Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/testing-the-bounds-of-loyalty-for-tsmcs-inner-circle.17401/
        )

    [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
)

Testing the bounds of loyalty for TSMC's inner circle

nghanayem

Well-known member
One of the great benefits of being fabless has been the ability to not worry about process, and just use the best node for your product. If you are fabbing bleeding edge chips, since N7 TSMC has been the unquestioned leader of customer service and process. At 16FFL we saw close partners of TSMC like Apple dual source with their old partner Samsung when 14LPP was working and 16FFL didn't have the volumes that Apple demanded. After all of the FUD on 14LPP iPhones being worse than the 16FFL iPhones apple stopped dual sourcing and probably never will again (at the very least for the same products). For the loyalty of AMD, APPLE, and to a lesser extent MediaTek TSMC rewards them with dibs on wafer capacity and prefered pricing. In the specific case of Apple they help shape the PDKs and get the new version of TSMC nodes 3-4 Qs before everyone else. With benefits like these moving part or all of your future capacity to a non TSMC fab WON'T happen even if Samsung or IFS have a node with say 5-10% better PPAC.

This is further supported by intel's claim that "We have engagements with 7 out of 10 of the largest fabless customers.". It is a safe guess that Apple and AMD are two out of those three. A year ago I would have assumed the third was MediTek because they only do business with Samsung on their lower end products where the better pricing is critical for competing with qualcomm and ending up in Samsung's phones. Other than this they seemed to be pretty loyal to TSMC, however the i16 deal obviously debunked that. Time will tell if they will ever leave TSMC for their more high end products. But now I couldn't even give a guess as to who the third might be.

With this information in mind how far would IFS or Samsung need to pull ahead to snag the likes of AMD or Apple for at least some of their business? Two full nodes; what about three? Would Samsung be able to snag AMD without too much difficulty given the two don't compete? Would IFS have an easier time snagging Apple than Samsung?

To be clear I don't expect these firms to ever leave the TSMC inner circle anytime soon, but it is interesting to think about what it would take for them to leave. Folks like Qualcomm and Nvidia will always be involved with all players, and cloud providers will go to whomever can give them the best chip. I suspect Broadcom and CISCO are also in a similar place. Snagging AMD or Apple would require these upstart foundries to be best in class across a wide variety of metrics.
 
Last edited:
Isn't Intel competing with Apple's Mac processors, and Samsung with Apple's phones? That's the issue with foundries who are also IDMs.
 
Yes and no. I would say that Apple competes with google directly and Samsung a little more indirectly. If Samsung makes a better phone it might encourage people to switch; but my understanding is that is very rare. People tend to be IOS people or Android people. I feel like this is even more the case with Macs vs PCs given the difference in software availability and price floor.

To me the biggest red flag for Apple going back to Samsung would be the whole IP theft thing that happened when Apple was using Samsung's 28nm process. Moving to Samsung again would again presents this risk, but it does seem like Apple can bend on this given they used 14LPP. Intel is less risky in this respect because they are more heavily invested in x86. For this same reason, I assume AMD would have an easier time being convinced to move to Samsung than intel.

Either way I agree they would never want to go to intel or Samsung, but theoretically how far behind can TSMC be before these companies can't make a competitive product anymore? For example Qualcomm was losing share to MediaTek and had to move from Samsung back to TSMC for their flagship products, and in the port process they got like 30% better efficiency as well as enough dies to more widely launch their flagship product. Another example was when N5 came out while intel was still stuck on 14nm, everyone and their mother was asking for a fabless transition (and many folks still are even after 10nm is out and working well). In these scenarios Qualcomm was about 1.5 years behind TSMC and Intel was like 2-2.5 years behind TSMC. If somehow TSMC fell 1 node behind, it is hard to imagine the inner circle deserts, 1.5 or 2 nodes probably, and if they fall any further beyond that I feel like there can be no question.
 
Last edited:
I'm not surprised by the MediaTek's contract with Intel Foundry Service because Mediatek has relationship with Intel in providing WiFi and 5G chips for a while.

In terms of who are those three companies out of the top 10 fabless who are not actively engaging with Intel Foundry Service, we can find some clues from the attached photo taken at the First Tool In ceremony at TSMC Fab in Phoenix Arizona in early December 2022.
 

