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

[DisEnchantment]

TSMC's N7 EUV is now in its second year of production and N5 is contributing to revenue for TSMC this quarter. N3 is scheduled for 2022 and I believe they have a good chance to reach that target.


N7 performance is more or less understood.


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

TSMC Radically Boosts CapEx to Expand Production Capacities, To Reach $14B For 2019

www.anandtech.com

For Samsung the nodes are basically the same from 7LPP to 4 LPE, they just add incremental scaling boosters while the bulk of the tech is the same.

Samsung is already shipping 7LPP and will ship 6LPP in H2. Hopefully they fix any issues if at all.
They have two more intermediate nodes in between before going to 3GAE, most likely 5LPE will ship next year but for 4LPE it will probably be back to back with 3GAA since 3GAA is a parallel development with 7LPP enhancements.

Samsung’s Aggressive EUV Plans: 6nm Production in H2, 5nm & 4nm On Track

www.anandtech.com

Samsung's 3GAA will go for HVM in 2022 most likely, similar timeframe to TSMC's N3.
There are major differences in how the transistor will be fabricated due to the GAA but density for sure Samsung will be behind N3.
But there might be advantages for Samsung with regards to power and performance, so it may be better suited for some applications.
But for now we don't know how much of this is true and we can only rely on the marketing material.

This year there should be a lot more available wafers due to lack of demand from Smartphone vendors and increased capacity from TSMC and Samsung.
Lots of SoCs which dont need to be top end will be fabbed with N7 or 7LPP/6LPP instead of N5, so there will be lots of wafers around.

Most of the current 7nm designs are far from the advertized density from TSMC and Samsung. There is still potential for density increase compared to currently shipping products.
N5 is going to be the leading foundry node for the next couple of years.

For a lot of fabless companies out there, the processes and capacity available are quite good.

 

[A///]

Raptor Lake looks like it beats Zen 3 (desktop) quite easily to me.

AMD had to bump up power limits for Zen 4 (which is on 5nm) just to compete, even though Intel was technically at a disadvantage. Intel still wins many benchmarks in terms of absolute performance, and does quite well in terms of efficiency if you cap the power limits.

Remember, AMD’s first N7 product was Zen 2.

I was discussing Zen 4. Not Zen 3. Zen 4's power can be cut with minimal to no loss in performance. Both Intel and AMD bumped their power specs this generation "just to compete". It's very telling when there's little loss in performance while cutting back significant power. In other worse they both pushed the power to gain scant performance. Ignore the power and Intel still needs more cores and more power. I'm not sure why this is difficult for some people to understand.

The 13900K locked at 253 watts still needs its 8+16 setup to match or just edge out the 7950X. It's a win for Intel, but it's a pyrrhic victory. Needing 24 cores against 16 while clocking higher just to match to beat AMD by a very small margin is not a good win.

 

[igor_kavinski]

Needing 24 cores against 16 while clocking higher just to match to beat AMD by a very small margin is not a good win.

I've said before in some thread that the best thing Intel can do right now is to license the AMD tech and start churning out their version of Ryzen in their fabs. The upside will be, it will be way more power efficient than anything Intel can come up with in the near term (still waiting to see if they have conjured up a miracle with MTL). The world will be a better place with only Intel Ryzen and AMD Ryzen. While Intel is eating humble pie, it can go back to the drawing board and try to design a brand new power efficient core. The world of today is so full of this useless pride that makes people and corporations do absolutely stupid stuff, just to keep up appearances of being strong and self reliant.

 

[Dayman1225]

I've said before in some thread that the best thing Intel can do right now is to license the AMD tech and start churning out their version of Ryzen in their fabs. The upside will be, it will be way more power efficient than anything Intel can come up with in the near term (still waiting to see if they have conjured up a miracle with MTL). The world will be a better place with only Intel Ryzen and AMD Ryzen. While Intel is eating humble pie, it can go back to the drawing board and try to design a brand new power efficient core. The world of today is so full of this useless pride that makes people and corporations do absolutely stupid stuff, just to keep up appearances of being strong and self reliant.

