Discussion Leading Edge Foundry Node advances (TSMC, Samsung Foundry, Intel) - [2020 - 2025]

[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.

AnandTech Forums: Technology, Hardware, Software, and Deals

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

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.

AnandTech Forums: Technology, Hardware, Software, and Deals

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

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.

---------------------------------------------------------------------------------------------------------------------------------------------------


FEEL FREE TO CREATE A NEW THREAD FOR 2025+ OUTLOOK, I WILL LINK IT HERE

 

[NTMBK]

It's wrong.

What is?

 

[OneEng2]

Why spin off fabs when 18A is so good and competition is losing so badly it has to attack that process? Use it for your own products to wipe the floor with competitors, think about spinning off later once everybody sees Intel is winning - more money that way.

Seems more likely that 18A run into serious problems that will become apparent to public only next year, and since that was the main project Pat pushed sacking him would make sense.

[Sarcasm] Intel has always been very transparent about new process issues and timing. I'm sure 18 A is no exception [/sarcasm]

I suspect that TSMC will fill all capacity they have for N2, but I agree they will not realize much revenue in 2025 (which is now officially "this year" so HAPPY NEW YEAR EVERY ONE!).

 

[Win2012R2]

Because their products are not in high enough demand to fill the fabs, and provide the funding to progress beyond 18A

Hang on, 18A products not shipping yet, how can you say they are not in high demand? Are you saying they won't be in high demand? And if they will be in high demand, why sell those valueable fabs and pay ~60% gross margin to TSMC?

A good performant hypothenical Xeon 7 that is priced well (meaning a lot cheaper than AMD - 64 high clocked cores for $3k max) will sell - as long as TCO based on perf/watt is in ballpark Turin, which it should be better if 18A is indeed N2 class stuff. Small e-cores in Xeon might also work - but only if they price them well, as in cheap.

Intel won't be able to sell just 18A fabs, and the rest of them ain't really sellable for at least few more years when they convert them to industry standard EDA tools.

 

[fkoehler]

Intel is sliding down to the basement bargain bin, and the rumors about 18A are all over the place. IF 18A were in fact some hidden comeback punch with relevence, Pat would not have been shown the door.
Hope I'm wrong, but Intel has 0 cred.

 

[DrMrLordX]

Hang on, 18A products not shipping yet, how can you say they are not in high demand?

He's probably referring to wafer lots, not the finished products themselves.

 

[Win2012R2]

He's probably referring to wafer lots, not the finished products themselves.

Because their products are not in high enough demand to fill the fabs, and provide the funding to progress beyond 18A. Their GPUs are not doing great, their AI business is a flop, Altera is not high volume, etc. Their next gen server chips should be competitive

Seems to me he is referring to current Intel's product line and next gen servers which should be on 18A

18A is the node that supports industry standard EDA tools, thus allowing 3rd parties to fab there. Given how expensive and booked out TSMC is there are only few possible reasons why we don't hear about 3rd parties using 18A -

1) yields are crap
2) too expensive - can be result of #1 or if it is an amazing node that is in high demand (I think we can cross the latter one out)
3) volume too small - Intel will obviously want priority for their own announced stuff, even though in theory IDM 2.0 should take highest bidder
4) secrecy - customers don't want to admit it, seems to be it would be very hard to hide at this point

Anything else?

I am inclined to believe 1,2 and 3 are true.

 

[DrMrLordX]

Seems to me he is referring to current Intel's product line and next gen servers which should be on 18A

Maybe, but as of now there are no end-user products, there's only 18a PDKs for people to look at and decide whether they want to produce anything on that process. And depending on whom you ask, there may not be enough people taking the bait to produce sufficient revenue to keep the fabs alive. That's where the demand side would come into the picture.

Of course you'd want to ask WHY potential customers (including a very-vocal Broadcom) are turning up their noses. We can only guess. Plus Broadcom may have been engaging in gamesmanship to knock Intel down on prices. Why else would they make anything like that public?

 

[Hitman928]

Seems to me he is referring to current Intel's product line and next gen servers which should be on 18A

18A is the node that supports industry standard EDA tools, thus allowing 3rd parties to fab there. Given how expensive and booked out TSMC is there are only few possible reasons why we don't hear about 3rd parties using 18A -

1) yields are crap
2) too expensive - can be result of #1 or if it is an amazing node that is in high demand (I think we can cross the latter one out)
3) volume too small - Intel will obviously want priority for their own announced stuff, even though in theory IDM 2.0 should take highest bidder
4) secrecy - customers don't want to admit it, seems to be it would be very hard to hide at this point

Anything else?

I am inclined to believe 1,2 and 3 are true.

Intel has said that they don’t expect meaningful revenue for 18a until 2027.