Attachments

  • TSMC AZ Tool Move In 2022.jpg
    TSMC AZ Tool Move In 2022.jpg
    169.9 KB · Views: 151
From an outsiders perspective TSMC and NVIDIA seem to have a rocky relationship. Couple this with NVIDIA still being a big Samsung customer (and for that small period of time being mostly Samsung). If anything they seem like a prime target for going back to Samsung for a larger percentage of their product mix the second their nodes sub 7LPP nodes become competitive or sending out feelers for IFS.

Who knows though. Maybe NVIDIA is trying to get back into the inner circle like they were before blaming TSMC and Apple/other OEMs for their failures at designing a chip that didn’t cook itself, and blaming TSMC for poor yields when they didn’t follow design rules.
 
From an outsiders perspective TSMC and NVIDIA seem to have a rocky relationship. Couple this with NVIDIA still being a big Samsung customer (and for that small period of time being mostly Samsung). If anything they seem like a prime target for going back to Samsung for a larger percentage of their product mix the second their nodes sub 7LPP nodes become competitive or sending out feelers for IFS.

Who knows though. Maybe NVIDIA is trying to get back into the inner circle like they were before blaming TSMC and Apple/other OEMs for their failures at designing a chip that didn’t cook itself, and blaming TSMC for poor yields when they didn’t follow design rules.

Nvidia must do more business with TSMC to stay relevant in the TSMC Premier Club. Otherwise they were already a small TSMC customers in terms of volume. At the same time Nvidia knows Samsung can't meet the volume, performance, and quality requirement demanded by Nvidia.
 
Nvidia must do more business with TSMC to stay relevant in the TSMC Premier Club. Otherwise they were already a small TSMC customers in terms of volume. At the same time Nvidia knows Samsung can't meet the volume, performance, and quality requirement demanded by Nvidia.
Sure for data center. But even when they were majority SF data center was done on TSMC. For automotive and low end consumer graphics I don’t see any reason why they can’t still use Samsung for these purposes because 7LPP and above are very competitive with TSMC.

Either way that wasn’t the point of this query. Maybe a better way to articulate my query would have been how valuable is TSMC’s superior foundry/customer service (especially for the members of the inner circle)? Even IF TSMC fell behind, their other qualities could keep the fabs full. By asking how far behind TSMC could fall behind before their superior foundry services and customer service can’t make up the difference, I guess I really want to know how much technology plays into things. Presumably if at some point technology gaps cannot be ignored, and even members of the inner circle would start to take some of their business elsewhere? I’m curious where that point is, because as Samsung’s foundry is experiencing, being “not TSMC” also has it’s limits, and people won’t buy wafers from unusable nodes.
 
The game may be much different at the end of the decade than it is now. The density shrinks are certainly slowing and costs per transistor not dropping nearly as quickly as they used to.

That could drive either increased loyalty if someone is able to get to a node that others can't justify economically or decrease loyalty if we end up with a 3-way commodity where everyone's 10A process is roughly the same and no one fab company has enough money to justify 8A... (For example).
 
It's very difficult to give the answer to how fabless still want to place orders even if tsmc still fall behind. Here are few thoughts:
1. The reputation of guarantee in capacity/yield/timeline is so much better than other foundries. Morris Chang, the tsmc founder said tsmc is more like a service industry rather than technology industry.
2. Even the shrinkage is slowing. The technology/cost gaps in advanced package between tsmc and other foundries are even larger than in logic nodes from what I heard. The customers love to have the train service and worry less.
One thing about Intel IFS' strategy is they tend to emphasize once Intel gets back the node leadership, they can get big checks from other tycoons.
Being "interested"/giving small amount of 2nd/3rd source order is one thing, to take IFS as the major manufacturer is another thing. I'm still very pessimistic about Intel IFS' future.
 
Yes and no. I would say that Apple competes with google directly and Samsung a little more indirectly. If Samsung makes a better phone it might encourage people to switch; but my understanding is that is very rare. People tend to be IOS people or Android people. I feel like this is even more the case with Macs vs PCs given the difference in software availability and price floor.

To me the biggest red flag for Apple going back to Samsung would be the whole IP theft thing that happened when Apple was using Samsung's 28nm process. Moving to Samsung again would again presents this risk, but it does seem like Apple can bend on this given they used 14LPP. Intel is less risky in this respect because they are more heavily invested in x86. For this same reason, I assume AMD would have an easier time being convinced to move to Samsung than intel.

