TSMC “Apple-first” 3nm policy leads to AMD and Qualcomm mutiny

[Spindel]

Summary

Rumours concerning TSMC’s and Apple’s close relationship continue to bubble up in Taiwanese media. A freshly-baked industry-sourced confection, from broadcaster TTV, suggests that both AMD and Qualcomm are now so fed up with TSMC’s purported “Apple-first” policy that they have decided to transition their businesses to Samsung Foundry services for 3nm..

 

Quotes

Quote

The Taiwanese media report characterises the battle for advanced semiconductor processes as “fierce,” and the ongoing tussle for capacity has led to a significant rupture, meaning that long-time TSMC customers AMD and Qualcomm will become the first major customers of Samsung’s 3nm process. Samsung is expected to have its first-gen 3nm process ready for mass production in H1 next year, but we won’t see flagship processors from them until 2023, according to the report.

 

My thoughts

While I can understand AMDs and Qualcomms frustration this is an issue of economics. If they want a piece of the cake I’m sure they can get it, if they out pay Apple. This is not some ”evil apple” concpiracy. Apple simply pay more to TSMC than AMD and Qualcomm do/can do and reap the benefits from it.

 

Sources

https://www.club386.com/tsmc-apple-first-3nm-policy-leads-to-amd-and-qualcomm-mutiny/

[Moonzy]

2 minutes ago, Spindel said:

Apple simply pay more to TSMC than AMD and Qualcomm do/can do and reap the benefits from it.

So, does AMD and Qualcomm expect tsmc to take lower offers or? Throw money, not tantrums

[Roswell]

It could also be a result of preference on TSMC's end as well. I imagine Apple's limited amount of chip SKUs is way easier to maintain than dealing with the amount of retooling and such that comes with Qualcomm and AMD's larger lineup.

 

Either way, welcome to capitalism, AMD and Qualcomm. 

[leadeater]

8 minutes ago, Roswell said:

It could also be a result of preference on TSMC's end as well. I imagine Apple's limited amount of chip SKUs is way easier to maintain than dealing with the amount of retooling and such that comes with Qualcomm and AMD's larger lineup.

 

Either way, welcome to capitalism AMD and Qualcomm. 

AMD only make CCDs, APU and GPUs at TSMC. The CCDs are literally a single design, desktop and laptop APUs are the same design, GPUs have (current) 3 designs. Apple across all the M1 and A series designs probably isn't much less. Qualcomm? Yea I care soooo little about them I have no idea lol.

 

17 minutes ago, Roswell said:

Either way, welcome to capitalism AMD and Qualcomm. 

Well the problem is it's probably more than just money that is the issue, TSMC for a long time essentially co-designs the next generation process with Apple and puts in contractual benefits due to that. I doubt TSMC is giving that same opportunity to AMD at all and additionally there is zero way to pay in more for earlier access otherwise that would be a breach of contract with Apple.

[LAwLz]

10 minutes ago, leadeater said:

Well the problem is it's probably more than just money that is the issue, TSMC for a long time essentially co-designs the next generation process with Apple and puts in contractual benefits due to that. I doubt TSMC is giving that same opportunity to AMD at all and additionally there is zero way to pay in more for earlier access otherwise that would be a breach of contract with Apple.

I think this is it too. 

Apple helps TSMC develop their process nodes and as a result Apple gets the benefit of getting it first. Also worth noting that it's not always as simple as one process node being good for all types of chips. The process node that Apple helped TSMC develop might not be suitable for what AMD wants. If TSMC has to prioritize development of one process they will pick the one that will have the most orders (which is probably Apple). 

 

Anyway, AMD have co-developed things with TSMC before. It was Google, TSMC and AMD that together developed the technology that is going into Zen 3D. 

 

If they helped develop it (which Apple probably did for N5 and N3, and AMD for 3D cache) then I think it's fair that they get short term exclusivity. I'm sure people would be mad at Intel and TSMC if Intel released a 3D cache version of Alder Lake on TSMC N7 like the same week as Zen 3D was released, and it was the exact same process node and technology. 