 

[Doug S]

Seems to me he is referring to current Intel's product line and next gen servers which should be on 18A

18A is the node that supports industry standard EDA tools, thus allowing 3rd parties to fab there. Given how expensive and booked out TSMC is there are only few possible reasons why we don't hear about 3rd parties using 18A -

1) yields are crap
2) too expensive - can be result of #1 or if it is an amazing node that is in high demand (I think we can cross the latter one out)
3) volume too small - Intel will obviously want priority for their own announced stuff, even though in theory IDM 2.0 should take highest bidder
4) secrecy - customers don't want to admit it, seems to be it would be very hard to hide at this point

Anything else?

Companies trust TSMC because they have a proven track record as a foundry, and don't yet trust Intel because they have zero track record as a foundry. That would be true even if Intel had never had any process problems and had been a year or two ahead of TSMC the whole time. There's a lot more to being a good foundry than the quality of your process, and Intel is just starting down that path. You might have the best pizza recipe in the world, but that doesn't mean you could operate a successful pizza place.

Yeah I know Intel sort of half pretended to be a foundry a decade ago and had a few token customers, but they basically forced those customers to do everything their way with proprietary tooling and Intel's products had priority in the fabs. I don't count that which is why I'm saying that today they have zero track record as a foundry.

I am inclined to believe 1,2 and 3 are true.

We know 3 is true, and that's all that is needed to hear much about third parties because they can't support much outside demand beyond their own in 2025. Doesn't mean 1 and/or 2 aren't also true but 3 alone is enough to cloud the issue as far as trying to determine success based on "who is reported to be using 18A".

When most of us here think of "third parties" we're thinking the big players like Apple, Nvidia and Qualcomm. Intel couldn't supply a fraction of their needs. If there was a rumor that Mediatek was having some SoCs made in Intel fabs would any of us here think "wow Intel must be succeeding to win an up and coming Android SoC vendor" or think "wow Intel is probably offering big discounts to win some token midrange products from a second tier Android SoC vendor".

Some customers might be a bit shy about wanting it known they're using Intel, because they are ALSO using TSMC for other products and don't want them finding out any sooner than they have to. It would be suicide to put all your eggs into the Intel basket now when 18A is unproven (though if you're a customer you should know whether it is in good shape or not by now) and Intel's ability to operate a foundry business is also unproven. Intel has to take that risk itself, it has no choice, but everyone else does.

 

[dttprofessor]

When APPLE on N3E，NV on N4P，TSMC will provide N3P 2025. Nobody will use 18A or N2 except INTEL who always used the top node in last 50 years.

 

[LightningZ71]

Wasn't there a strong rumor recently that Intel's Foundry design tools were still lagging behind as well. It doesn't matter much how good the physical node is if no one can actually develop a design for it...

 

[OneEng2]

One issue I keep hearing about is hot spots in BSPDN produced chips.

This would apparently limit frequency scaling due to TJMAX in transistors in those hot spots.

This provides reasoning to why Intel would first use 18A for Clearwater Forest (cloud computing DC processor) and Panther Lake (laptop processor).

It is likely that 18A will be quite power efficient, but my current gut feeling is that there will be clock scaling issues. This will become a problem for desktop and HPC.

I think the transition to TSMC's N2 (which does not include BSPD) will be slow. The large price increase for these wafers is tough to make a business case for over N3P IMO. TSMC also wisely (IMO) separated out BSPDN to A16 while making the libraries compatible with each other so customers would have options on which process to use.

Currently Intel lacks this feature with 18A.

 

[Josh128]

Seems rumored that the SOC for the Nintendo Switch 2 is using the Samsung 5nm node. Is this essentially the same node that the Exynos 2200 RDNA2 mobile phone CPU used? I remember it to be not very competitive after much hype. I think they marketed it as Samsung 4nm back then. Doesnt bode extremely well for Switch 2 performance wise against other consoles and handhelds, but pricing should be decent as well as availability. Considering Switch 1 used 2017 era 20nm/16nm process tech, even this Samsung 5 should allow for very nice gains in comparison to its predecessor at least.

 

[Meteor Late]

Seems rumored that the SOC for the Nintendo Switch 2 is using the Samsung 5nm node. Is this essentially the same node that the Exynos 2200 RDNA2 mobile phone CPU used? I remember it to be not very competitive after much hype. I think they marketed it as Samsung 4nm back then. Doesnt bode extremely well for Switch 2 performance wise against other consoles and handhelds, but pricing should be decent as well as availability. Considering Switch 1 used 2017 era 20nm/16nm process tech, even this Samsung 5 should allow for very nice gains in comparison to its predecessor at least.

