Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/what-are-tsms-plans-for-lower-end-chips.17374/
        )

    [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
)

What are TSM's Plans for lower end chips?

Arthur Hanson

Well-known member
I read that TSM was considering a fab in China for 28nm chips. Since TSM has been leading edge for years, should they not be able to use their older equipment for lower end chips and win market share since the equipment has already paid for itself. Any observations or comments would be appreciated. TSM should have a huge advantage in this area if they so choose.
 
TSMC already does this. Fabs are almost never converted to new nodes; rather they just continue making the same node for years to come. TSMC building extra trailing edge fabs is because their current trailing edge fabs cannot keep up with the explosive growth that happened/is happening with these nodes (you can see this with GF being profitable/having high utilization for the first time ever during the chip shortage). Building up in China is also a nice way to throw a bone at the PRC and placate them.
 
TSMC already does this. Fabs are almost never converted to new nodes; rather they just continue making the same node for years to come. TSMC building extra trailing edge fabs is because their current trailing edge fabs cannot keep up with the explosive growth that happened/is happening with these nodes (you can see this with GF being profitable/having high utilization for the first time ever during the chip shortage). Building up in China is also a nice way to throw a bone at the PRC and placate them.

This is true but the mature node bubble has burst. Last year companies were double and triple ordering to build up inventory. Car makers for example went from Just in Time inventory to Just in Case inventory. Fab utilization was at a record 100%+. Today fab utilization is closer to 90% (my guesstimate) and it will keep falling as companies get comfortable with their inventory levels.

Remember, orders were cancelled in 2019 and we had to make up for lost fab utilization and the new inventory levels. We are past that now so it is business as usual with a little more inventory than previous years.
 
This is true but the mature node bubble has burst. Last year companies were double and triple ordering to build up inventory. Car makers for example went from Just in Time inventory to Just in Case inventory. Fab utilization was at a record 100%+. Today fab utilization is closer to 90% (my guesstimate) and it will keep falling as companies get comfortable with their inventory levels.

Remember, orders were cancelled in 2019 and we had to make up for lost fab utilization and the new inventory levels. We are past that now so it is business as usual with a little more inventory than previous years.
Why build a mature fab in Germany then, what's up with folks like infineron/wolfspeed, and why are automakers still complaining about 40+nm? My understanding is that while the middle node bubble burst the trailing edge is to some extent still struggling. Of course I could totally be wrong here since you are more plugged into the fabless ecosystem than I am.
 
Why build a mature fab in Germany then, what's up with folks like infineron/wolfspeed, and why are automakers still complaining about 40+nm? My understanding is that while the middle node bubble burst the trailing edge is to some extent still struggling. Of course I could totally be wrong here since you are more plugged into the fabless ecosystem than I am.

TSMC and other fabs expanded mature nodes during the pandemic and ran the fabs at full capacity. It is my understanding that the fab builds in the EU are more part of the arms race to have semiconductor manufacturing in country. The problem with Automotive and foundries is that their orders are small potatoes in comparison to other markets so they were not prioritized. TSMC's automotive revenue grew by 74% in 2022. That is an unheard of amount for automotive so those orders are getting through and auto suppliers are stocking up. Based on this month's call, automotive shortages are now blamed on the "golden screw" which is not chips but other parts that are caught up in the supply shortage.
 
Dan, with autonomous and near autonomous driving coming on at an accelerating rate, will auto makers and many other areas of transportation from cars, trucks, aircraft becoming self-piloting could this not become a major market, especially as you add advanced robotics and autonomous everything into the equation? I also see robotic assisted surgery also providing a large and growing market. A fully automated blood drawer and tester was made a few years ago.

 
TSMC already does this. Fabs are almost never converted to new nodes; rather they just continue making the same node for years to come. TSMC building extra trailing edge fabs is because their current trailing edge fabs cannot keep up with the explosive growth that happened/is happening with these nodes (you can see this with GF being profitable/having high utilization for the first time ever during the chip shortage). Building up in China is also a nice way to throw a bone at the PRC and placate them.

Strategically the TSMC new 28nm fab in Nanjing China not only serves the needs from mainland China but also squeezes any new 28nm Chinese competitors out.

It's an investment that the authority in mainland China loves to have and hates to happen.
 
Last edited:
TSMC and other fabs expanded mature nodes during the pandemic and ran the fabs at full capacity.

Cite sources if possible? I have seen no evidence that TSMC expanded their capacity above 28nm. They definitely did run at full capacity, which helps... but it's not a good long-term solution. Almost all of the traditional industrial companies (TI/Analog Devices/NXP/ST/Renesas/Infineon/Microchip/
onsemi/Diodes Inc) did announce in earning calls that they expanded their own fab capacity... but nobody mentioned how much, except TI which is bringing new fabs online and I believe they gave some estimates of wafer capacity. And very little of it outside the foundries is in that 40nm - 90nm range; the only ones that have sub-90nm capacity are TI and ST (and Bosch), and maybe Infineon. NXP gave up after 90nm.