That is probably the craziest thing I’ve ever heard. Think just about every Intel investor would sell if they announced such a deal and I imagine a hell of a lot of people would leave the company for lack of confidence.

 

[controlflow]

I'm aware of the stupid naming conventions. While TSMC 7 and Intel "7" (10nm++ SF) are similar in density and pitch, one still out performs the other. Intel's node gives them the ability to clock higher at the cost of heat and power. Are we really ignoring that Intel needs that plus 24 total cores with 8 of them having HT to compete or uncomfortably beat a 7950X while using more power?

Intel can notch down a few nodes in the next few years but their design choices leave a lot to be desired. Needing to pump out higher frequency and more cores to match 16 SMT cores is sad.

I think Raptor Lake does a reasonable job at competing with Zen 4 when you consider that the 7950x is on TSMC N5. Yes Raptor Lake is less power efficient then Zen 4 and it is pushed to pretty high power consumption to match or beat Zen 4 which has a node advantage.

Raptor Lake seems to match or beat TSMC N7 products like Zen 2 and 3 though. I think it is fair to criticize Intel for their design choices with Golden Cove being too large or even the big core small core strategy with the incompatible ISAs, but I'm not understanding the criticism of Intel 7 as being some kind of misleading name when we are willing to accept TSMC N7. I think Intel 7 is generally as advertised.

 

[igor_kavinski]

That is probably the craziest thing I’ve ever heard. Think just about every Intel investor would sell if they announced such a deal and I imagine a hell of a lot of people would leave the company for lack of confidence.

The new investors will get rich in time and the people leaving will hopefully be the ones who have put Intel in its current condition by promoting their crappy ideas over sensible ones that got vetoed coz those engineers didn't have enough political power within the company. I don't care what anyone says but I firmly believe that Jim Keller left due to a toxic culture.

 

[A///]

I think Raptor Lake does a reasonable job at competing with Zen 4 when you consider that the 7950x is on TSMC N5. Yes Raptor Lake is less power efficient then Zen 4 and it is pushed to pretty high power consumption to match or beat Zen 4 which has a node advantage.

Raptor Lake seems to match or beat TSMC N7 products like Zen 2 and 3 though. I think it is fair to criticize Intel for their design choices with Golden Cove being too large or even the big core small core strategy with the incompatible ISAs, but I'm not understanding the criticism of Intel 7 as being some kind of misleading name when we are willing to accept TSMC N7. I think Intel 7 is generally as advertised.

Don't forget the 8 extra cores!

 

[A///]

I've said before in some thread that the best thing Intel can do right now is to license the AMD tech and start churning out their version of Ryzen in their fabs. The upside will be, it will be way more power efficient than anything Intel can come up with in the near term (still waiting to see if they have conjured up a miracle with MTL). The world will be a better place with only Intel Ryzen and AMD Ryzen. While Intel is eating humble pie, it can go back to the drawing board and try to design a brand new power efficient core. The world of today is so full of this useless pride that makes people and corporations do absolutely stupid stuff, just to keep up appearances of being strong and self reliant.

That is probably the craziest thing I’ve ever heard. Think just about every Intel investor would sell if they announced such a deal and I imagine a hell of a lot of people would leave the company for lack of confidence.

Igor that is the nuttiest suggestion you've ever made. Go home you're drunk on bananas and banana liqueur.

 

[igor_kavinski]

Igor that is the nuttiest suggestion you've ever made. Go home you're drunk on bananas and banana liqueur.

Why wouldn't the world be a better place with AMD as the ARM of the x86 market?

 

[controlflow]

Don't forget the 8 extra cores!

The 8 extra cores that take the same die area as 2 P cores? Intel could have gone with a 12 P core raptor lake sku with the same die area but instead they went 8+16. Seems that was a wise choice given their node disadvantage. Going with a second PPA optimized architecture turned out to be a pretty decent trade off to boost MT performance within the same die space budget.

