Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/are-samsung%E2%80%99s-3nm-gaa-volume-production-yields-closer-to-25-or-70.16269/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2020970
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

Are Samsung’s 3nm GAA Volume Production Yields Closer to 25% or 70%

Daniel Nenni

Admin
Staff member
Samsung has said they are building (3) 3nm fabs so they have high hopes. Based on the numbers I have seen Samsung 3nm is not competitive to TSMC N3 or Intel 3 and the design learning curve for GAA is still an issue so I do not expect the big semiconductor companies to bet their HVM products on Samsung 3nm.

In my opinion TSMC N2, Samsung 3nm, and Intel 20A will be yield learning nodes with few customers. Intel proper will be the only 20A customer and it will be CPU tiles only. TSMC N2 will be a like like 20nm and 10nm where only Apple did HVM and if you look at the latest TSMC financials there is no revenue attached to those nodes.
 

hist78

Well-known member
What are the odds that after years of pathetically low yields from Samsung's 5nm and 4nm chips, which are using DECADES OLD FinFET technology, that suddenly their soon to be released BRAND NEW Gate-All-Around (GAA) technology will have yields greater than 35%.

http://koreabizwire.com/samsung-expected-to-announce-mass-production-of-3nm-chip-next-week/221816

The bragging right is important to certain companies and Samsung is one of them. Does that change any competitive landscape? Probably not. Foundry customers spend big money and take so much risk to develop and make a product. They are sophisticated enough to tell what's real and what's important to them.
 

triceratops24

New member
They claimed 3nm GAA will be in production in 1H/22, so they'll make an announcement at the end of June just to say that they hit the milestone, regardless of what the yield is (which is probably low).
 

fansink

Member
Samsung has said they are building (3) 3nm fabs so they have high hopes. Based on the numbers I have seen Samsung 3nm is not competitive to TSMC N3 or Intel 3 and the design learning curve for GAA is still an issue so I do not expect the big semiconductor companies to bet their HVM products on Samsung 3nm.

In my opinion TSMC N2, Samsung 3nm, and Intel 20A will be yield learning nodes with few customers. Intel proper will be the only 20A customer and it will be CPU tiles only. TSMC N2 will be a like like 20nm and 10nm where only Apple did HVM and if you look at the latest TSMC financials there is no revenue attached to those nodes.

Eight weeks ago, sources were claiming Samsung’s 3nm GAA yields were 15%, including sources published here at SemiWiki.

It’s hard to imagine those yields have improved much since then, considering the additional complexity of their GAA construction.

https://wccftech.com/samsung-3nm-gaa-worse-yields-than-4nm/?beta=1
 

hist78

Well-known member
I'm wondering how much impacts or involvement on Samsung's 3nm from IBM?

It will be interesting if the reported Japan-US 2nm research and manufacturing project will use IBM's process technology.

IBM keeps conducting semiconductor manufacturing process research since it became a fabless company several years ago. As far as I know IBM is the only fabless company doing that.
 
Last edited:

Daniel Nenni

Admin
Staff member
They claimed 3nm GAA will be in production in 1H/22, so they'll make an announcement at the end of June just to say that they hit the milestone, regardless of what the yield is (which is probably low).

Samsung did the same at 10nm. The had to be first at all costs. When the announcement went live I talked to some friends at QCOM and they said 10nm yield was in single digits and they were getting good die versus wafers from Samsung to hide it.

It is a cultural thing, being first no matter what the costs. As I have said before, Samsung's biggest foundry market segment is the NOT TSMC business. Intel is now competing for that market so Samsung is in serious foundry trouble, my opinion.

I don't see either Samsung or Intel as a serious threat to TSMC with the FinFET business. GAA may be another story but if I had to bet on a winner it would be TSMC. The TSMC ecosystem and huge customer support is a force of nature, absolutely.
 

fansink

Member
Samsung did the same at 10nm. The had to be first at all costs. When the announcement went live I talked to some friends at QCOM and they said 10nm yield was in single digits and they were getting good die versus wafers from Samsung to hide it.

