Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/tsmcs-top-customers-2019-2021.13925/
        )

    [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] => 2021370
            [XFI] => 1050270
        )

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

TSMC's Top Customers 2019-2021

Daniel Nenni

Admin
Staff member
Not a big surprise but it will be interesting to see how much of a bump Intel will see at 3nm. I'm also guessing Mediatek will jump up a bit as well.


TSMC Top Customers 2019-2021.jpg
 
it look like others fill in for the lost from Hi Silicon (Huawei). the Qualcomm and Nvidia shrink. i'm guessing that's when they turn to Samsung.
 
it look like others fill in for the lost from Hi Silicon (Huawei). the Qualcomm and Nvidia shrink. i'm guessing that's when they turn to Samsung.

Yes, Nvidia and QCOM use both TSMC and Samsung. Samsung for better pricing if they yield. TSMC if they don't yield. You should see a big shift at 3nm to TSMC since Samsung 3nm is not doing very well with yield or density. Mediatek is growing because they have a lock on the China SoC business instead of HSilicon. TSMC CAPEX may rise again!
 
View attachment 404
This is tsmc Fab 18 construction site image on Nov.30 2020 from DACIN Construction Co., Ltd. It explains the customer requirement and CapEx spending in 2021.

TSMC 3nm is going to be another big node. In fact, it will be the biggest node in the history of TSMC, my opinion. Apple, Intel and AMD will dominate it. QCOM and Nvidia will be back at TSMC for 3nm (Samsung 3nm called in sick).
 
TSMC 3nm is going to be another big node. In fact, it will be the biggest node in the history of TSMC, my opinion. Apple, Intel and AMD will dominate it. QCOM and Nvidia will be back at TSMC for 3nm (Samsung 3nm called in sick).
The big node typically means more than 100kwpm and becomes >20% revenue in a year. To achieve that, CapEx would be >$20B for the whole node building up. It seems you are right but it definitely is not easy to be successful in tech development, ramp and HVM. For 3nm 100kwpm, tsmc might need >50 sets of EUV but ASML can produce only 40 sets per year(2020 31set, 4 tools short of planned, 2021 would be ~36 sets only). Let's see how tsmc, Samsung and intel play to secure EUV tools in 2021 and 2022.
 
The big node typically means more than 100kwpm and becomes >20% revenue in a year. To achieve that, CapEx would be >$20B for the whole node building up. It seems you are right but it definitely is not easy to be successful in tech development, ramp and HVM. For 3nm 100kwpm, tsmc might need >50 sets of EUV but ASML can produce only 40 sets per year(2020 31set, 4 tools short of planned, 2021 would be ~36 sets only). Let's see how tsmc, Samsung and intel play to secure EUV tools in 2021 and 2022.

Do you know how many EUV tools Intel, TSMC, and Samsung have today? I believe TSMC has 50 with more on order for 3nm. I also heard that Intel and ASML are in disagreement about EUV tool shipments.

TSMC EUV Tools 2020.jpg
 
1616295399785.png

This was report from Japanese media last year but it is wrong. Now tsmc, Samsung, intel and SK Hynix order some tools from total ~36 tools in 2021. I believe it could be very challenging for tsmc to get 50 new EUV tools by end of 2021.
 
From assumed 14 EUV layers for TSMC 5nm, with Fab 18 3 phases, we might estimate 42 tools (3 x 14). That is about half the installed base, in fact. But we know some 5nm TSMC layers are not single patterning, so there could be more, unless the additional patterning is DUV.
 
Here are the TSMC EUV nodes and supposed EUV layers and capacity for 2021: 6&7+nm/1/70kwpm, 5nm/12/100kwpm, and 3nm/20/20kwpm. Estimation of total EUV machines required in TSMC this year would be interesting.
 
Last edited:
The situation in manufacturing is very screwed at the moment. If TSMC was not practically a monopoly for the high-end nodes, the distribution of revenue in 2021 would be very worrisome. It demonstrates that TSMC is too dependent on Apple. Now since, TSMC has to accomodate all customers asking to manufacture at the bleeding edge, it is in a very comfortable position and can use the dominant position of Apple to play with others who are trying to secure supply. Removing HiSilicon from the equation has freed quite significant capacity for other players. Comparing the 2020 and 2021 columns of the table, the top 10 in 2020 accounted for 81,3% of revenue, but this drops to 71,5% in 2021, which shows that smaller players seem to picking up capacity.