[porina]

I think the way the source puts it may be a bit more confrontational than needs be to get the eyeballs in that direction. Like a breakup, we wont know what happened without facts, which are lacking.

 

Good companies will try to do what is best for them in the future. Working with only one partner, just because you've worked with them significantly in the past, is not necessarily the only way forward. If Samsung's 3nm is competitive and cheaper and/or more available, why not use it?

 

On the flip side, things are certainly are not static. We have Intel also using TSMC. First consumer product on 6nm with the impending Arc GPUs. They're also using TSMC for parts of Ponte Vecchio, with 7nm and 5nm in use. Rumours suggest Intel is going against Apple for 3nm capacity, since Intel do not plan to regain fab leadership until 2025 with sub 2nm process. It might be AMD and Qualcom are unable to compete for TSMC's capacity.

 

On that last note, I did wonder about the relative size of the companies. Historically I've used market cap as a quick way to compare companies. Right now, apparently Intel and Qualcom are about the same around 200B. AMD are only slightly behind, at 188B. Ryzen certainly has been a success, but is AMD really that close to the scale of Intel now? P/E ratio might be more interesting here, where AMD is >4x higher than Intel. Or am I just looking for an easy to compare single number when I warn against such very often?  They'll do what they think is best.

[dizmo]

I really don't think it's as simple as just throwing money at them. Apple has invested hundreds of millions, of not billions into TSMC aside from what they buy from them as a product. That likely gets them some bargaining chips the others simply don't have. Apple also likely has far more expendable income, so they can afford to pay more for said capacity. It'll be interesting see what happens once all these new fabs come online.

[Fasterthannothing]

1 hour ago, Moonzy said:

Throw money, not tantrums

[pas008]

2 hours ago, porina said:

I think the way the source puts it may be a bit more confrontational than needs be to get the eyeballs in that direction. Like a breakup, we wont know what happened without facts, which are lacking.

 

Good companies will try to do what is best for them in the future. Working with only one partner, just because you've worked with them significantly in the past, is not necessarily the only way forward. If Samsung's 3nm is competitive and cheaper and/or more available, why not use it?

 

On the flip side, things are certainly are not static. We have Intel also using TSMC. First consumer product on 6nm with the impending Arc GPUs. They're also using TSMC for parts of Ponte Vecchio, with 7nm and 5nm in use. Rumours suggest Intel is going against Apple for 3nm capacity, since Intel do not plan to regain fab leadership until 2025 with sub 2nm process. It might be AMD and Qualcom are unable to compete for TSMC's capacity.

 

On that last note, I did wonder about the relative size of the companies. Historically I've used market cap as a quick way to compare companies. Right now, apparently Intel and Qualcom are about the same around 200B. AMD are only slightly behind, at 188B. Ryzen certainly has been a success, but is AMD really that close to the scale of Intel now? P/E ratio might be more interesting here, where AMD is >4x higher than Intel. Or am I just looking for an easy to compare single number when I warn against such very often?  They'll do what they think is best.

does nvidia have a piece of that pie too or are they all samsung now?

[Quackers101]

3nm or 14nm?

when does it become monopolistic?

at least if they do or when they hit 3nm, hope its stable and all that. also hopes on what intel new fabs or other chip giants do in future.

[leadeater]

1 hour ago, pas008 said:

does nvidia have a piece of that pie too or are they all samsung now?

Nvidia GA100 dies are TSMC 7nm, all the lesser stuff is Samsung

[porina]

35 minutes ago, Quackers101 said:

when does it become monopolistic?

If Intel were to try to buy TSMC, that could be monopolistic. Given TSMC are roughly 2 to 3x the market cap of Intel, suffice to say that isn't likely to happen even ignoring all the other problems that would attract.

