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.

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.

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


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

 

[maddie]

Reports were also that yields had improved. Anandtech reported that 4nm yields had surpassed 75% back in June.

This forum is like an echo chamber sometimes.

I was just pointing out to you that stating high utilization in isolation can be misleading.

 

[Doug S]

It was widely reported a few months ago in Korean tech media that Samsung Foundry had around 90% utilization of its 5nm capacity and 80% of its 4nm capacity.

When was the last time you saw a press release from a non-consumer-facing company about which foundry and node a given product was made on? That includes virtually every HPC/AI/networking/datacenter business.

Over the summer, there were numerous reports about significantly improved 4nm yields and large orders from (unnamed) HPC and AI customers.

Combined with Samsung's Q3 reports showing record Foundry revenue and solid future guidance, there appears little reason to doubt the validity of the reports.

It may well be highly utilized, but the fact that none of the sort of devices that will always utilize leading edge capacity like higher end smartphones, GPUs, whatever are using Samsung is telling. Who matters in the HPC and AI market beyond AMD, Intel, and Nvidia? There are a lot of smaller companies (like Nuvia was until Qualcomm bought them) pushing ARM CPUs for servers and HPC, but all of them collectively probably don't add up to 10% of an AMD or Intel. There are tons of startups especially in the AI market where the current hype is, so maybe they are all going to Samsung because they have better available, or offer better pricing due to having worse performance than TSMC. But again they would be a drop in the bucket compared to NVidia.

Utilization is also meaningless without knowledge of available capacity. If foundry A has 90% utilization of 20k wpm and foundry B has 70% utilization of 200k wpm, for instance. So the is the concept of 'large' orders. Obviously the Korean tech media is going to paint a rosy picture for Samsung, so without any numbers I take their claims with a grain of salt. If Samsung was competitive we would hear about one of the major players working with them. Or heck, Samsung phones using Exynos SoCs would show equal or better performance for TSMC fabbed SoCs using the same ARM designed cores.

 

[qmech]

I was just pointing out to you that stating high utilization in isolation can be misleading.

Yet we have, as I pointed out, more data points.

It may well be highly utilized, but the fact that none of the sort of devices that will always utilize leading edge capacity like higher end smartphones, GPUs, whatever are using Samsung is telling. Who matters in the HPC and AI market beyond AMD, Intel, and Nvidia? There are a lot of smaller companies (like Nuvia was until Qualcomm bought them) pushing ARM CPUs for servers and HPC, but all of them collectively probably don't add up to 10% of an AMD or Intel.

(Talk about moving goal posts... To a claim that Samsung Foundry had "hardly any customers" for 4nm, I pointed out the high utilization rates. That resulted in a "but low yields", to which I pointed out reports of high yields, and now we're at "but total capacity".)

Anyway.

I think you vastly overestimate the size of AMD here. Apple is huge, but there are more companies than you might think who are similar to AMD in terms of foundry spending.

Broadcom alone spends almost as much money with TSMC as AMD does. Are they single source? What about their competitors?

Amazon uses mainly TSMC for their millions of custom chips (Graviton3, Trainium, Inferentia), but Tenstorrent and Baidu have been reported as using Samsung Foundry for years. What about Facebook's? Google's TPU chips are, I believe, fabbed by TSMC via Google's ASIC partner (Broadcom).

IBM also has a wide array of custom silicon and their primary foundry partner is Samsung.

What we do know is that Samsung Foundry was reported to have landed "tier one" (whatever that means) clients in networking, HPC, and AI earlier this year. I don't know who any of these clients are, although I'm hearing one partiular company keep being mentioned in the networking space (although details seem to match, I still consider it to be water cooler gossip), and I don't think anyone else here does either. At least not anyone able to say .

All this isn't to say that Samsung Foundry is going to challange TSMC in anything anytime soon. I'm simply pointing out that the echo chamber comments here that Samsung isn't able to get any 4nm clients appears to be unsubstantiated and that reports would seem to indicate that their 4nm business is actually quite healthy at this stage, despite not having one of the major consumer-facing whales.

 

[Geddagod]

(Talk about moving goal posts... To a claim that Samsung Foundry had "hardly any customers" for 4nm, I pointed out the high utilization rates. That resulted in a "but low yields", to which I pointed out reports of high yields, and now we're at "but total capacity".)

Is it really that deep? Lol

Anyway.

Igh bro

I think you vastly overestimate the size of AMD here. Apple is huge, but there are more companies than you might think who are similar to AMD in terms of foundry spending.

Broadcom alone spends almost as much money with TSMC as AMD does. Are they single source? What about their competitors?

Amazon uses mainly TSMC for their millions of custom chips (Graviton3, Trainium, Inferentia), but Tenstorrent and Baidu have been reported as using Samsung Foundry for years. What about Facebook's? Google's TPU chips are, I believe, fabbed by TSMC via Google's ASIC partner (Broadcom).IBM also has a wide array of custom silicon and their primary foundry partner is Samsung.