I understand that 2021 numbers are projections, but I would expect Apple and AMD to score even higher percentages. ARM is expected to introduce more M1-like chips for a broader range of laptops/desktops and AMD seems not to be able to offer enough CPUs/GPUs. I do not understand how these numbers have not risen more.
 
Amd at 7nm will become tsmc largest customer.


Agreed. Samsung has a good 7nm so QCOM and Nvidia stayed there. TSMC won the 5nm node and will win 3nm, my opinion. From what I hear now Intel will not outsource as much to TSMC as previously expected so AMD may the #2 TSMC 3nm customer after the dust settles.
 
Agreed. Samsung has a good 7nm so QCOM and Nvidia stayed there. TSMC won the 5nm node and will win 3nm, my opinion. From what I hear now Intel will not outsource as much to TSMC as previously expected so AMD may the #2 TSMC 3nm customer after the dust settles.
I had read Nvidia was still on 8nm, maybe the large die was a bigger risk with EUV.

Qualcomm is already trying to move beyond 7nm, but it was suspected 5nm (888)/4nm (X65/X62 modems) is with Samsung.

4nm is a surprise, I don't think Samsung really wants to go through with it for long; it does not have enough tooling for it.
 
Last edited:
Given the news of the lack of Snapdragon 888 processors and rumors of Samsung cancelling their 2021 Note smartphones, I do wonder if Samsung's 5nm has good yields.

Ambarella last earnings call seems to suggest that Samsung is able to currently meet their foundry needs at least.

One thing I noticed is that both Qualcomm and Ambarella mentions that they use "the most advanced 5nm process technology" in the Qualcomm press release/Ambarella earnings call yet based on the Scotten Jones reports (and foundry estimates) we got, it would appear that not-yet-released Samsung's 3nm may not beat TSMC's current 5nm in terms of performance, density or cost. Is Samsung paying their customers to make these statements?!
 
One thing I noticed is that both Qualcomm and Ambarella mentions that they use "the most advanced 5nm process technology" in the Qualcomm press release/Ambarella earnings call yet based on the Scotten Jones reports (and foundry estimates) we got, it would appear that not-yet-released Samsung's 3nm may not beat TSMC's current 5nm in terms of performance, density or cost. Is Samsung paying their customers to make these statements?!

Yes.

And I don't differentiate between Samsung 8nm and 7nm, it is the same fab, same process with a little tuning. As is Samsung 5nm and 4nm. They used to put + behind the number when they optimized a process. Now they change it completely because most people don't know any better.
 
Given the news of the lack of Snapdragon 888 processors and rumors of Samsung cancelling their 2021 Note smartphones, I do wonder if Samsung's 5nm has good yields.

That was indeed very interesting, I also wondered why the Note was being delayed, if the 5nm had something to do with it. If that's the case, the 4nm modems are also in trouble, could be even worse.
 
Last edited:
In an interview with CNET, incoming Qualcomm CEO Cristiano Amon was asked "when is the supply chain issue was going to end" (at approximately 3:00 mark in the linked recording). Amon stated that "it will get better at the end of this year. We have line of sight and given our scale, we are very fortunate and very well-positioned...and we have line-of-sight that this is going to get resolved by the end of the year".

He sounded rather confident when making the statement. Rumor has it that the Snapdragon 888's successor (codenamed 'Waipio') may be built using a 4nm process. The interesting question is where will it be built (Samsung or TSMC)? While both Samsung and TSMC have announced a 4nm process, Fred might have unintentionally answered the question for us ("it does not have enough tooling for it"). Assuming Waipio will use 4nm, a Qualcomm move to TSMC would seem to add to the "Samsung can't yield" narrative.

Given that TSMC's capital expenditures will raise between $25-28 billion this year, the question on everybody's mind is what caused the raise? Qualcomm? Perhaps...but even with their "scale" it wouldn't be enough. Intel was the popular theory but it seems Daniel just poured some cold water on that. So instead of Intel...perhaps the return of Nvidia and Tesla? I'm hoping we'll get clues on April 15 (TSMC's next earnings call).
 
Back
Top