I don't see how any of this relates to your argument that Intel 7 is somehow a misleading name and should be singled out relative to TSMC N7.
Raptor Lake on Intel 7 is able to compete with or beat AMD Zen2 and Zen3 on TSMC N7 but can't match TSMC N5 based Zen 4 in efficiency. It seems to be inline with what is expected...

 

[H433x0n]

I've said before in some thread that the best thing Intel can do right now is to license the AMD tech and start churning out their version of Ryzen in their fabs. The upside will be, it will be way more power efficient than anything Intel can come up with in the near term (still waiting to see if they have conjured up a miracle with MTL). The world will be a better place with only Intel Ryzen and AMD Ryzen. While Intel is eating humble pie, it can go back to the drawing board and try to design a brand new power efficient core. The world of today is so full of this useless pride that makes people and corporations do absolutely stupid stuff, just to keep up appearances of being strong and self reliant.

If this isn’t trolling I’m convinced this forum is developing brain worms. There’s this perception here that AMD is like 10 years ahead.

If you measure raw IPC at iso clocks, RPC has an IPC advantage (wins fp, ties int). From a raw performance standpoint, Zen 4 doesn’t tally a single win in client. Both products use approximately 250mm^2 of silicon but Zen 4 has significantly more transistors to work with. Despite this RPL has a 10% ST advantage and ties in MT performance. Yeah, it's less efficient but it's also on an older node and has less transistors to work with - so yeah?

As far as the core sizes go - first it’s a borderline dumb argument since GLC was developed for a big.Little product with a larger size budget due to gracemont. Intel also owns their own fabs, their primary motivator isn't to save silicon space since they have a comparative advantage in this aspect. Ignoring that Intel 7 UHP has half (~64Mtr) of the density that Zen 4 has (~120Mtr). Zen 4 is using a custom flavor N5P with liberal use of HD cells and much better sram scaling. Despite this, everybody is spiking the football over a smaller core with worse performance on a much more dense node?

 

[A///]

The 8 extra cores that take the same die area as 2 P cores? Intel could have gone with a 12 P core raptor lake sku with the same die area but instead they went 8+16. Seems that was a wise choice given their node disadvantage. Going with a second PPA optimized architecture turned out to be a pretty decent trade off to boost MT performance within the same die space budget.

I don't see how any of this relates to your argument that Intel 7 is somehow a misleading name and should be singled out relative to TSMC N7.
Raptor Lake on Intel 7 is able to compete with or beat AMD Zen2 and Zen3 on TSMC N7 but can't match TSMC N5 based Zen 4 in efficiency. It seems to be inline with what is expected...

We're not discussing space here are we? A core is still a core even if it lacks ht. Either cut them out of the equation for performance tests and meet parity with Zen 4 and fail, or continue to use the extra cores to stay above water. Intel needs 24 cores with 32 threads to just dribble past AMD's 16 core 32 thread processor. At a much higher power. That's a pyrrhic win.

You can have the best node available to but if your design sucks, it sucks. Let me say this in plain English. In 2-3 years when Intel will be on 18A, provided they get there, will still be matching or just edging out AMD if they're lucky. That's assuming they can manufacture everything themselves without relying on TSMC in addition to their own manufacturing abilities, present or not.

 

[lightisgood]

If this isn’t trolling I’m convinced this forum is developing brain worms. There’s this perception here that AMD is like 10 years ahead.

If you measure raw IPC at iso clocks, RPC has an IPC advantage (wins fp, ties int). From a raw performance standpoint, Zen 4 doesn’t tally a single win in client. Both products use approximately 250mm^2 of silicon but Zen 4 has significantly more transistors to work with. Despite this RPL has a 10% ST advantage and ties in MT performance. Yeah, it's less efficient but it's also on an older node and has less transistors to work with - so yeah?

As far as the core sizes go - first it’s a borderline dumb argument since GLC was developed for a big.Little product with a larger size budget due to gracemont. Intel also owns their own fabs, their primary motivator isn't to save silicon space since they have a comparative advantage in this aspect. Ignoring that Intel 7 UHP has half (~64Mtr) of the density that Zen 4 has (~120Mtr). Zen 4 is using a custom flavor N5P with liberal use of HD cells and much better sram scaling. Despite this, everybody is spiking the football over a smaller core with worse performance on a much more dense node?

