In unusual move, TSMC not charging Apple for defective 3nm chips

[tonyget]

In unusual move, TSMC not charging Apple for defective 3nm chips

In an unusual move contrary to the normal way of doing business, chip fabricator TSMC is not charging Apple for defective 3nm chips ahead of…

macdailynews.com

In an unusual move contrary to the normal way of doing business, chip fabricator TSMC is not charging Apple for defective 3nm chips ahead of the introduction of the iPhone 15 Pro models which are said to be powered by 3nm Apple A17 Bionic SoCs, The Information reports.

Introduction of upgraded chip technology like ‌3nm‌ involves the production of a high number of defective chips until the manufacturing process can be perfected. According to The Information, TSMC is only charging Apple for “known good dies,” with no fee for defective chips. This is highly unconventional, since TSMC clients usually have to pay for the wafer and all of the dies it contains, including any defective ones.

Since Apple’s orders from TSMC are so large, it can apparently justify absorbing the cost of defective chips. Apple’s willingness to be the supplier’s first customer for new manufacturing processes helps it pay for the research and development of new nodes, as well as the facilities to make them.

The size of Apple’s orders also enable TSMC to more quickly learn how to improve and scale up a node during mass production. Once production and yield issues with manufacturing ‌3nm‌ chips improves and other customers seek the technology, TSMC can demand higher prices from those clients, as well as charge for defective dies.

 

[benb]

If this is true (which seems unlikely), TSMC would now be within the Apple event horizon, and no longer engaging at arms length. A massive strategic degradation.

Maybe TSMC just has excess capacity and offered it to Apple, that I could believe. It has an effect like KGD but maintains arms length.

 

[Daniel Nenni]

False.

 

[blueone]

False.

What's false?

 

[hskuo]

"Die-Buy" is rare for leading technology provider. It typically happens on catching-up supplier to attract customers for cost of switch. As we remembered, when ASML launched EUV, it used to provide service contract "paid by wafer output" which is similar to die-buy model in wafer manufacturer sites. It encouraged customers to adopt solutions earlier and eventually broke even to cover the cost which was not easy to justify then. I am not sure it is false or not, but possible.

 

[Daniel Nenni]

What's false?

That report from MAC Daily News is false.

Dr. CC Wei has reported on N3 yield many times in investor calls etc... I trust CC more than a MAC rumor site. I have also talked to people in the ecosystem who know. You can confirm by looking at the TSMC financial for FY23. TSMC cannot hide yield problems as big as what is rumored in this article. The only foundry I know of selling good logic die is Samsung. Not once have I seen TSMC do it.

 

[nghanayem]

Dr. CC Wei has reported on N3 yield many times in investor calls etc... I trust CC more than a MAC rumor site. I have also talked to people in the ecosystem who know. You can confirm by looking at the TSMC financial for FY23. TSMC cannot hide yield problems as big as what is rumored in this article. The only foundry I know of selling good logic die is Samsung. Not once have I seen TSMC do it.

One of those two people has a legal obligation to tell the truth... and the other; not so much. So yeah when CC talks about this kind of stuff it can't be patently false.

One potentially interesting thought is that the two claims could in theory both be true. As much as TSMC likes to dance around the issue, I think it is pretty clear at some point in time N3 was having issues of some kind (be they DD, performance, reliability, cost, cycle time, something else, or some combination of factors). Wafer contracts are also done in advance of production. So I suppose it isn't impossible that when TSMC started missing development milestones, Apple was able to renegotiate the contract to be know good dies to hedge against the possibility of further issues. If this is what happened, then known good dies and N3 having DD be in a healthy place are not mutually exclusive. There is also the side effect that if DD is in a good place, you would not expect known good dies to have a major negative impact on TSMC's financials.

However given the poor track record of process technology "leaks" akrihms razor would suggest we shouldn't put a ton of confidence behind any such claims.

 

[Daniel Nenni]

One of those two people has a legal obligation to tell the truth... and the other; not so much. So yeah when CC talks about this kind of stuff it can't be patently false.

