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

Page 4 - 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.
 

DrMrLordX

Lifer
Apr 27, 2000
21,583
10,785
136
@LightningZ71

At least based on TSMC's public commentary about the performance advantages of advancing nodes (N7P, N7+, and N5), AMD's continued success hinges on using the latest node available. Obviously it will take them longer to get to N5 than the phone SoC firms, but they have to go there to continue moving forward at the same pace that they established three years ago.
 

Doug S

Platinum Member
Feb 8, 2020
2,203
3,407
136
@LightningZ71

At least based on TSMC's public commentary about the performance advantages of advancing nodes (N7P, N7+, and N5), AMD's continued success hinges on using the latest node available. Obviously it will take them longer to get to N5 than the phone SoC firms, but they have to go there to continue moving forward at the same pace that they established three years ago.

There's no particular reason it should take AMD longer to get on a node than Apple does, other than that it is widely believed Apple pays TSMC up front for a guaranteed number of wafers off a new process which leaves few for others when it first ramps.

So they wouldn't be shipping 5nm parts early this fall like Apple will be, but there's no reason they shouldn't be able to ship 5nm stuff a few months later like early 2021 if they wanted. This is made easier by the fact they wouldn't need anywhere near the number of wafers Apple does to start with, because they only need the lower volume high end / high profit SKUs to be made on the leading edge, and the midrange can be made on N+1 and the low end on N+2. That's basically how Intel rolls things out, they use a new process selectively at first where it can make the most money early on i.e. laptop CPUs and high end desktop stuff. Server CPUs take longer to validate so they are never going to be first out the gate for a new process for anyone.

Obviously I'm only talking about AMD CPUs here, since thanks to chiplets AMD isn't using any ridiculously large dies for those like they do for GPUs (or have they gone chiplet there too...I don't really pay attention to the GPU world these days) You can't make 700 mm^2 dies off a brand new process, you want to give it time enough to be well tuned so your yields are decent, so those probably can't come until the following summer at best.
 

DrMrLordX

Lifer
Apr 27, 2000
21,583
10,785
136
There's no particular reason it should take AMD longer to get on a node than Apple does, other than that it is widely believed Apple pays TSMC up front for a guaranteed number of wafers off a new process which leaves few for others when it first ramps.

Apple comes first, ahead of everyone. Even Qualcomm, though not by much.
 

moinmoin

Diamond Member
Jun 1, 2017
4,934
7,620
136
I don't think anybody disputes it's in AMD's best interest to use the latest nodes. But I personally think Zen 2 set a precedence in regards to how AMD goes about it: AMD doesn't just push for some silicon getting the benefits of the new node as soon as possible (notable exceptions are the pipe cleaner GPU releases), instead Zen 2 showed an intricate adaption to the actual node, making the most of it right with the first design. This means the CPU designs on the respective node arrive later, when yield is already better and competition for the capacity is not as tight anymore. So AMD doesn't jump on bleeding edge nodes for mass production but for a phase of development that is significantly longer than that of other customers of TSMC.
 

uzzi38

Platinum Member
Oct 16, 2019
2,566
5,575
146
Apple comes first, ahead of everyone. Even Qualcomm, though not by much.
Not on N7P they didn't :^P

Navi10 came first there.

(Personally I think the node designed for AMD that's rumoured for N5 is actually just N5P, and we'll see AMD just release products on N5P directly as opposed to N5 first, at least on RTG's side anyway).
 

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
IIRC AMD has stated to not expect them to always be on the leading node, however. It would be counterproductive for AMD to move to a brand new, high cost node with limited benefits.

I personally believe the silicon landscape will change within the next 5 years. Node shrinks are creating their own issues. I wouldn’t be surprised to see AMD combine dies from 2 different processes, low power on 5nm and high performance on N7+ for example, on one chip.

Furthermore I expect newer, higher performance nodes to be introduced from 7nm-22nm depending on market demand. I expect the sweet spot for AMD will be 7nm.