I would love to know how much extra capacity in the 40-90nm range was added in 2021 and 2022. My guess is that it's significant, but maybe in the 10-15% range at most, not too much ahead of the rate at which automotive/industrial/IoT is increasing.
 
Last edited:
TSMC's automotive revenue grew by 74% in 2022. That is an unheard of amount for automotive so those orders are getting through and auto suppliers are stocking up. Based on this month's call, automotive shortages are now blamed on the "golden screw" which is not chips but other parts that are caught up in the supply shortage.
How good are they at tracing the supply chain from foundry -> semi manufacturer -> Tier 1 supplier and identifying parts as "automotive"?

Pardon my skepticism but I don't see the auto chip shortage as over. I'll wait to hear what Kurt Sievers has to say in NXP's earnings call next week.

We *may* have neared/reached the point where TSMC is not the bottleneck and can deliver enough wafers for their own demand, but the semi manufacturers still have to push them through packaging and test. Some of the companies like NXP and Microchip have mentioned they have multiple quarters of noncancelable backlog, and are fairly sure that at least the noncancelable orders are a good measure of true demand and are not over-ordering.
 
Last edited:
Cite sources if possible? I have seen no evidence that TSMC expanded their capacity above 28nm. They definitely did run at full capacity, which helps... but it's not a good long-term solution. Almost all of the traditional industrial companies (TI/Analog Devices/NXP/ST/Renesas/Infineon/Microchip/

One way to look at fab expansion is through the CAPEX reports. How much CAPEX is going to mature nodes now versus before etc... I know TSMC has mentioned this in the past but I don't remember the split. It is something like 10-20% for mature nodes out of $30-40B CAPEX? UMC spent $3B or so last year and that is all 28nm and above and GF probably spent that much or more in Singapore. Same thing for SMIC. So add it all together and you get billions of dollars for mature node expansion in 2022.

Some people track this but they charge money for those reports (Scott Jones for example). I can get reports for free but cannot share or I just ask the questions.

I remember someone saying capacity optimization versus expansion meaning they can produce more wafers now than before but didn't necessarily buy new equipment. Sometimes it is a software upgrade or a process optimization. Fabs are also using AI to get cost reductions and capacity increases but that is very hush hush.

I'm not an automotive expert but the question I always have is why was there not an automotive chip shortage before the pandemic? Did people really buy more cars during the pandemic? The numbers I have seen do not support that.

Back to TSMC, the bottom line is that they do not want customers on mature nodes because the other fabs make TSMC "like" copies so the design GDS is easily taken from one fab to another. If customers move to FinFETs then it's TSMC's business forever. TSMC N16 down to N5 is automotive qualified which is a big tell. ADAS and other compute centric applications are already on FinFETs and there will be more to come. But of course TSMC listens to customers which is why they continue to invest in mature nodes even though they really don't want to.
 
Dan, with autonomous and near autonomous driving coming on at an accelerating rate, will auto makers and many other areas of transportation from cars, trucks, aircraft becoming self-piloting could this not become a major market, especially as you add advanced robotics and autonomous everything into the equation? I also see robotic assisted surgery also providing a large and growing market. A fully automated blood drawer and tester was made a few years ago.


Agreed. A friend of mine had surgery recently and it was robotic assisted. The doctor literally was behind a shield controlling the surgery. Pretty soon we will not need the doctor. I had a recent health issue with an abundance of tests and conflicting doctor opinions on the surgical solution. I got a third opinion, a doctor who pored over my records and test results and he came up with the answer and it was not surgical. It took months, AI could have done it in an hour.

It really is all about AI and AI has an insatiable compute requirement and that means leading edge nodes (FinFETs) and bigger die. TSMC owns FinFETs and they will own AI wafers, absolutely.
 
It is something like 10-20% for mature nodes
Yes, and that's the 22-28nm capex (for the Kumamoto fab), because that's what TSMC considers "mature".

I'm not an automotive expert but the question I always have is why was there not an automotive chip shortage before the pandemic? Did people really buy more cars during the pandemic? The numbers I have seen do not support that.
It's not more cars, it's more electronics in the cars. I don't think the worldwide # of cars is going up except maybe in China and India --- USA car sales are pretty much level over the last 40 years with some fluctuations (between 12.5 million and 17.5 million vehicles per month) dependent on the economy, and a short sharp dip in early 2020 with COVID --- see https://fred.stlouisfed.org/series/TOTALSA from the St. Louis Federal Reserve.

And honestly this was a disaster waiting to happen, with 200mm capacity getting more and more heavily utilized and automotive electronics growing by around 10-12% on average per year since around 2011. (see IC Insights bulletins, although Bill McClean has taken all that information off his site since he retired and sold the works to TechInsights. :-( ) Presumably 300mm capacity in 40nm and larger nodes has had similar increases in utilization but nobody breaks down those numbers beyond TSMC's/UMC's revenue-by-node disclosures in their quarterly reports.