Agree.
Intel 7 Ultra, world first 6GHz process, remind me and us that density is not all.
Yes, this is the bitter lesson of Intel 10nm, too.

I think that Intel is going right path now.
Especially, the additional process of Intel 3 and GNR/SRF are great plans because TSMC N3 is ruined.

 

[A///]

I think that Intel is going right path now.
Especially, the additional process of Intel 3 and GNR/SRF are great plans because TSMC N3 is ruined.

They're great but only if Intel can manufacture enough of them without problems. And TSMC N3 is ruined how? Intel's packaging is about to get very complex over the next few years and validation is going to eat up a lot of their manufacturing time. Time will tell if their choices are up to snuff or not.

 

[lightisgood]

They're great but only if Intel can manufacture enough of them without problems. And TSMC N3 is ruined how? Intel's packaging is about to get very complex over the next few years and validation is going to eat up a lot of their manufacturing time. Time will tell if their choices are up to snuff or not.

TSMC made too confident decision about EUV SADP and N3.
In 2H21~1H22, we could be aware that there was no hope of success.

 

[SiliconFly]

My point here was even if both companies had node access parity, Intel would still lose due to poor design. If AMD manufactured a Zen product on Intel's IFS in the future and Intel used that same node, AMD would still be ahead of them. Intel is an ego filled company that's experiencing brain drain and keeps hacking off bits and pieces.

That part is true. Thats because of only one reason. The super-fat & inefficient P-core. It's an outdated architecture that shouldn't even exist anymore. Intel is well aware of it and they're actively trying to ditch it asap.

Remember there is some confusion with ARL's P-core being RWC+ or LNL? Rumors said that they have (or had) both the projects going (dunno for how long) until one showed more promise. Not sure what that exactly means cos Intel's still keeping us in the dark. But if things work out, Intel may ditch the old P-cores in favor of LNL sooner than expected.

And there's a small chance Intel may use TSMC N3 for it's client tiles as well which can give them some more advantage.

 

[Thibsie]

The 8 extra cores that take the same die area as 2 P cores? Intel could have gone with a 12 P core raptor lake sku with the same die area but instead they went 8+16. Seems that was a wise choice given their node disadvantage. Going with a second PPA optimized architecture turned out to be a pretty decent trade off to boost MT performance within the same die space budget.

I don't see how any of this relates to your argument that Intel 7 is somehow a misleading name and should be singled out relative to TSMC N7.
Raptor Lake on Intel 7 is able to compete with or beat AMD Zen2 and Zen3 on TSMC N7 but can't match TSMC N5 based Zen 4 in efficiency. It seems to be inline with what is expected...

Yep but this is at first a die size economy thing.
If they could have design P core a little less fat, they could have done 16 P cores as AMD.

 

[Thibsie]

Igor that is the nuttiest suggestion you've ever made. Go home you're drunk on bananas and banana liqueur.

Yet, that would be quite enjoyable.
Once upon a time... AMD made Intel licensed designed CPUs as second source for the industry.
Intel licensing AMD design would habe been fun xD

Never gonna happen though...

 

[A///]

TSMC made too confident decision about EUV SADP and N3.
In 2H21~1H22, we could be aware that there was no hope of success.

The past is the past, we're in the now.

That part is true. Thats because of only one reason. The super-fat & inefficient P-core. It's an outdated architecture that shouldn't even exist anymore. Intel is well aware of it and they're actively trying to ditch it asap.

Remember there is some confusion with ARL's P-core being RWC+ or LNL? Rumors said that they have (or had) both the projects going (dunno for how long) until one showed more promise. Not sure what that exactly means cos Intel's still keeping us in the dark. But if things work out, Intel may ditch the old P-cores in favor of LNL sooner than expected.

And there's a small chance Intel may use TSMC N3 for it's client tiles as well which can give them some more advantage.