Either way I agree they would never want to go to intel or Samsung, but theoretically how far behind can TSMC be before these companies can't make a competitive product anymore? For example Qualcomm was losing share to MediaTek and had to move from Samsung back to TSMC for their flagship products, and in the port process they got like 30% better efficiency as well as enough dies to more widely launch their flagship product. Another example was when N5 came out while intel was still stuck on 14nm, everyone and their mother was asking for a fabless transition (and many folks still are even after 10nm is out and working well). In these scenarios Qualcomm was about 1.5 years behind TSMC and Intel was like 2-2.5 years behind TSMC. If somehow TSMC fell 1 node behind, it is hard to imagine the inner circle deserts, 1.5 or 2 nodes probably, and if they fall any further beyond that I feel like there can be no question.
Right, but the issue is IP theft across the Ios/Android or x86/Mac boundary.
 
From an outsiders perspective TSMC and NVIDIA seem to have a rocky relationship. Couple this with NVIDIA still being a big Samsung customer (and for that small period of time being mostly Samsung). If anything they seem like a prime target for going back to Samsung for a larger percentage of their product mix the second their nodes sub 7LPP nodes become competitive or sending out feelers for IFS.

Who knows though. Maybe NVIDIA is trying to get back into the inner circle like they were before blaming TSMC and Apple/other OEMs for their failures at designing a chip that didn’t cook itself, and blaming TSMC for poor yields when they didn’t follow design rules.

Morris Chang and Nvidia CEO Jensen Huang are VERY good friends, this is well known, I have seen them together at different events. In fact, the success of Nvidia and TSMC were closely tied together for a very long time. Nvidia's competitor was also a TSMC customer and TSMC treated them as equals. It was the same with Altera/Xilinix, QCOM/Mediatek, etc... That is how TSMC does business. Some companies like Nvidia and QCOM went to Samsung to leverage TSMC. Altera went to Intel for the same reason. Things changed when FinFETs came about. TSMC is MUCH more protective of their IP and rightly so. Now there is an inner circle for TSMC exclusive customers. Now that Samsung screwed up 5/4/3nm Nvidia has no choice. QCOM is going to Intel Foundry but Nvidia can't do that. Nvidia is back at TSMC until Samsung get's back on track. My opinion.
 
Morris Chang and Nvidia CEO Jensen Huang are VERY good friends, this is well known, I have seen them together at different events. In fact, the success of Nvidia and TSMC were closely tied together for a very long time.
Well that's nice. The early histories of these two firms is so interwoven, and they both contributed to each other's immense success today. It would have been a shame if NVIDIA being too arrogant at 40nm and moving the majority of their products to Samsung for a time burned that relationship that in the past brought so much success. Unfortunately for Jensen NVIDIA's relationship with Apple does not seem to have been nearly as resilient.

While I agree that betting on TSMC is certainly the safe bet, if intel does prove itself what is preventing NVIDIA from going to intel for some of it's product stack? Isn't NVIDIA's secret sauce their software? Their hardware only seems slightly ahead of intel/AMD's data center offerings and obviously far less sticky than the software that runs on it.
 
While I agree that betting on TSMC is certainly the safe bet, if intel does prove itself what is preventing NVIDIA from going to intel for some of it's product stack? Isn't NVIDIA's secret sauce their software? Their hardware only seems slightly ahead of intel/AMD's data center offerings and obviously far less sticky than the software that runs on it.

With the advent of chiplets I think using multiple fabs for chip design is again an option. Remember, back in the day QCOM could and did use multiple fabs for the same design to get the best cost and/or best delivery. Today Intel is using TSMC for chiplets intermixed with Intel chiplets. Nvidia and others can and will probably do the same. That way they can get the best performance or the best cost or the best availability for different chiplets and spread the wealth amongst foundries. It is in everyone's best interest to have multiple fab sources and chiplets can certainly help with that.
 
The challenge of new nodes are getting higher and higher. There is no guarantee each player will succeed, including TSMC.
Why would customers take huge risk changing foundries?

TSMC is like IBM decades ago, even if it fail, everybody fail, your boss would blame you.
 
Multi foundries strategy has its limitation on solving the supply chain interruptions.

For the same model of chip, unless a fabless company has multiple teams doing parallel development and did tapeouts at multiple foundries(companies), they can't just switch from foundry A to Foundry B overnight. It will take months if not years to switch. But even if they do have such capability, it doesn't mean the foundry will have the capacity readily available for them. More than often, the capacity won't be there at least for another 3 to 6 months. There are only a few fabless companies in a rare chance can afford to do that.

A more workable approach for a fabless company is to use the same foundry but at multiple sites.
 