One potentially interesting thought is that the two claims could in theory both be true. As much as TSMC likes to dance around the issue, I think it is pretty clear at some point in time N3 was having issues of some kind (be they DD, performance, reliability, cost, cycle time, something else, or some combination of factors). Wafer contracts are also done in advance of production. So I suppose it isn't impossible that when TSMC started missing development milestones, Apple was able to renegotiate the contract to be know good dies to hedge against the possibility of further issues. If this is what happened, then known good dies and N3 having DD be in a healthy place are not mutually exclusive. There is also the side effect that if DD is in a good place, you would not expect known good dies to have a major negative impact on TSMC's financials.

However given the poor track record of process technology "leaks" akrihms razor would suggest we shouldn't put a ton of confidence behind any such claims.

From the most recent call:

Now let me talk about our N3 and N3E status. Our 3-nanometer technology is the most advanced semiconductor technology in both PPA and transistor technology. N3 is already involved in production with good yield.

If a foundry is selling die versus wafers due to bad yield, which as I said TSMC has never in my knowledge done, that is not considered good yield. If I was the SEC or a big investor I would have a serious problem with this statement.

 

[nghanayem]

From the most recent call:

Now let me talk about our N3 and N3E status. Our 3-nanometer technology is the most advanced semiconductor technology in both PPA and transistor technology. N3 is already involved in production with good yield.

If a foundry is selling die versus wafers due to bad yield, which as I said TSMC has never in my knowledge done, that is not considered good yield. If I was the SEC or a big investor I would have a serious problem with this statement.

I agree if TSMC was still having yield issues while selling known good dies you should be able to see that on the books. Lying to investors is also a great way to end up in prison, so I do believe that TSMC is telling the truth that they are now in an okay spot. What I meant was that IF that contract was signed, it would have happened years ago back when TSMC was blowing past their original development milestones. I was just pointing out that a theoretical contract for known good dies being negotiated back in like 2020/21 doesn't mean that TSMC must be having yield problems now. Rather it means there were issues back whenever that alleged contract was signed.

 

[Daniel Nenni]

I agree if TSMC was still having yield issues while selling known good dies you should be able to see that on the books. Lying to investors is also a great way to end up in prison, so I do believe that TSMC is telling the truth that they are now in an okay spot. What I meant was that IF that contract was signed, it would have happened years ago back when TSMC was blowing past their original development milestones. I was just pointing out that a theoretical contract for known good dies being negotiated back in like 2020/21 doesn't mean that TSMC must be having yield problems now. Rather it means there were issues back whenever that alleged contract was signed.

Apple and TSMC jointly freeze the Apple specific process in Q4 for HMV in 2H so yield is known. The Apple TSMC wafer agreement is not the norm, not even close. It is a very demanding partnership for TSMC but it is also the reason TSMC now has dominant market share, absolutely.

 

[blueone]

Apple and TSMC jointly freeze the Apple specific process in Q4 for HMV in 2H so yield is known. The Apple TSMC wafer agreement is not the norm, not even close. It is a very demanding partnership for TSMC but it is also the reason TSMC now has dominant market share, absolutely.

It also seems like this relationship makes it very difficult and expensive for Apple to "diversify" its foundry business. Duplicating this process work with another parallel foundry sounds like it would be prohibitively expensive and probably need duplicated scarce people within Apple. So if TSMC had a hiccup, even a big one, going to another foundry also seems like it would be much worse in cost and schedule than just sticking with TSMC.

I'm guessing that when Intel uses TSMC to fab chiplets there is nothing like the Apple complexity involved. I would guess Intel just uses the standard process variant they select. Is that a good guess?

 

[hskuo]

It also seems like this relationship makes it very difficult and expensive for Apple to "diversify" its foundry business. Duplicating this process work with another parallel foundry sounds like it would be prohibitively expensive and probably need duplicated scarce people within Apple. So if TSMC had a hiccup, even a big one, going to another foundry also seems like it would be much worse in cost and schedule than just sticking with TSMC.

I'm guessing that when Intel uses TSMC to fab chiplets there is nothing like the Apple complexity involved. I would guess Intel just uses the standard process variant they select. Is that a good guess?

For intel, I would say it depends on the volume intel will commit and make order for customer specific process development(Phase-in products). The loyalty costs. The swtich cost in current advanced nodes are very high and you also need a strong team with track record to make it be successful. It is very challenging but not impossible.

 

[Daniel Nenni]

For intel, I would say it depends on the volume intel will commit and make order for customer specific process development(Phase-in products). The loyalty costs. The swtich cost in current advanced nodes are very high and you also need a strong team with track record to make it be successful. It is very challenging but not impossible.

