TSMC is scaling back its plans to produce 3nm chips next year after Intel postponed a major order

[Obioban]

Summary

Looks like Intel is no longer planning to use TSMC to 3nm production, at least this year.

 

Quotes

Quote

Intel reportedly planned to outsource production of its Meteor Lake tGPU chipset to TSMC, with mass production scheduled for the second half of 2022, before being delayed to the first half of 2023 due to design and verification issues. Now, Intel is said to have delayed mass production to the end of 2023, virtually cancelling the 3nm chip production capacity that it had booked with TSMC for most of next year. As a result, TSMC has apparently been "greatly affected" by the move, forcing it to slow its expansion of 3nm chip production to ensure that production capacity "is not excessively idle, leading to massive cost amortization pressure." Apple is believed to be the main customer of TSMC's initial 3nm chip mass production. According to the report, Apple is now the only major company among the first wave of 3nm chip production clients with orders scheduled between the second half of 2022 and the start of 2023. Apple's upcoming 3nm chips reportedly include new M-series chips and the "A17 Bionic." It is not clear if the disruption at TSMC caused by Intel will impact Apple's 3nm chip production volume or schedule. The first 3nm chip from Apple is rumored to be the M2 Pro, debuting in the 14-inch MacBook Pro, 16-inch MacBook Pro, and a high-end Mac mini model.

 

My thoughts

 Sounds like good news for Apple, as it was previously soundling like they were going to be a bit supply constrained on 3nm. Presumably this previously booked, now unused capacity will come cheaper, too.

 

Sources

https://www.macrumors.com/2022/08/04/apple-chip-supplier-hit-by-intel-delay/

https://www.trendforce.com/presscenter/news/20220804-11330.html

https://www.businesswire.com/news/home/20220804005450/en/Intel-Orders-Delayed-TSMC-Slows-Three-Nanometer-Expansion-Says-TrendForce

[porina]

I'm wondering what the implications of this are. Meteor Lake is/was supposed to be 14th gen expected late 2023, and was to be Intel's first "tile" based consumer offering (like AMD chiplets turned up to 11). I'm not aware of them indicating anything different, so if we assume it is still going ahead on time, it then flips over to why the difference? Are they now planning to sell a lot less of them than before? Is it delayed? Was TSMC single source for the GPU tile, or are they also doing a portion in house? Raptor Lake CPU cores are supposed to be build on Intel 4, but Intel 3 is also expected to be available in a similar timeframe. If Intel expects surplus of their own fab capacity maybe they're bringing it in, assuming they parallel designed the GPU tile to be made on both their own and TSMC "3" class prcesses.

 

Also I recall from previous rumour threads that TSMC were repurposing a site just for Intel 3 orders, so presumably that is no longer happening. Apple would be unaffected since they have dibs on TSMC's original 3nm capacity. We also had past news about both AMD and nvidia allegedly cutting and/or delaying their orders with TSMC, so the outlook over the next year or so is interesting. Feels like no one wants to overproduce leading edge silicon.

[DANK_AS_gay]

21 minutes ago, porina said:

Intel's first "tile" based consumer offering (like AMD chiplets turned up to 11).

Maybe the savings produced from the tile design offset the cost of additional fab equipment? It would make sense to move production in house ASAP, in order to save money and eventually drop prices to blow AMD out of the water in terms of Price-to-Performance (I really hope AMD can keep up, a duopoly is bad enough).

[AnonymousGuy]

Prettttty sure it's the other way around.  TSMC was overselling how far along N3 was and it hit problems that were going to delay it by 12 months so Intel had to move to other processes.

[LAwLz]

Seems like everyone is trying to pull out of their TSMC orders. 

First AMD. 

Then Apple.

Then Nvidia. 

Now Intel. 

 

I wonder if these companies are preparing for a recession where people won't be buying expensive computer components. 

[hishnash]

17 hours ago, Obioban said:

 Sounds like good news for Apple, as it was previously soundling like they were going to be a bit supply constrained on 3nm

very unlikely apple was every constrained by other orders, the nature of apple status for TSMC means for the most part if apple asks for supply TSMC will delay others orders to meet apples depends. This is what you get for in effect bankrolling the new node production. But yes if everyone else cancels apple can likly get cheaper prices, apple has not really been constrained when it comes to the core SOC chip volumes they have said in sales calls the chip shortages for them have been getting `legacy` nodes chips that are auxiliary to the main SOC silicon like little controlers for the webcam or thunderbolt etc.  

Of course with more 3nm (and 5nm) capacity not the table apple might well opt to just rule their own for these parts that are in demand. Apple clearly have already started down this path with the M2 MBA now having not only apples own power controlers but also now using apples own TB chips and removing some auxiliary prossing chips from parts like the trackpad by moving that logic within the M2 SOC.  

[porina]

 

Generally speaking, companies take a "no comment" approach to rumours and speculation, and it is exceptional for them to respond. I'm not sure what to make of that. 