 

At most there could be scenarios where Intel may be accuse of being anti-competitive, such as buying out capacity purely as a strategy to block competitors from using it. That wouldn't make any sense though. In the current market if you secure any allocated capacity, you're going to make good use of it.

 

12 minutes ago, leadeater said:

Nvidia GA100 dies are TSMC 7nm, all the lesser stuff is Samsung

"lesser" would depend on your chosen measure. If gaming performance is your thing, GA100 is not even capable of it.

[leadeater]

49 minutes ago, porina said:

"lesser" would depend on your chosen measure. If gaming performance is your thing, GA100 is not even capable of it.

Lesser in size, SKU, cost, importance to Nvidia, take your pick 

[porina]

3 minutes ago, leadeater said:

Lesser in size, SKU, cost, importance to Nvidia, take your pick 

For this forum, gaming performance is king?  

 

As for "importance to Nvidia" I actually had to look it up. Historically gaming is bigger in revenue than datacentre products to Nvidia, until last quarter results when datacentre overtook gaming. That was about 3 months ago, so we might be due another set of results any day now. See if that trend continues. I'm not familiar with nvidia's offerings in datacentre, so if the GA100 itself isn't all/most of it, gaming as a whole might still be more important for revenue. Unless you count per specific product, and at that point I have to stop myself over-thinking things. Again.

[leadeater]

30 minutes ago, porina said:

As for "importance to Nvidia" I actually had to look it up. Historically gaming is bigger in revenue than datacentre products to Nvidia

Yea it's bigger but Nvidia designs top down so all their archs are top end datacenter focused and everything is cut down from that.

 

The lesser dies are used in datacenter products too i.e. A40 = GA102

[porina]

13 minutes ago, leadeater said:

Yea it's bigger but Nvidia designs top down so all their archs are top end datacenter focused and everything is cut down from that.

I'm not sure the GA102 and "lesser" are a direct descendant of the GA100. GA100 supports an older version of CUDA and lacks many features in GA102. I don't know if it was omitted from hardware, or simply never enabled for differentiation. It is past my bed time and I'm not about to go down another Google hole. Without evidence my guess is that GA100 was designed for its specific task. The gaming class Ampere GPUs were an evolution onwards from GA100. Smaller, yes, but not the same. You can't cut down a GA100 to make a GA102. You could possibly cut down GA102 to lesser GPUs if desired. It might be somewhat comparable to Volta vs Turing in previous generation.

 

Then again I might have totally made all that up  

[Zodiark1593]

29 minutes ago, porina said:

I'm not sure the GA102 and "lesser" are a direct descendant of the GA100. GA100 supports an older version of CUDA and lacks many features in GA102. I don't know if it was omitted from hardware, or simply never enabled for differentiation. It is past my bed time and I'm not about to go down another Google hole. Without evidence my guess is that GA100 was designed for its specific task. The gaming class Ampere GPUs were an evolution onwards from GA100. Smaller, yes, but not the same. You can't cut down a GA100 to make a GA102. You could possibly cut down GA102 to lesser GPUs if desired. It might be somewhat comparable to Volta vs Turing in previous generation.

 

Then again I might have totally made all that up  

It wouldn’t make a lot of sense to have a product of the same line (and supposedly a flagship model at that) have a lesser feature set unless requested specifically, or perhaps bugged (Intel’s Haswell TSX, for example). 
 

Amazon AWS is a customer of Nvidia, potentially a large enough one that can request certain features or other changes. 

[leadeater]

43 minutes ago, porina said:

You can't cut down a GA100 to make a GA102

No you can't but the actual core arch is designed around this biggest die, the actual FP/INT and caches etc. The groupings of them are different between Gx100 and the rest. Gx100 also has FP64 units and the rest do not. Gx100 has HBM memory controller.

 

Wasn't really the point though, when Amprere (or any in the past) was designed it was designed for datacenter. Like we don't have Tensor cores in our gaming GPUs because those were added for gaming, Nvidia found a use for them in that use case but they were 100% designed for datacenter.