It is a cultural thing, being first no matter what the costs. As I have said before, Samsung's biggest foundry market segment is the NOT TSMC business. Intel is now competing for that market so Samsung is in serious foundry trouble, my opinion.

I don't see either Samsung or Intel as a serious threat to TSMC with the FinFET business. GAA may be another story but if I had to bet on a winner it would be TSMC. The TSMC ecosystem and huge customer support is a force of nature, absolutely.

After years of poor Exynos performance and user complaints, and according to news/rumors, Samsung may put their Exynos SoC on hold, or even cancel, and use outside SoC's from Qualcomm, MediaTek, or similar.

If the demise of Exynos is true, and if sufficient foundry customers can't be convinced to come back, what internal Samsung device might use their 3nm process?
 

Xebec

Active member
After years of poor Exynos performance and user complaints, and according to news/rumors, Samsung may put their Exynos SoC on hold, or even cancel, and use outside SoC's from Qualcomm, MediaTek, or similar.

If the demise of Exynos is true, and if sufficient foundry customers can't be convinced to come back, what internal Samsung device might use their 3nm process?
They could still court other makers of SoC's to use their fabs and then pull some kind of the very old IBM trick of being allowed to fab the exact same thing for themselves for a small licensing fee.

Question - when ARM creates a new ARM core or ARM IP and then goes to TSMC to fab test versions of it on the latest nodes.. do they also do this at Samsung or not?
 

Daniel Nenni

Admin
Staff member
They could still court other makers of SoC's to use their fabs and then pull some kind of the very old IBM trick of being allowed to fab the exact same thing for themselves for a small licensing fee.

Question - when ARM creates a new ARM core or ARM IP and then goes to TSMC to fab test versions of it on the latest nodes.. do they also do this at Samsung or not?

Yes, Samsung is an Arm customer and partner. Synopsys SRAM as well and some SerDes stuff All belong on leading edge process test chips for yield and performance check.
 
Question - when ARM creates a new ARM core or ARM IP and then goes to TSMC to fab test versions of it on the latest nodes.. do they also do this at Samsung or not?
Some clarifying points:
1) ARM develops physical IP design in addition to Soft IP (Physical is their std cell libraries and memory IP, which they license out as POP kits for their CPU / GPU soft IP; sometimes also GPIOs, but that is about it). ARM have made POP kits available on Samsung S7LPP and S5LPE from what I've seen announced at previous SFFs, but unclear about anything more recent.
2) ARM very occasionally will develop a POC or development vehicle, that includes physical implementations of their CPU or GPU IP; typically this is done only when there is a new ISA change and they want to seed development community with this device for early software development outside of their own internal migration.

Otherwise, ARM only develops their IP on FPGAs / emulation environments and licenses that soft IP to customers, who are the first to physically implement it, either with Foundry physical IP, or with POP kits. I can think of one instance in the past decade where ARM put an early CPU on N10 in support of a specific customer, but otherwise when you see claims from TSMC or Samsung Foundry about ARM CPU performance, it is the Foundry doing the implementation of that IP (Cortex A72, A53, etc), not ARM. They might use a Foundry PDK to validate that their IP meets specific PPA claims, but that doesn't mean they are going to drop that onto a shuttle.

You would see ARM cores used for PPA claims at events like SNUG / CDNLive / DAC, but again these are typically a customer (Mediatek), the EDA companies, or the Foundries themselves doing the presentation, and maybe ARM is there as a partner on the project. ARM has also done some interesting testchips to test concepts like 3D stacking / core folding (Trishul) but they are using proven CPU or fabric IP for that (keeps the bugs limited to their POC vs the IP), and again it is unclear to me if they took those projects from a converged design (what they share in publications) to actual silicon.
 
Top