The ultimate game is to balance cost, power consumption, and performance. Node shrinks don’t necessarily help with all three of those things.
 

piokos

Senior member
Nov 2, 2018
554
206
86
All those tables comparing the technical specs of each foundry processes with each other all miss one important detail, the resulting yield rate.
The resulting yield depends on the product you make.
Node is characterized by density of issues which just happen randomly. In this aspect Intel is probably on par with everyone else, since production of small mobile chips looks perfectly fine.
Monolithic designs make it harder to make large desktop/server chips.

Overall: such a waste this - potentially very interesting topic - became another AMD-vs-Intel CPU discussion. Like if there weren't enough threads where you could show your love for Zen2.
 

soresu

Platinum Member
Dec 19, 2014
2,617
1,812
136
The resulting yield depends on the product you make.
Node is characterized by density of issues which just happen randomly. In this aspect Intel is probably on par with everyone else, since production of small mobile chips looks perfectly fine.
Monolithic designs make it harder to make large desktop/server chips.

Overall: such a waste this - potentially very interesting topic - became another AMD-vs-Intel CPU discussion. Like if there weren't enough threads where you could show your love for Zen2.
It's also dependent on the maturity of the node too - brand new nodes will always be poor and a node 2+ years old should be fairly dependable if other customers have been satisfied with it.
 

jpiniero

Lifer
Oct 1, 2010
14,511
5,159
136
Node is characterized by density of issues which just happen randomly. In this aspect Intel is probably on par with everyone else, since production of small mobile chips looks perfectly fine.

Not really, Icelake is tiny volume. The economics of 10 nm become convoluted when there's a ton of money that has already been spent that needs to be depreciated.
 
Mar 11, 2004
23,031
5,495
146
IIRC AMD has stated to not expect them to always be on the leading node, however. It would be counterproductive for AMD to move to a brand new, high cost node with limited benefits.

I personally believe the silicon landscape will change within the next 5 years. Node shrinks are creating their own issues. I wouldn’t be surprised to see AMD combine dies from 2 different processes, low power on 5nm and high performance on N7+ for example, on one chip.

Furthermore I expect newer, higher performance nodes to be introduced from 7nm-22nm depending on market demand. I expect the sweet spot for AMD will be 7nm.

The ultimate game is to balance cost, power consumption, and performance. Node shrinks don’t necessarily help with all three of those things.

Have they said that? I think they've said its actually pretty important to their overall product portfolio and continuing to offer high performance chips, although I think they've talked about offering a full product stack and so not everything would be on the most advanced, but they've already been doing that. I think they've said that due to issues moving to new processes that they need to look at a variety of things, with the move to chiplets being one way of dealing with it, and then chip packaging (2.5/3D), hence why they're doing a lot of research there. They even talked about how much they had to rework the packaging for Zen 2.

That would be a possibility, especially in mobile, but I think they'd be better off mixing those in each die (so they use smaller more efficient cores to space out the larger hotter cores, maybe ditch SMT and instead put that towards going 2-3 small cores per 1 big core, so they still get lots of cores for multi-threaded stuff, but then get big cores for things that don't make as much use of it, while they space out the chips to alleviate hotspots). And then just doing what they've been doing of having the number of chiplets be determined by other factors. Although if maybe they look to mix x86 with ARM or something (possibly they buy someone else's ARM SoC to put on package, where on the go it'd just use the ARM stuff, but then plugged in it'd use the x86 Zen, with the ARM just handling networking especially say cellular

I would guess though that it'd be quite different. EPYC (and their HPC GPU/compute, which I think is going chiplet too) chips would start to be lead, and they'd be on the most advanced node, with the customers that can afford/need that performance can/would pay for it. Then consumer stuff comes later. I think AMD was actually trying to do that starting with Zen 2, but maybe with Zen 4 it really starts (maybe we see Zen 4 EPYC in 2021, but consumer Ryzen/Threadripper isn't until 2022). I kinda have a hunch that once they go GPU chiplet (even if its just for APU) that most of their consumer stuff will go that route. For the users that want more CPU, make them pony up for Threadripper.

