
2,201
3,405
136

Ajay

Lifer
Jan 8, 2001
15,332
7,792
136
I don't get it. That post didn't provide any new information, TSMC has been saying for well over a year that N3 would be H2 2022. They are talking like the delay is something that just happened a couple months ago.
Well, the author should have made the point that this was retrospective. That's how I took it for obvious reasons.
 

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
I don't get it. That post didn't provide any new information, TSMC has been saying for well over a year that N3 would be H2 2022. They are talking like the delay is something that just happened a couple months ago.

It reveals that Intel may be getting some N4 wafers, and provides some otherwise non-existent commentary on how Intel might be getting N3 before AMD without AMD really being bothered. Sort of puts to rest the idea that Intel is trying to buy out AMD's share of N3. I still don't know if I agree that TSMC is trying to cripple Intel's manufacturing by making of them addicts.
 

Panino Manino

Senior member
Jan 28, 2017
813
1,010
136
Seems like Samsung is still lagging behind TSMC.


Samsung reportedly encountered issues with integrated graphics frequency, claims “Ice Universe”. From the initial target of 1.9 GHz, the chip has been subsequently lowered to 1.3 GHz. The company is allegedly working around the clock to at least reach the 1.49 GHz frequency target. claims the leaker.
 

LightningZ71

Golden Member
Mar 10, 2017
1,627
1,898
136
Samsung hasn't exactly been closing the gap with TSMC lately. I'm not even sure that they have a node that is fully competitive with Intel's 10ESF/Intel7 in high volume production. From their releases, they can hit the density mark, but, power/performance numbers at those density levels seem to be... less than stellar.
 

DisEnchantment

Golden Member
Mar 3, 2017
1,590
5,722
136
Seems like Samsung is still lagging behind TSMC.

In the meanwhile, RDNA3 on TSMC N5 rumored to be hitting 3GHz stock, twice that of this Exynos GPU, and faster than the main CPU.:( (granted with a massive thermal envelope)
Architecturally it makes no sense to take this GPU to mobile, all the SFU in it will not even be triggered by mobile games. All that RT, massive geometry throughput, async compute etc will not even get triggered.
I hope they stripped it all off or at least planned some new use cases around it.

Samsung needs to keep trying, someday they will find success.
In the meanwhile this chip could be sent to Automotive.
 
  • Like
Reactions: Tlh97

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
Samsung hasn't exactly been closing the gap with TSMC lately. I'm not even sure that they have a node that is fully competitive with Intel's 10ESF/Intel7 in high volume production. From their releases, they can hit the density mark, but, power/performance numbers at those density levels seem to be... less than stellar.

We'll see how 4LPE fares once Snapdragon 8 Gen 1 hits the market/gets benchmarked some more. I think the Chinese market already has some Gen 1 phones coming this month. Funny how Qualcomm has already planned to move the same design to N4 for ~Q3 2022 though!

Also, @Panino Manino


Samsung refutes that report. Feb 8th is when we'll see it, and the phone, for the first time. Assuming they don't postphone the reveal for the phone.

More than most of us realize, I think. Isn't this chip on a new node? Not even competitive with N7 (unless it is an architectural issue).

The Exynos 2200 should be on Samsung 4LPE, which is meant to be a competitor to N5 and/or N4. Based on Anandtech's Exynos 2100/Snapdragon 888 review:


it looks like Samsung's 5LPE is roughly equivalent to N7 in performance.
 
Last edited:

Panino Manino

Senior member
Jan 28, 2017
813
1,010
136
Architecturally it makes no sense to take this GPU to mobile, all the SFU in it will not even be triggered by mobile games. All that RT, massive geometry throughput, async compute etc will not even get triggered.
I hope they stripped it all off or at least planned some new use cases around it.

I agree.
Wasn't Mali too big to the performance that it was giving?
Then why switch to RDNA2 that have those big CUs? Even if all went well with the process and performance is good, wouldn't this be at the coast of big chips?
I'm clad for AMD but it's hard to understand Samsung's choice.
 
  • Like
Reactions: scineram

DisEnchantment

Golden Member
Mar 3, 2017
1,590
5,722
136


TSMC is reportedly repurposing Fab12P8 and Fab12P9 for Intel N3 orders (including CPUs). These two are R&D fabs previously.
So there is big money indeed involved and Intel practically bankrolled these two fab phases being upgraded to N3. :oops:
One tidbit I can share is that both Fab12P8 (this one first) and Fab12P9 (later) have started construction already since last year and they are getting converted to N3 which means they are not too far from coming online.
Probably Fab12P8/9 will come online together with Fab18P7/8 late 2023.
So we can expect major capacity boost for N3 at TSMC in 2H23, starting early with Fab18P4/5/6 in 1H23 and then the ones above next

Good thing though is that Fab18P4/5/6/7/8 are still there unaffected for other customers using N3.