Don't know much about Broadcom tbh. Do the use leading edge or next to leading edge nodes primarily? Idk. But the other examples you listed, (the other companies custom chips, and IBM) aren't really high volume compared to AMD, Nvidia, etc etc are they)

What we do know is that Samsung Foundry was reported to have landed "tier one" (whatever that means) clients in networking, HPC, and AI earlier this year.

I don't pay that much attention to Samsung. Rumor or did Samsung say so themselves?

I'm simply pointing out that the echo chamber comments here that Samsung isn't able to get any 4nm clients appears to be unsubstantiated and that reports would seem to indicate that their 4nm business is actually quite healthy at this stage, despite not having one of the major consumer-facing whales.

It is contradictory to say their 4nm business is healthy at this stage while simultaneously not having any major consumer facing whales IMO.

 

[Ajay]

I don't pay that much attention to Samsung. Rumor or did Samsung say so themselves?

It is contradictory to say their 4nm business is healthy at this stage while simultaneously not having any major consumer facing whales IMO.

They said that, but "Tier one" is something debatable. Apparently some Tesla chips and Tenstorrent chips are being manufactured by SS. The quantities aren't anything like one would see with ARM bases SoCs or AMD's portfolio. Apparently, their yields are up on 4LPP and 3GAE (with 3GAP supposedly coming next year)**. It's really hard to find anything but rumors on Samsung for some reason.

** https://www.anandtech.com/show/18960/samsung-foundry-s-3nm-and-4nm-yields-are-improving-report

 

[moinmoin]

My impression of Samsung Semiconductor is that while they are not at the leading edge they also don't occupy the space of older node gens that e.g. GloFo and Tower Semiconductor serve. So while Samsung Semiconductor apparently is not competitive against TSMC's current leading edge (and as such nobody seems willing to have press about it) they still offer the "next best thing" on the open market, very likely at a competitive cost (which is one reason Nvidia went all in previously before pivoting back to TSMC).

 

[qmech]

It is contradictory to say their 4nm business is healthy at this stage while simultaneously not having any major consumer facing whales IMO.

If they are operating at > 90% capacity for 5nm and > 80% for 4nm, does it matter if their customers are consumer-facing or not?

AI is seeing *insane* amounts of money being poured into it and it has been that way for a few years now. There are not enough NVIDIA chips to go around and they are extremely expensive, so everyone is rolling their own. The newest generations of those chips are all 4-5nm (very little aside from Apple is on 3nm right now, obviously). Those chips need serious networking chips (often with stacked HBM). There was a recent announcement by a design firm announcing a deal for an upcoming chip on Samsung 3nm (likely 3GAP), but the name of the (US) networking company was not mentioned. The current generation of those chips are also, like the AI chips, on 4 and 5nm. Not a single one of those companies publicizes who their foundry partner is.

 

[igor_kavinski]

Samsung lands orders for AMD's 4nm CPUs as chipmaker reportedly seeks to diversify production

AMD's Prometheus to be made both by TSMC and Samsung.

www.tomshardware.com

Basic Zen 5c on Samsung 4nm. If true, AMD must have big plans for this architecture to want to derisk its production.

 

[Markfw]

Samsung lands orders for AMD's 4nm CPUs as chipmaker reportedly seeks to diversify production

AMD's Prometheus to be made both by TSMC and Samsung.

www.tomshardware.com

Basic Zen 5c on Samsung 4nm. If true, AMD must have big plans for this architecture to want to derisk its production.

from all I have heard, there is a good chance that Zen 5 will totally dominate in every area. They already have server, HTPC, gaming, and for the most desktop, and laptop is right in the heat of the battle. If so, TSMC may not have enough production capability to supply all of their needs. Is that what you meant by de-risk ?

 

[igor_kavinski]

If so, TSMC may not have enough production capability to supply all of their needs. Is that what you meant by de-risk ?

That and pricing may be an issue too. When TSMC sees significant enough quantity of AMD wafers being produced by Samsung, they might offer better pricing options to AMD.

 

[Geddagod]

from all I have heard, there is a good chance that Zen 5 will totally dominate in every area. They already have server, HTPC, gaming, and for the most desktop, and laptop is right in the heat of the battle.

Cool story bro

When TSMC sees significant enough quantity of AMD wafers being produced by Samsung, they might offer better pricing options to AMD.

^^^
If this deal is for an actual CPU die rather than just IO stuff, then ye I can see some low end client Zen 5C parts come to fruition. Maybe as a Mendocino type successor for the future.

 

[FlameTail]

Samsung 4nm is inferior to TSMC 4nm. That is certain.

 

[H433x0n]

Samsung 4nm is inferior to TSMC 4nm. That is certain.

Then why does Samsung label it as 4nm? That's something that has always annoyed me with Samsung. On paper they've had "4nm" since 2021 and yet tt wasn't even in the same zip code as TSMC's N5 in that same time period.

