Apple's Huge 3nm Chip Orders for iPhone 15 Pro and M3 Macs Helping to Maintain TSMC's Fab Utilization Rates

[_Spinn_]

Apple’s orders are so huge that it can literally make and break companies. Good ending ones include Corning, while bad ending ones include that company with sapphire glass.

Yeah it’s pretty crazy when you think about it.

 

[deconstruct60]

It's all very confusing. We've got a lot of conflicting statements/rumors about N3:

1) TSMC faced overcapacity, and had to scale back: Reports that Intel cancelled most of its 2023 N3 orders with TSMC, causing TSMC to scale back N3 production to avoid being over-capacity (https://www.electronicsweekly.com/n...-orders-tsmc-pares-back-n3-expansion-2022-08/).

2) TSMC is undercapacity: TSMC's January 2023 earnings call with investors, in which they said they expect to be under-capacity for N3 and 2023, and thus unable to meet all customer orders. ("As our customers' demand for N3 exceeds our ability to supply, we expect the N3 to be fully utilized in 2023.")

3) TSMC is overcapacity: Digitimes (which probably should never be taken seriously) saying that TSMC is significantly over-capacity on N3.

As I said in another post I suspect that the way 'N3' is being used is a bit sloppy and not being used consistently across all of these stories. Some folks want to cast N3 as only being N3E . I don't think TSMC uses it that way. I also don't think TSMC completely forks N3E completely out of the "N3 family" either.

My suspicions. #1 is partially related to ....

"Rialto Bridge cancelled ..."
https://www.servethehome.com/intel-announces-it-is-ending-traditional-hpc-platforms/

I don't think this was a huge amount relative to N3 being fully scaled up. But could have been a substantive amount in the first 1-2 quarters of the N3B production ramp. There were other reports about Intel buying up all the very early N3 production. Well, the very early N3 production is going to be smaller. So it is easier for a smaller buy to cover the a larger fraction early on.

Riato Bridge was mainly going to crank up the core counts and pull back on some of the excessive disaggregation of Ponte Vecchio. N3B isn't easy to work with and they are having trouble getting Ponte Vecchio out the door in numbers (in part due to the high complexity).

In short, that was only "oversubscription" in the first quarter ( maybe part of the second quarter) of HVM.


All compounded by Intel discrete GPU business scale out being a hot mess.


#2 is likely also talking about different quarters at a different stage of roll out process. Too many folks are skipping N3B and so there is giant log jam building for N3E which will only happen in the second half of the year. Q2 will likely all be under capacity and then in June-July it all breaks buck-wild. The slow baking N3B stuff will soak up a disproportionate EUV machines and and everyone who wanted to jump over N3B will soak up any slack.

Here TSMC is talking about the combo of N3B and N3E as 'N3'. If the A17 is on N3B then the iPhone demand bubble is going to overlap some with launch of N3E.

"we expected N3 to be fully utilized in 2023"

Doesn't necessarily mean the whole year. Just a substantive part of the year. Before N3E gets started (which has most of the tape outs ) there isn't going to be a way to saturate it.


#3 again , talking about Q1 '23 and parts of Q2 '23 isn't the same thing as talking about the whole year.


There is a bit of games played here also. When TSMC plays the "scarce wafers better get your reservations in" some folks block out capacity they may not use "just in case". That was chronic during peak of pandemic driven shortage.
A bit of the oversubscription is likely going on here at the end of 2023. If economy slows down in Q3 some of that will disappear.


Throughout much of 2022 there was lots of talk by some about how horrible N3 was and that everyone should skip until TSMC came up with something better ( i.e., N3E cheaper and shorter bake times). N3B didn't start to many peoples expectations. TSMC said 2H '22 and many took that to be likely July-September. It was closer to November-December (technically still in 2H '22 , but with longer bake times ... really won't deliver heavy volume until 2023. )


Have to locate which part of the ramp folks are talking about. Some folks think the HVM (high volume manufacturing) mark is 'late stage' rollout milestone as opposed to 'pre mid-life' . There is still capacity ramp after HVM.