Also TSMC is reportedly planning a huge investment in N5 capacity expansion.

From latest earnings PR, TSMC investment for 2022 shot up from a projected 30B USD to 40-44B USD. Almost 50% jump. Holy smokes.
1642064345536.png

Good thing is that this will really push Intel and AMD competition architecturally to another level.

My estimates here
Fab18P5/6/7/8 = 4*40 --> 160k WPM
Fab12P8/9 = 2*20 --> 40k WPM
Fab18P4 = 0.5*40 --> 20k WPM (assuming P4 shared between N3 and N5)
220k WPM by end of 2023/Early 2024 for N3. That is bonkers

One thing for Intel though, over reliance on TSMC means their Fabs will die a slow death. But they are probably counting on the US Govt to bail them out on that if it happens.
 
Last edited:

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
Fab12P8/9 = 2*20 --> 40k WPM

Interesting. Initial rumours seemed to indicate that Intel would only be getting 20 kwpm N3. TSMC may have doubled their allocation.

One thing for Intel though, over reliance on TSMC means their Fabs will die a slow death. But they are probably counting on the US Govt to bail them out on that if it happens.

Uncle Sam never does a good job of checking the balance sheet of companies receiving Federal subsidies. Intel is indirectly funneling public money to TSMC (or will be).
 

maddie

Diamond Member
Jul 18, 2010
4,722
4,627
136


TSMC is reportedly repurposing Fab12P8 and Fab12P9 for Intel N3 orders (including CPUs). These two are R&D fabs previously.
So there is big money indeed involved and Intel practically bankrolled these two fab phases being upgraded to N3. :oops:
One tidbit I can share is that both Fab12P8 (this one first) and Fab12P9 (later) have started construction already since last year and they are getting converted to N3 which means they are not too far from coming online.
Probably Fab12P8/9 will come online together with Fab18P7/8 late 2023.
So we can expect major capacity boost for N3 at TSMC in 2H23, starting early with Fab18P4/5/6 in 1H23 and then the ones above next

Good thing though is that Fab18P4/5/6/7/8 are still there unaffected for other customers using N3.

Also TSMC is reportedly planning a huge investment in N5 capacity expansion.

From latest earnings PR, TSMC investment for 2022 shot up from a projected 30B USD to 40-44B USD. Almost 50% jump. Holy smokes.
View attachment 55865

Good thing is that this will really push Intel and AMD competition architecturally to another level.

My estimates here
Fab18P5/6/7/8 = 4*40 --> 160k WPM
Fab12P8/9 = 2*20 --> 40k WPM
Fab18P4 = 0.5*40 --> 20k WPM (assuming P4 shared between N3 and N5)
220k WPM by end of 2023/Early 2024 for N3. That is bonkers

One thing for Intel though, over reliance on TSMC means their Fabs will die a slow death. But they are probably counting on the US Govt to bail them out on that if it happens.
As a general rule of thumb, over reliance on any single source for critical materials is insane, especially in this case of political and possibly military conflict. Is there a word beyond insane?
 

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136


TSMC is reportedly repurposing Fab12P8 and Fab12P9 for Intel N3 orders (including CPUs). These two are R&D fabs previously.
So there is big money indeed involved and Intel practically bankrolled these two fab phases being upgraded to N3. :oops:
One tidbit I can share is that both Fab12P8 (this one first) and Fab12P9 (later) have started construction already since last year and they are getting converted to N3 which means they are not too far from coming online.
Probably Fab12P8/9 will come online together with Fab18P7/8 late 2023.
So we can expect major capacity boost for N3 at TSMC in 2H23, starting early with Fab18P4/5/6 in 1H23 and then the ones above next

Good thing though is that Fab18P4/5/6/7/8 are still there unaffected for other customers using N3.

Also TSMC is reportedly planning a huge investment in N5 capacity expansion.

From latest earnings PR, TSMC investment for 2022 shot up from a projected 30B USD to 40-44B USD. Almost 50% jump. Holy smokes.
View attachment 55865

Good thing is that this will really push Intel and AMD competition architecturally to another level.