Multi foundries strategy has its limitation on solving the supply chain interruptions.

For the same model of chip, unless a fabless company has multiple teams doing parallel development and did tapeouts at multiple foundries(companies), they can't just switch from foundry A to Foundry B overnight. It will take months if not years to switch. But even if they do have such capability, it doesn't mean the foundry will have the capacity readily available for them. More than often, the capacity won't be there at least for another 3 to 6 months. There are only a few fabless companies in a rare chance can afford to do that.

A more workable approach for a fabless company is to use the same foundry but at multiple sites.
What about the NVIDIA or MediaTek model of doing some parts of their product stack at one firm and other parts at another? Does putting one SOC or say 16FFL and another on 14LPP not accomplish a similar result? Obviously since 5/3nm is kind of a one man race at the moment you are forced to single source for products that need the latest nodes. However you can (and many firms do) still throw Samsung bones at 14-7LPP.
 
What about the NVIDIA or MediaTek model of doing some parts of their product stack at one firm and other parts at another? Does putting one SOC or say 16FFL and another on 14LPP not accomplish a similar result? Obviously since 5/3nm is kind of a one man race at the moment you are forced to single source for products that need the latest nodes. However you can (and many firms do) still throw Samsung bones at 14-7LPP.
16nm/14nm cost and risk was a lot lower than 5nm/3nm now.
 
What about the NVIDIA or MediaTek model of doing some parts of their product stack at one firm and other parts at another? Does putting one SOC or say 16FFL and another on 14LPP not accomplish a similar result? Obviously since 5/3nm is kind of a one man race at the moment you are forced to single source for products that need the latest nodes. However you can (and many firms do) still throw Samsung bones at 14-7LPP.

Not sure the "parts" you mentioned are having the identical in design, features, and performance? If not then they are different products to begin with. Basically it's still be a single sourcing for a particular product.

Apple, Nvidia, AMD, Qualcomm, Mediatek and many fabless companies always source their chips from multiple foundries. The problem is that when a disaster happens like a earthquake in Taiwan, a fire in Japan, or a power grid collapse in Texas that stops a foundry to producing their chips, can a multi foundries practice save them?

In a short term they can't. In the longer term, the fabless company will be forced to use another foundry if the fab can't come back alive. But it's still a single sourcing for a particular product at a given time.

That's why to use a foundry who can provide the same node manufacturing across multiple nations for a particular product is more practical.
 
Not sure the "parts" you mentioned are having the identical in design, features, and performance? If not then they are different products to begin with. Basically it's still be a single sourcing for a particular product.

Apple, Nvidia, AMD, Qualcomm, Mediatek and many fabless companies always source their chips from multiple foundries. The problem is that when a disaster happens like a earthquake in Taiwan, a fire in Japan, or a power grid collapse in Texas that stops a foundry to producing their chips, can a multi foundries practice save them?

In a short term they can't. In the longer term, the fabless company will be forced to use another foundry if the fab can't come back alive. But it's still a single sourcing for a particular product at a given time.

That's why to use a foundry who can provide the same node manufacturing across multiple nations for a particular product is more practical.

Agreed, TSMC and Samsung having US based fabs is key. Same thing with packaging, test and everything else a wafer succumbs to before being a system ready chip? What about at the system level? Can Apple go from design to finished product in the US or any other country? Is it going to get up to that level or will we stop at redundant wafer manufacturing?
 
Multi foundries strategy has its limitation on solving the supply chain interruptions.

For the same model of chip, unless a fabless company has multiple teams doing parallel development and did tapeouts at multiple foundries(companies), they can't just switch from foundry A to Foundry B overnight. It will take months if not years to switch. But even if they do have such capability, it doesn't mean the foundry will have the capacity readily available for them. More than often, the capacity won't be there at least for another 3 to 6 months. There are only a few fabless companies in a rare chance can afford to do that.

A more workable approach for a fabless company is to use the same foundry but at multiple sites.

If you want to push the same device through multiple foundries (e.g. TSMC and IFS or Samsung) at advanced nodes, you have no choice except to have separate development teams, because the foundry contracts require that there must be zero chance of any technology leak -- engineers who work on a TSMC process have to be totally firewalled from employees who work on a competitor's process, to the extent of having completely separated teams, tools and licenses.

This pretty much doubles the back-end development effort and cost, so very few companies can justify doing it due to the cost and resources needed -- only truly massive projects with huge revenue would even consider it.
 
Back
Top