 

Looking at the Trendforce piece, it doesn't actually say a lot. There's a lot of loose wording suggesting things. If we assume that Meteor Lake is on schedule for a late 2023 launch, I do feel manufacturing in volume earlier in 2023 is a more natural timescale to enable that. How much earlier depends on the quantity of product required, since there is quite some lead time from wafer start to wafer finish, of the order of a month or two depending on complexity. How many can TSMC push through in parallel? How much launch volume would Intel need? Maybe we could do a fermi estimation on that. If Intel now sees reduced demand than before, it makes sense to start volume production later so you don't have to sit on inventory too early.

[Donut417]

6 hours ago, LAwLz said:

I wonder if these companies are preparing for a recession where people won't be buying expensive computer components. 

I think its that and its the fact that demand is slowing just because the supply chain is starting to catch up. Considering crypto crashes and a whole load of used GPU's are hitting the market and there are only so many people now who need a GPU, it makes sense for companies like Nvidia to pull back for instance. 

 

[adarw]

wait they're going to do 3nm now? interesting

[ewitte]

Seems they like to scale up when demand is about to go down and scale down when demand is about to go up.  Somewhere in probably 2024 mining interest is going back up.

[williamcll]

On 8/5/2022 at 1:41 PM, LAwLz said:

Seems like everyone is trying to pull out of their TSMC orders. 

First AMD. 

Then Apple.

Then Nvidia. 

Now Intel. 

 

I wonder if these companies are preparing for a recession where people won't be buying expensive computer components. 

Could also be them looking for alternative manufacturers or setting their own up. Also the chances of a PLAN live fire exercise "accidentally" hitting a TSMC fab has somewhat increased.

[KaitouX]

On 8/5/2022 at 7:33 AM, porina said:

 

Generally speaking, companies take a "no comment" approach to rumours and speculation, and it is exceptional for them to respond. I'm not sure what to make of that. 

 

Looking at the Trendforce piece, it doesn't actually say a lot. There's a lot of loose wording suggesting things. If we assume that Meteor Lake is on schedule for a late 2023 launch, I do feel manufacturing in volume earlier in 2023 is a more natural timescale to enable that. How much earlier depends on the quantity of product required, since there is quite some lead time from wafer start to wafer finish, of the order of a month or two depending on complexity. How many can TSMC push through in parallel? How much launch volume would Intel need? Maybe we could do a fermi estimation on that. If Intel now sees reduced demand than before, it makes sense to start volume production later so you don't have to sit on inventory too early.

Intel probably wants to avoid confusion that their side is having issues. Meteor lake is supposed to use the Intel 4 process, so they might be trying to imply that their new node is on track, which would make sense that they want to avoid the image caused by the 10nm delays. There seems to be some confusion on where TSMC 3nm would be used too, rumors were initially indicating that the GPU tile on Meteor Lake would be the one using it, but some people said it Arrow/Lunar Lake would be the ones using it.

It seems like this slide is what caused the rumors:

Apparently some took this as a confirmation that Meteor Lake would use TSMC 3nm.

[porina]

1 hour ago, KaitouX said:

It seems like this slide is what caused the rumors:

I don't think I've seen that one before. Source? I find it interesting the lack of mention of Intel 3 on it. Unless I mis-remember or misunderstood, Intel 4 is a limited stepping stone node and will be mainly used internally, whereas Intel 3 is of a similar time frame and will be offered as part of IFS. We know they planned to use tiles, with parts being obtained from multiple sources, as demonstrated by Ponte Vecchio. 

[KaitouX]

45 minutes ago, porina said:

I don't think I've seen that one before. Source? I find it interesting the lack of mention of Intel 3 on it. Unless I mis-remember or misunderstood, Intel 4 is a limited stepping stone node and will be mainly used internally, whereas Intel 3 is of a similar time frame and will be offered as part of IFS. We know they planned to use tiles, with parts being obtained from multiple sources, as demonstrated by Ponte Vecchio. 

https://www.techpowerup.com/292143/intel-meteor-lake-and-arrow-lake-use-gpu-chiplets

The other slide there actually seem to imply the GPU is also on the Intel 4.

 

 

Meteor Lake was always on Intel 4, seems like consumer CPUs are going to skip the 3, the second slide below seems to imply Intel 3 is going to be in servers.

https://www.techpowerup.com/284936/intel-rebadges-10nm-enhanced-superfin-node-as-intel-7-invents-other-creative-node-names

https://www.techpowerup.com/292126/intel-tick-tock-alive-again-company-announces-new-intel-18a-node-1-8-nm-class

 

Granite Rapids seems to have gone from Intel 4 to Intel 3, so it's probably it.

https://videocardz.com/newz/intel-confirms-15th-gen-core-arrow-lake-xeon-granite-rapids-and-sierra-forest-2024-series

[porina]

A piece disagreeing with trendforce's claims: https://semiwiki.com/semiconductor-manufacturers/316286-intel-and-tsmc-do-not-slow-3nm-expansion/

 

I'm not familiar with this site either, but suffice to say they don't seem friends with trendforce.