But the thing is, Nvidia is more efficient than AMD and Intel ISO node, something people always forget.

 

[Josh128]

But the thing is, Nvidia is more efficient than AMD and Intel ISO node, something people always forget.

The difference is not nearly as large as it once was. 7900XTX is less efficient than 4080 Ada by around 16%. Only by 10% vs 4090. 16% is probably not going to completely make up for the node disavantage Samsung 5 has vs TSMC 4. We'll see.

AMD Radeon RX 7900 XTX Review - Disrupting the GeForce RTX 4080

Navi 31 is here! The new $999 Radeon RX 7900 XTX in this review is AMD's new flagship card based on the wonderful chiplet technology that made the Ryzen Effect possible. In our testing we can confirm that the new RX 7900 XTX is indeed faster than the GeForce RTX 4080, but only with RT disabled.

www.techpowerup.com

 

[jpiniero]

Seems rumored that the SOC for the Nintendo Switch 2 is using the Samsung 5nm node. Is this essentially the same node that the Exynos 2200 RDNA2 mobile phone CPU used? I remember it to be not very competitive after much hype. I think they marketed it as Samsung 4nm back then. Doesnt bode extremely well for Switch 2 performance wise against other consoles and handhelds, but pricing should be decent as well as availability. Considering Switch 1 used 2017 era 20nm/16nm process tech, even this Samsung 5 should allow for very nice gains in comparison to its predecessor at least.

See the other thread I bumped... it is almost certainly using Samsung 8 nm, the same that Ampere is on.

 

[Josh128]

See the other thread I bumped... it is almost certainly using Samsung 8 nm, the same that Ampere is on.

Yeah I saw David Huangs take on it this morning. Makes sense, I doubt Samsung5 could hit Nintendos price point.

 

[del42sa]

https://wccftech.com/nvidia-is-rumored-to-switch-towards-samsung-foundry-for-2nm-process/

NVIDIA is rumored to ditch TSMC's 2nm process, and instead turn towards Samsung Foundry, given the high costs associated and the Taiwan giant's growing influence in controlling prices.

https://wccftech.com/qualcomm-shifting-to-samsung-from-tsmc-for-2nm-chips/

Qualcomm reportedly has its eyes on Samsung Foundry too for all its production needs

however it´s kind of questionable regarding rumours about low yield of SF

 

[DrMrLordX]

Samsung, huh? They'd better be ready to set some wafers on fire unless they can get their yield problems under control.

 

[Win2012R2]

They'll both go with TSMC, only difference will be that some managers will get a hefty bonus for negotiating price down by 10%, and TSMC managers will also get bonus for inflating target price by 10%. Win-win, even Samsung won't complain because they'll get some PR in the meantime.

 

[LightningZ71]

Just keep in mind that not every product needs a leading edge node. With the cost of wafers for new nodes skyrocketing, the financial juggling act between the cost of spinning a chip architecture on multiple nodes vs. simplifying physical layer design and staying on one node takes on a different look. If you have a lower end product that is going to be produced in high numbers, say a xx60 series card, that can comfortably live on a mature 5nm class node, and notably lower volume parts in the xx70+ range that can live on two leading edge dies with various levels of die recovery and also sell at a notable market premium, you may decide to walk that road.

We also have no idea how badly Samsung wants to eat it's own face by entering into "known good die" contracts on it's more advanced nodes. Having machines sit idle is less desirable than having them working and producing some product that at least covers some of you fixed costs and all of your variable costs.

 

[Win2012R2]

Just keep in mind that not every product needs a leading edge node.

True, but we need leading edge GPUs and CPUs, must have!!!

 

[igor_kavinski]

TSMC's wafer pricing now $18,000 for a 3nm wafer

At advanced nodes, chips are not getting cheaper.

www.tomshardware.com

 

[Meteor Late]

True, but I need leading edge GPUs and CPUs, must have!!!

Fixed

 

[Curious_Inquirer]

Dylan Patel of SemiAnalysis on the rumors of Nvidia and Qualcomm moving some of their production to Samsung from TSMC.

https://twitter.com/x/status/1875961510251606515

https://twitter.com/x/status/1875961526244520029

https://x.com/dylan522p/status/1875961528861827212?s=61&t=6s7eyGwxMGt7pKMzBnBjmw
https://x.com/dylan522p/status/1875961531499979074?s=61&t=6s7eyGwxMGt7pKMzBnBjmw

Makes sense that they wouldn’t move any of their most important designs over to Samsung if they(Samsung) can’t get their process nodes up to high yields. A lot of money is riding on these designs to work and work well it makes a lot of business sense to rely on a tried, tested and most important of all Trusted partner in TSMC.

Edit: can’t add more than two tweets.