At the risk of exposing my own flaws: I have written an article on the state of the automotive chip shortage which is what I have been able to gather about the situation from various published documents. You probably have a LOT more intuition than I do w/r/t the foundries, not to mention the scuttlebutt from customers; I have no idea who to talk to who would be willing/able to shed more light on the situation. I wish TSMC were more forthcoming about the 40-90nm range; there are a lot of MCU makers that depend on TSMC. I guess one of my conclusions is that there are some very unintuitive phenomena going on in the last few years.
 
Last edited:
Back to TSMC, the bottom line is that they do not want customers on mature nodes because the other fabs make TSMC "like" copies so the design GDS is easily taken from one fab to another. If customers move to FinFETs then it's TSMC's business forever.

I think mature nodes are low margin compare to advanced nodes,big companies usually trim/get ride of low margin part of businesses over time
 
Leading edge is the low margin bit. Trailing edge is where the profits are.

In the TSMC case, for a period time at the beginning of starting a new leading edge node, that new node's gross profit margin might be less than other existing leading edge nodes or mature nodes. TSMC calls it "below corporate average gross margin".

It really depends on the progress of a particular leading edge node and it may take two or more years for a new leading edge node to catch up the so-called "corporate average gross margin". On the other hand although a brand new leading edge node may not have the same high margin as other existing nodes (leading edge and mature), it is not necessarily a "low" margin one. For example TSMC N7 started with 1% to 2 % dilution to TSMC's "corporate average gross margin".

With the complexity introduced by the N3 and the already the high gross profit margin (59.6%), TSMC may take longer than two years to let N3 gross profit margin to catch up its corporate average.
 
In the TSMC case, for a period time at the beginning of starting a new leading edge node, that new node's gross profit margin might be less than other existing leading edge nodes or mature nodes. TSMC calls it "below corporate average gross margin".

It really depends on the progress of a particular leading edge node and it may take two or more years for a new leading edge node to catch up the so-called "corporate average gross margin". On the other hand although a brand new leading edge node may not have the same high margin as other existing nodes (leading edge and mature), it is not necessarily a "low" margin one. For example TSMC N7 started with 1% to 2 % dilution to TSMC's "corporate average gross margin".

With the complexity introduced by the N3 and the already the high gross profit margin (59.6%), TSMC may take longer than two years to let N3 gross profit margin to catch up its corporate average.
Yes, but think of the fraction of the profit. N7 didn’t have all the capacity it has now back then. So to dilute their corporate margin by 1-2% means that it might have only been sold at like a 20 or 30% margin. If you walk it back to the old days I thought I remembered seeing that in the very early days of a new node TSMC would even sell wafers at or below cost.
 
This is true but the mature node bubble has burst. Last year companies were double and triple ordering to build up inventory. Car makers for example went from Just in Time inventory to Just in Case inventory. Fab utilization was at a record 100%+. Today fab utilization is closer to 90% (my guesstimate) and it will keep falling as companies get comfortable with their inventory levels.

Remember, orders were cancelled in 2019 and we had to make up for lost fab utilization and the new inventory levels. We are past that now so it is business as usual with a little more inventory than previous years.

Certainly not by what I heard just in December. Small lot 200mm is still 9m to 12m lead time depending on a fab.

40nm-65nm supply has mostly eased, but few MCUs reach even those nodes.

And I don't see that even relatively high volume chips like display controllers, or display interface converters jump onto cheapest 300mm for budget offerings.

Another big thing that EVs are less modular, and have more original parts. This encourages designers to put control of more parts onto a single MCU. No need for CanBUS, and no need to go specifically for automotive MCUs for them, nor use CanBus converters.
 
Last edited:
At the risk of exposing my own flaws: I have written an article on the state of the automotive chip shortage which is what I have been able to gather about the situation from various published documents. You probably have a LOT more intuition than I do w/r/t the foundries, not to mention the scuttlebutt from customers; I have no idea who to talk to who would be willing/able to shed more light on the situation. I wish TSMC were more forthcoming about the 40-90nm range; there are a lot of MCU makers that depend on TSMC. I guess one of my conclusions is that there are some very unintuitive phenomena going on in the last few years.

The amount of electronics is only growing in high-end cars, and very high-end cars, I keep repeating it.

The trend with mainstream, and low-end cars, is that after an initial model is released, subsequent refresh models have smaller, more optimised BOMs than the initial generation. That's how the car industry works world-wide.


When electric cars will go mainstream in much of the world, I do not preclude them having even less ICs than piston engine powered cars, as more stuff will be moving from hardware to software.


For example, cheapest EVs in China already have two motors, and a single gear reduction, and NO ABS sensors, no torque sensors, since they can read axle speeds directly from the motor, and do basic ABSing by only using motors themselves.
 
Back
Top