I do think AMD has said they'll be utilizing 7nm for some time. But then TSMC seems to be changing what is 7nm and its all kinda blurring together, and I think some of that might just be more marketing games. But the IOD is still 12nm, so there's plenty of ways they can advance while still using multiple process nodes.
 

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
Have they said that? I think they've said its actually pretty important to their overall product portfolio and continuing to offer high performance chips, although I think they've talked about offering a full product stack and so not everything would be on the most advanced, but they've already been doing that. I think they've said that due to issues moving to new processes that they need to look at a variety of things, with the move to chiplets being one way of dealing with it, and then chip packaging (2.5/3D), hence why they're doing a lot of research there. They even talked about how much they had to rework the packaging for Zen 2.

That would be a possibility, especially in mobile, but I think they'd be better off mixing those in each die (so they use smaller more efficient cores to space out the larger hotter cores, maybe ditch SMT and instead put that towards going 2-3 small cores per 1 big core, so they still get lots of cores for multi-threaded stuff, but then get big cores for things that don't make as much use of it, while they space out the chips to alleviate hotspots). And then just doing what they've been doing of having the number of chiplets be determined by other factors. Although if maybe they look to mix x86 with ARM or something (possibly they buy someone else's ARM SoC to put on package, where on the go it'd just use the ARM stuff, but then plugged in it'd use the x86 Zen, with the ARM just handling networking especially say cellular

I would guess though that it'd be quite different. EPYC (and their HPC GPU/compute, which I think is going chiplet too) chips would start to be lead, and they'd be on the most advanced node, with the customers that can afford/need that performance can/would pay for it. Then consumer stuff comes later. I think AMD was actually trying to do that starting with Zen 2, but maybe with Zen 4 it really starts (maybe we see Zen 4 EPYC in 2021, but consumer Ryzen/Threadripper isn't until 2022). I kinda have a hunch that once they go GPU chiplet (even if its just for APU) that most of their consumer stuff will go that route. For the users that want more CPU, make them pony up for Threadripper.

I do think AMD has said they'll be utilizing 7nm for some time. But then TSMC seems to be changing what is 7nm and its all kinda blurring together, and I think some of that might just be more marketing games. But the IOD is still 12nm, so there's plenty of ways they can advance while still using multiple process nodes.

I could be mistaken, but I remember reading it either here on AnandTech or elsewhere. When I have time I will have to dig it up.

I will say that I have seen people criticizing TSMC for things like the naming of nodes and the like, but thus far they are the only ones delivering both high performance and low power chips.

Intel 10nm is getting there. I still stand by the belief that Intel isn’t actually having problems with 10nm, but rather, the Core architecture doesn’t port well so they lose a ton of performance. I am willing to be that a new architecture built specifically around 10nm would put the hurt on AMD and TSMC.
 

piokos

Senior member
Nov 2, 2018
554
206
86
Not really, Icelake is tiny volume. The economics of 10 nm become convoluted when there's a ton of money that has already been spent that needs to be depreciated.
You have some data to support this?
Because I haven't seen any data for actual node production, but looking at all the Ice Lake SoCs in popular laptop lineups, I wouldn't really risk the word "tiny".
In fact, looking just on laptops, I wouldn't even risk "less than AMD" at this point.
There are also 10nm products in the AI and networking divisions that we often forget about.

Sure, server 10nm server chips will arrive later this year and that's when we'll be able to say with full confidence that Intel is confident with both supply and quality of this node.
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,483
14,434
136
You have some data to support this?
Because I haven't seen any data for actual node production, but looking at all the Ice Lake SoCs in popular laptop lineups, I wouldn't really risk the word "tiny".
In fact, looking just on laptops, I wouldn't even risk "less than AMD" at this point.
There are also 10nm products in the AI and networking divisions that we often forget about.