The top TSMC customers have input to process development but only Apple gets a custom process that only they can use. It has been like that since 20nm.

 

[benb]

The top TSMC customers have input to process development but only Apple gets a custom process that only they can use. It has been like that since 20nm.

Dan this made my jaw drop a little bit but at the same time makes sense. And maybe charts the future development of this industry.
What this implies, I think, is that process integration doesn't exist at TSMC, at least as regards Apple, because Apple owns process integration.
A less aggressive interpretation, and more in line with industry norms, is process unlocks are very infrequent (once a year for example) and improvements pour into that window, at different paces depending on customer circumstances, which results in highly customer specific process flows.

 

[hist78]

"Recent media reports that TSMC is not asking Apple to pay for the cost of defective 3nm chips are different from my understanding.

There are two ways that customers buy chips from TSMC: finished goods buy and wafer buy.

The vast majority of deals are wafer-buy because TSMC's yields are good enough to ignore the cost of defective chips.

But Apple is a special case. Apple has always asked TSMC to provide the world's latest advanced node production services. Because the latest advanced node has many defective chips in the early stage of production, Apple has always purchased finished goods. And TSMC allocates most of the cost of defective chips to the selling price of each finished chip. The best evidence is that the cost of the new processor used in new iPhones increases significantly every year, and this year's A17 is no exception."

https://twitter.com/i/web/status/1689332407999270912

 

[Paul2]

"Recent media reports that TSMC is not asking Apple to pay for the cost of defective 3nm chips are different from my understanding.

There are two ways that customers buy chips from TSMC: finished goods buy and wafer buy.

The vast majority of deals are wafer-buy because TSMC's yields are good enough to ignore the cost of defective chips.

But Apple is a special case. Apple has always asked TSMC to provide the world's latest advanced node production services. Because the latest advanced node has many defective chips in the early stage of production, Apple has always purchased finished goods. And TSMC allocates most of the cost of defective chips to the selling price of each finished chip. The best evidence is that the cost of the new processor used in new iPhones increases significantly every year, and this year's A17 is no exception."

https://twitter.com/i/web/status/1689332407999270912

It reminds me how the industry was functioning when chips were still expensive

There was an era of enormous cheapening of commodity nodes in 200X, and when prices started to recover, 2008 struck, extending that into around 2015, when waves of rollups started

 

[Daniel Nenni]

"Recent media reports that TSMC is not asking Apple to pay for the cost of defective 3nm chips are different from my understanding.

There are two ways that customers buy chips from TSMC: finished goods buy and wafer buy.

The vast majority of deals are wafer-buy because TSMC's yields are good enough to ignore the cost of defective chips.

But Apple is a special case. Apple has always asked TSMC to provide the world's latest advanced node production services. Because the latest advanced node has many defective chips in the early stage of production, Apple has always purchased finished goods. And TSMC allocates most of the cost of defective chips to the selling price of each finished chip. The best evidence is that the cost of the new processor used in new iPhones increases significantly every year, and this year's A17 is no exception."

https://twitter.com/i/web/status/1689332407999270912

As I have mentioned before, TSMC and Apple develop the process as partners so there are no yield surprises. Apple knows the intimate process details and to meet time to market they may sacrifice yield to get chips for the iProduct launch. Or they may delay the process introduction and use an optimized version of a previous node. It is Apples decision. Now that TSMC does packaging for Apple you could say they only buy good die since Apple gets packaged chips but they are not buying die, they are buying packaged chips.

When the Apple/TSMC relationship was first announced (iPhone 6/20nm) I remember the doubters who said Apple would crush TSMC, buy TSMC, control TSMC, etc... In my opinion the Apple TSMC relationship will go down in semiconductor history as one of the best partnerships, absolutely.

 

[hskuo]

As I have mentioned before, TSMC and Apple develop the process as partners so there are no yield surprises. Apple knows the intimate process details and to meet time to market they may sacrifice yield to get chips for the iProduct launch. Or they may delay the process introduction and use an optimized version of a previous node. It is Apples decision. Now that TSMC does packaging for Apple you could say they only buy good die since Apple gets packaged chips but they are not buying die, they are buying packaged chips.

If it is "buying packaged chips", it will be defined as "turn-key service" which covers costs in wafer, substrate, assembly and testing. Will tsmc do full packaged AP chips or just InFo only?