My estimates here
Fab18P5/6/7/8 = 4*40 --> 160k WPM
Fab12P8/9 = 2*20 --> 40k WPM
Fab18P4 = 0.5*40 --> 20k WPM (assuming P4 shared between N3 and N5)
220k WPM by end of 2023/Early 2024 for N3. That is bonkers

One thing for Intel though, over reliance on TSMC means their Fabs will die a slow death. But they are probably counting on the US Govt to bail them out on that if it happens.

Intel won’t even use TSMC to produce half it’s products, let alone most. Right now, TSMC is supposed to make Intel GPUs. There are rumors of them making Intel CPUs, but if that happens (which I doubt) it will likely only be for one generation.
 

hemedans

Member
Jan 31, 2015
189
95
101
In the meanwhile, RDNA3 on TSMC N5 rumored to be hitting 3GHz stock, twice that of this Exynos GPU, and faster than the main CPU.:( (granted with a massive thermal envelope)
Architecturally it makes no sense to take this GPU to mobile, all the SFU in it will not even be triggered by mobile games. All that RT, massive geometry throughput, async compute etc will not even get triggered.
I hope they stripped it all off or at least planned some new use cases around it.

Samsung needs to keep trying, someday they will find success.
In the meanwhile this chip could be sent to Automotive.

Mali drivers are so bad, i dont know if samsung can fix that, with Amd drivers even if it perform worse on paper in real life its win For Samsung.

Like in Emulation flagship Exynos/kirin Lose to even lowend/midrange Qualcomm soc.
 

DisEnchantment

Golden Member
Mar 3, 2017
1,590
5,722
136
I have never seen mobile Gpu with That frequency. Adreno 650 (TSMC 7nm) boost to 670Mhz, most Gpu in mobile space are under 1ghz. What so special that samsung fail to overclock Gpu to 1.9ghz? Or i Miss something here?
It is about Samsung Foundry not able to meet design frequency and efficiency targets
For instance,
Samsung Foundry tells designer don't worry we can hit 2 GHz and power efficiency is so and so, then designer put 4 CUs because they expect 4 CUs to hit 2 GHz and the performance expectation would be met and design cooling accordingly.
However when the chip came back from the fab those 4 CUs can only hit 1.3 GHz and has bad efficiency. Then there is a problem of performance and throttling due to insufficient cooling as designed.

On the other hand, if Samsung Foundry tells max we can do is 1 GHz, then designer says OK fine we will put more CUs to get same throughput and beef up cooling.
Then there is no problem if said targets are worse because it is known from start and designer can adapt accordingly albeit with compromises but device will perform as expected.

Mali drivers are so bad, i dont know if samsung can fix that, with Amd drivers even if it perform worse on paper in real life its win For Samsung.
That may be be true, but the RDNA architecture is quite complex and bloated for mobile, in addition to what I just said like RT, Geometry Engine, Async compute, it has so many other features which make zero sense in mobile like GPU virtualization, mixed precision and RAS which ends up making the CU big.
To make a compelling selling point, Samsung must have more unique use cases to show which leverages these things that are present in RDNA2 otherwise with fewer CUs the Exynos will lose in a pure fp throughput due to Mali being able to be packed with more units for given area.
 

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
I have never seen mobile Gpu with That frequency. Adreno 650 (TSMC 7nm) boost to 670Mhz, most Gpu in mobile space are under 1ghz. What so special that samsung fail to overclock Gpu to 1.9ghz? Or i Miss something here?

AMD Rembrandt can hit 2.2 ghz with 12CUs and an 8 core CPU. That is 15W. Steam Deck has 8CUs that boost up to 1.6ghz with 4 Zen 2 cores. 4-15W.

The issue is the foundry. Had the chip been made on TSMC N5, they would have hit their targets.
 
  • Like
Reactions: Tlh97

hemedans

Member
Jan 31, 2015
189
95
101
AMD Rembrandt can hit 2.2 ghz with 12CUs and an 8 core CPU. That is 15W. Steam Deck has 8CUs that boost up to 1.6ghz with 4 Zen 2 cores. 4-15W.

The issue is the foundry. Had the chip been made on TSMC N5, they would have hit their targets.
But mobile soc rarely reach those figures, usually they consume between 3W to 5W when you play heavy game like Genshini Impact.

Probably Samsung overpromise with 2Ghz figures but still its impressive if they can Cool 1.2ghz Gpu in smartphone.
 
  • Like
Reactions: Tlh97 and Kaluan