 

Cutting down/re-munging an arch and cutting down dies are different things.

[pas008]

2 hours ago, leadeater said:

Lesser in size, SKU, cost, importance to Nvidia, take your pick 

 

3 hours ago, porina said:

If Intel were to try to buy TSMC, that could be monopolistic. Given TSMC are roughly 2 to 3x the market cap of Intel, suffice to say that isn't likely to happen even ignoring all the other problems that would attract.

 

At most there could be scenarios where Intel may be accuse of being anti-competitive, such as buying out capacity purely as a strategy to block competitors from using it. That wouldn't make any sense though. In the current market if you secure any allocated capacity, you're going to make good use of it.

 

"lesser" would depend on your chosen measure. If gaming performance is your thing, GA100 is not even capable of it.

i was just asking if nvidia does have a piece of the 3nm pie though considering they did have some of tsmc 7nm

[Vishera]

Does TSMC hate money?

Because if AMD and Quallcom planned to use their 3nm process and now they go to SAMSUNG - That's a loss of really big customers

[Zodiark1593]

53 minutes ago, Vishera said:

Does TSMC hate money?

Because if AMD and Quallcom planned to use their 3nm process and now they go to SAMSUNG - That's a loss of really big customers

If their 3 nm capacity is literally 100% sold off to Apple (quite likely the highest bidder) already, what could they do? If AMD and Qualcomm wanted TSMC 3nm, they needed to have outbid Apple for capacity. 

[hishnash]

9 hours ago, Spindel said:

“Apple-first” policy that they have decided to transition their businesses to Samsung Foundry services for 3nm..

TSMC would not be so Appel first if Appel was not pre-purchasing the node capacity years before they are. Apple is using its massive amounts of $ to effectively fund the production of new nodes and therefore gets priority access. 

 

 

24 minutes ago, Zodiark1593 said:

If AMD and Qualcomm wanted TSMC 3nm, they needed to have outbid Apple for capacity. 

Note only do they need to outbid Apple they need to be willing to pay for this in advance to fund the node.

 

[leadeater]

4 hours ago, pas008 said:

 

i was just asking if nvidia does have a piece of the 3nm pie though considering they did have some of tsmc 7nm

They aren't even on any 5nm node yet and 3nm is much further off of anything like that.

[Spindel]

11 hours ago, leadeater said:

Well the problem is it's probably more than just money that is the issue, TSMC for a long time essentially co-designs the next generation process with Apple and puts in contractual benefits due to that. I doubt TSMC is giving that same opportunity to AMD at all and additionally there is zero way to pay in more for earlier access otherwise that would be a breach of contract with Apple.

Well it still boils down to "pay up or wait" for other companies than Apple. If AMD and Qualcomm would invest in node development like Apple do they would probably also be able to get access to it earlier.

 

Of course the end result would still be that Apple would have the rights to the latest node before AMD and Qualcomm simply because Apple can out invest AMD and Qualcomm anyway. It's a catch 22 of modern private sector economy; outpay the competition to get a technological advantage, but if the competition has more resources they can always out pay you and you are perpetually stuck behind unless you find another way to achieve your goals. 

 

In the end Apples investment benefits all from a technological standpoint. It's just that Apple reap the benefits of the development first. 

 

EDIT:// Canged "public sector" to "private sector" otherwise my post would make even less sense  

[Brooksie359]

12 hours ago, Moonzy said:

So, does AMD and Qualcomm expect tsmc to take lower offers or? Throw money, not tantrums

I think they are thinking Samsung is easier to get capacity at and maybe it's cheaper. I mean nvidia used Samsung for their rtx 3000 series and the performance of those cards was just fine. Tbh I think the biggest issue is there simply isn't enough capacity at tsmc so eventually the price premium they would have to pay over say Samsung will make people switch.