Sure, server 10nm server chips will arrive later this year and that's when we'll be able to say with full confidence that Intel is confident with both supply and quality of this node.
Nothing I have seen backs this up. We will just see.. Milan may be out by then.
 

RetroZombie

Senior member
Nov 5, 2019
464
386
96
The resulting yield depends on the product you make.
Node is characterized by density of issues which just happen randomly. In this aspect Intel is probably on par with everyone else, since production of small mobile chips looks perfectly fine.
Monolithic designs make it harder to make large desktop/server chips.
That's a red flag right there.
If even small chips monolithic dual core* cpus are near impossible to manufacture by intel 10nm, what do you have to say about it? (the process, the yield, the product and how intel is probably on par with everyone else)

*That is already smaller than it should.
 

piokos

Senior member
Nov 2, 2018
554
206
86
That's a red flag right there.
If even small chips monolithic dual core* cpus are near impossible to manufacture by intel 10nm, what do you have to say about it?
Are you writing from 2018 or what? I don't understand this comment.
Take the new 13" Macbook Pro. Intel couldn't even deliver enough Icelake chips despite it being a year out, that Apple had to continue to use Coffee Lake for the base model.
Did Apple say Intel couldn't deliver enough chips? What's your source?

MacBook Pro 13 2020 is available with different chips - true. Just like the models that came before it.

If 10nm supply was the only issue, Apple could have opted for Comet Lake. I don't know why they went for 8th gen.

Latest MacBook Air uses Ice Lake exclusively.
 

piokos

Senior member
Nov 2, 2018
554
206
86
Nothing I have seen backs this up. We will just see.. Milan may be out by then.
I don't understand why Milan appears in this topic. You want to visit or what? Beautiful city but kind of locked right now.
Anyway, the discussion is supposed to be about semiconductor nodes, so let's leave tourism for another one. :)
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,483
14,434
136
I don't understand why Milan appears in this topic. You want to visit or what? Beautiful city but kind of locked right now.
Anyway, the discussion is supposed to be about semiconductor nodes, so let's leave tourism for another one. :)
Don't play stupid with me. You are trying to say Intel is coming out with new server chips this year. I am saying I doubt that, and that even if they do The new AMD server chips will most likely blow them out of the water.
 

DrMrLordX

Lifer
Apr 27, 2000
21,583
10,785
136
Not on N7P they didn't :^P

Navi10 came first there.

(Personally I think the node designed for AMD that's rumoured for N5 is actually just N5P, and we'll see AMD just release products on N5P directly as opposed to N5 first, at least on RTG's side anyway).

Apple leapfrogged N7P for N5 N7+, no? And they will be first to N5.

Intel 10nm is getting there. I still stand by the belief that Intel isn’t actually having problems with 10nm, but rather, the Core architecture doesn’t port well so they lose a ton of performance. I am willing to be that a new architecture built specifically around 10nm would put the hurt on AMD and TSMC.

I don't think I agree. Otherwise DG1 would have been less of a letdown, and Tremont wouldn't be so badly-delayed in its many incarnations.

Don't play stupid with me. You are trying to say Intel is coming out with new server chips this year. I am saying I doubt that, and that even if they do The new AMD server chips will most likely blow them out of the water.

They will have IceLake-SP, albeit after massive delays. I'm also a bit skeptical about the platform . . . we will see what happens.
 
Last edited:
  • Like
Reactions: spursindonesia

piokos

Senior member
Nov 2, 2018
554
206
86
Don't play stupid with me. You are trying to say Intel is coming out with new server chips this year. I am saying I doubt that, and that even if they do The new AMD server chips will most likely blow them out of the water.
I said Intel will offer Ice Lake SP, which can be seen as confirmation of 10nm being well controlled and suitable for large chips. That's it.