 

[F-no]

Samsung 4nm is inferior to TSMC 4nm. That is certain.

I've not seen anyone do a comparison in a long time. And then there's TSMC's 4nm and 4P versus Samsung's 4 different versions of "4nm" at this point with the newest one having availability at the end of next year.

Samsung fired their previous foundry head and replaced him alongside investing a substantial amount into the division. It would be interesting to compare the two 5c variations when they come out.

 

[jpiniero]

Then why does Samsung label it as 4nm? That's something that has always annoyed me with Samsung. On paper they've had "4nm" since 2021 and yet tt wasn't even in the same zip code as TSMC's N5 in that same time period.

Cuz it's better than the node they named 5 nm.

 

[DrMrLordX]

Apparently, their yields are up on 4LPP and 3GAE (with 3GAP supposedly coming next year)**.

Is anyone actually ordering 3GAE wafers?

 

[Ajay]

Is anyone actually ordering 3GAE wafers?

Not that I could find - so, I guess just Samsung's internal numbers. I think SS's point is that 3GAP will be better than 3GAE, which is already yielding better.

Samsung fired their previous foundry head and replaced him alongside investing a substantial amount into the division. It would be interesting to compare the two 5c variations when they come out.

Hope it works out for their logic division. They need some good wins.

 

[Doug S]

Then why does Samsung label it as 4nm?

Marketing.

They want to preserve the illusion they are competing on the leading edge, rather than being a generation behind. Obviously that doesn't matter to the average person, since we aren't signing wafer purchase contracts, but Samsung wants to maintain a corporate image of being at the forefront of technology because it helps sell Galaxy phones, SSDs and other consumer products.

 

[qmech]

Not that I could find - so, I guess just Samsung's internal numbers. I think SS's point is that 3GAP will be better than 3GAE, which is already yielding better.

TechInsights had a report early this summer about a product on 3GAE in the wild, the MicroBT WhatsMiner M56S++. As the name implies, it is a cryptominer and these are often rather small and simple ASICs.

It is likely that Samsung 3GAE is quite limited in volume, as their Early node variants usually are.

 

[Ajay]

Well, saw an Update on Rapidus were they are working with TensTorrent and grabbed this from Wikipedia:

In January 2023, representatives of Rapidus and IBM participated in a meeting between Japanese Trade Minister Hagiuda's successor, Yasutoshi Nishimura, and Secretary Raimondo.[30]

In February 2023, Rapidus announced the selection of a site close to the New Chitose Airport in Chitose, Hokkaido for the location of its planned factory.[31][32] Rapidus also increased its partnership with imec by joining the latter's "Core Partner Program" in April 2023,[33][34] and in the same month received additional funding of 260 billion yen from the Japanese government.[35] An estimated $37 billion in total funding is needed before Rapidus can begin production.[36]

In July 2023, Rapidus and Tokyo Electron were discussed as companies that would "play a major role" in a new memorandum of understanding between Japan and India that is expected to involve $35.9 billion of investment by Japan in India by 2027[37] and was officially signed in New Delhi by ministers Yasutoshi Nishimura and Ashwini Vaishnaw for their countries respectively.[38][39] This agreement closely followed a similar deal between the United States and India which had been reached by President Joe Biden and Prime Minister Narendra Modi in June.[40][41]

Groundbreaking to begin construction on Rapidus' factory in Chitose was held in September 2023.[42] The company has reportedly been supported by around US$2.5 billion in government subsidies from the Kishida administration thus far.[43] Dutch equipment manufacturer ASML Holding announced they would also open an office in Chitose with 40-50 technical staff to assist Rapidus.[44] While full operation is not expected until 2027, a pilot line is expected to start in 2025.[45]

Apparently they are working with IBM and others to bring up a 2nm process in 2025 (pre-production I think).

 

[Mopetar]

That and pricing may be an issue too. When TSMC sees significant enough quantity of AMD wafers being produced by Samsung, they might offer better pricing options to AMD.

Only if someone else isn't right behind AMD to buy those wafers up. As long as TSMC has the best node available, expect them not to be overly concerned because plenty of companies want a piece of it.

 

[FlameTail]

Samsung Foundry to reportedly make 4nm processors for Tesla.

https://x.com/dnystedt/status/1727147844292038711?s=20

 

[controlflow]

A pretty interesting interview with Ann Kelleher.

 

[FlameTail]

https://x.com/Tech_Reve/status/1727887829391548702?s=20

Samsung to make LPDDR5T.

Seems like LPDDR5T-9600 seems to be the next stepping stone before LPDDR6.

Now all 3 major DRAM manufacturers are on it. (Micron calls theirs LPDDR5X-9600).

 

[lightisgood]

A pretty interesting interview with Ann Kelleher.

Intel 3 : 18% performance per watt gain (from Intel 4)
Intel 20A : 10% ppw gain (from Intel 3)
Intel 18A : 10% ppw gain (from Intel 20A)

Intel 18A should beat TSMC N2 on ppw and come earlier.