The sheer and utter confusion of what's coming down the pipeline unaided by dimwits like mlid doesn't help.

Yep but this is at first a die size economy thing.
If they could have design P core a little less fat, they could have done 16 P cores as AMD.

would have could have. I'm waiting to see how awful their yields are on bigger tiles.

Yet, that would be quite enjoyable.
Once upon a time... AMD made Intel licensed designed CPUs as second source for the industry.
Intel licensing AMD design would habe been fun xD

Never gonna happen though...

AMD would have to agree to such a preposterous deal in the first place. intel's problem is their ego and apple like advertising.

 

[H433x0n]

That part is true. Thats because of only one reason. The super-fat & inefficient P-core. It's an outdated architecture that shouldn't even exist anymore. Intel is well aware of it and they're actively trying to ditch it asap.

Remember there is some confusion with ARL's P-core being RWC+ or LNL? Rumors said that they have (or had) both the projects going (dunno for how long) until one showed more promise. Not sure what that exactly means cos Intel's still keeping us in the dark. But if things work out, Intel may ditch the old P-cores in favor of LNL sooner than expected.

And there's a small chance Intel may use TSMC N3 for it's client tiles as well which can give them some more advantage.

It’s not. RWC is a slightly larger core than RPC (iso node) and slightly more performant and yet RWC is 26% smaller despite packing in more transistors. The node makes a huge difference and this is seemingly lost on everybody.

MTL compute tile is 73mm^2, the same size as a Zen 4 CCD and it outperforms it. Did Intel suddenly become masters of silicon space efficiency?

 

[DrMrLordX]

Intel 7 Ultra, world first 6GHz process, remind me and us that density is not all.

Pfft Intel had 6 GHz+ chips years ago. Unless for some reason you're discounting overclockers, which renders the distinction moot if you do.

Yep but this is at first a die size economy thing.
If they could have design P core a little less fat, they could have done 16 P cores as AMD.

Careful, I've suggested that before but the e-core fans won't have any of it.

 

[lightisgood]

The past is the past, we're in the now.

Why didn't you see iPhone 15 Pro's small amount performance boost ?
This is why major customers (Qualcomm, MediaTek, NVIDIA, etc) avoid N3 and focus on N3E.
Also, Intel never rely on N3 process, poor-yield and high-wafer-cost... , in this year.

Generally, process development demands long lead time.
So, we may easily predict that it is good or not, especially foundry's process from major customers action.

For example, when Qualcomm left Intel 10nm in 2017-18, we cloud know it is rotten process.
Yes, after this, we got Ice Lake in 2019 with poor performance.

 

[A///]

Why didn't you see iPhone 15 Pro's small amount performance boost ?
This is why major customers (Qualcomm, MediaTek, NVIDIA, etc) avoid N3 and focus on N3E.
Also, Intel never rely on N3 process, poor-yield and high-wafer-cost... , in this year.

Generally, process development demands long lead time.
So, we may easily predict that it is good or not, especially foundry's process from major customers action.

And the poor bump in performance in prior iphones was because all those leading edge nodes sucked too? Did it ever occur to you Apple can't grab 40% IPC improvement or some other stupid astronomical figure you lot expect them to do with each generation?

 

[A///]

The node makes a huge difference and this is seemingly lost on everybody.

Litho is one part of the equation, design being another. You could have the best litho with the worst design.

 

[Thibsie]

Careful, I've suggested that before but the e-core fans won't have any of it.

Intel e-cores certainly need a lot of fans to move that heat xD

 

[SiliconFly]

It’s not. RWC is a slightly larger core than RPC (iso node) and slightly more performant and yet RWC is 26% smaller despite packing in more transistors. The node makes a huge difference and this is seemingly lost on everybody.

MTL compute tile is 73mm^2, the same size as a Zen 4 CCD and it outperforms it. Did Intel suddenly become masters of silicon space efficiency?

I think we're actually discussing about the size of a single p-core. It is larger than it needs to be (for example, compared to the size of a single Zen 4 core).