This has nothing to do with actual quality of the chip, because this topic is not about chips. And as such, it has even less to do with how these chips will stack against competition made by a company that outsources manufacturing.
2 years from now Intel may have the best semiconductor node and lag behind competition in actual chips (or the other way round). These are 2 separate matters.

Intel's CPU and GPU divisions compete with AMD and Nvidia on chip quality.
Intel's semiconductor division competes with TSMC and Samsung on node quality.
 
  • Like
Reactions: pcp7

jpiniero

Lifer
Oct 1, 2010
14,511
5,159
136
Did Apple say Intel couldn't deliver enough chips? What's your source?

Only other reason I can think of is that they were told Icelake was going to be EOLed very quickly and they needed something with a stable supply as the base model. Apple isn't like the other OEMS who have tons of different models and they are probably still mostly buying Whiskey Lake at this point.

Icelake Server is kind of the same deal. They may not basically cancel it like Cooper Lake but what OEMs will be buying is Cascade Lake/Refresh. Pretty obvious that Cooper Lake got cancelled because the single die version was redundant with Cascade Lake Refresh and the dual die version had embarrassing power consumption.

Ultimately Intel's future products hinge on what their chiplet strategy is and if they have the foresight to dual source if they have any problems with 7 nm.

If 10nm supply was the only issue, Apple could have opted for Comet Lake. I don't know why they went for 8th gen.

There's no Comet Lake with GT3. The alternative was Renior, which we know Apple looked at due to driver leaks but ultimately decided against it.
 

piokos

Senior member
Nov 2, 2018
554
206
86
Only other reason I can think of is that they were told Icelake was going to be EOLed very quickly and they needed something with a stable supply as the base model. Apple isn't like the other OEMS who have tons of different models and they are probably still mostly buying Whiskey Lake at this point.
If anything, it's Comet Lake (10th gen 14nm) that's going to be discontinued soon. Tiger Lake takes the key role and it'll be easier for Intel to support another 10nm arch.

Apple uses slightly customized SoCs. They already know that 8th gen SoC and gains from going Comet Lake are minimal.
Ultimately Intel's future products hinge on what their chiplet strategy is and if they have the foresight to dual source if they have any problems with 7 nm.
Assuming Xe will be made by a 3rd party manufacturer, it would make sense to outsource Xe IGPs as well.
So once Intel goes MCM, they would be able to increase their production capacity by 10-20%.
There's no Comet Lake with GT3. The alternative was Renior, which we know Apple looked at due to driver leaks but ultimately decided against it.
Iris-equipped 8th gen chips were made for Apple anyway. Apple could have ordered Comet Lake with GT3. But that would probably cost them more than getting the chip that was already developed.
 

RetroZombie

Senior member
Nov 5, 2019
464
386
96
Are you writing from 2018 or what? I don't understand this comment.
It's 2020 and here they are the Intel Core i3-1000, released in Q3-2019 and sold today.

It was you who mentioned this on your post to my post.
You want to talk about nodes but you cant do it without the other things i mentioned that you decided to ignore, like you want to ignore the inability of intel to manufacture one very small chip in high volume.
 

piokos

Senior member
Nov 2, 2018
554
206
86
It's 2020 and here they are the Intel Core i3-1000, released in Q3-2019 and sold today.
I still don't understand what you wanted to say here:
"If even small chips monolithic dual core* cpus are near impossible to manufacture by intel 10nm"

Ice Lake U lineup consists of 2 and 4-core SoCs.
And now, since OEMs started using Ice Lake U popular notebooks, it seems rather plausible that there are no problems with delivering them in large quantities.
You want to talk about nodes but you cant do it without the other things i mentioned that you decided to ignore, like you want to ignore the inability of intel to manufacture one very small chip in high volume.
Why would I not be able to compare nodes? Node is the product. TSMC doesn't even make in-house chips.