 

[theorist9]

As I said in another post I suspect that the way 'N3' is being used is a bit sloppy and not being used consistently across all of these stories. Some folks want to cast N3 as only being N3E . I don't think TSMC uses it that way. I also don't think TSMC completely forks N3E completely out of the "N3 family" either.

My suspicions. #1 is partially related to ....

"Rialto Bridge cancelled ..."
https://www.servethehome.com/intel-announces-it-is-ending-traditional-hpc-platforms/

I don't think this was a huge amount relative to N3 being fully scaled up. But could have been a substantive amount in the first 1-2 quarters of the N3B production ramp. There were other reports about Intel buying up all the very early N3 production. Well, the very early N3 production is going to be smaller. So it is easier for a smaller buy to cover the a larger fraction early on.

Riato Bridge was mainly going to crank up the core counts and pull back on some of the excessive disaggregation of Ponte Vecchio. N3B isn't easy to work with and they are having trouble getting Ponte Vecchio out the door in numbers (in part due to the high complexity).

In short, that was only "oversubscription" in the first quarter ( maybe part of the second quarter) of HVM.


All compounded by Intel discrete GPU business scale out being a hot mess.


#2 is likely also talking about different quarters at a different stage of roll out process. Too many folks are skipping N3B and so there is giant log jam building for N3E which will only happen in the second half of the year. Q2 will likely all be under capacity and then in June-July it all breaks buck-wild. The slow baking N3B stuff will soak up a disproportionate EUV machines and and everyone who wanted to jump over N3B will soak up any slack.

Here TSMC is talking about the combo of N3B and N3E as 'N3'. If the A17 is on N3B then the iPhone demand bubble is going to overlap some with launch of N3E.

"we expected N3 to be fully utilized in 2023"

Doesn't necessarily mean the whole year. Just a substantive part of the year. Before N3E gets started (which has most of the tape outs ) there isn't going to be a way to saturate it.


#3 again , talking about Q1 '23 and parts of Q2 '23 isn't the same thing as talking about the whole year.


There is a bit of games played here also. When TSMC plays the "scarce wafers better get your reservations in" some folks block out capacity they may not use "just in case". That was chronic during peak of pandemic driven shortage.
A bit of the oversubscription is likely going on here at the end of 2023. If economy slows down in Q3 some of that will disappear.


Throughout much of 2022 there was lots of talk by some about how horrible N3 was and that everyone should skip until TSMC came up with something better ( i.e., N3E cheaper and shorter bake times). N3B didn't start to many peoples expectations. TSMC said 2H '22 and many took that to be likely July-September. It was closer to November-December (technically still in 2H '22 , but with longer bake times ... really won't deliver heavy volume until 2023. )


Have to locate which part of the ramp folks are talking about. Some folks think the HVM (high volume manufacturing) mark is 'late stage' rollout milestone as opposed to 'pre mid-life' . There is still capacity ramp after HVM.

So to summarize: Your suspicion is there will be plenty of capacity for anything that Apple wants to put on N3B during the first half (or so) of 2023, but things could get tight for N3E during the 2H 2023 (because it will have many customers in addition to Apple).

Didn't follow this sentence: "The slow baking N3B stuff will soak up a disproportionate EUV machines and and everyone who wanted to jump over N3B will soak up any slack." Do you meant that those customers who can't get enough N3E in 2H 2023 might substitute N3B?

 

[gund1234]

it doesn’t seem real because it isn’t. The whole nm thing largely lost meaning and it has little to do with the actual transistor size. Sure there has been huge progress and they can squeeze so many more transistors in the same area than they used to, but 3nm is a misleading number. The truth is that we got to the limit of what’s physically possible: silicon atoms have a size and you can’t change that fact no matter how much research you do.

If 2023 iPhone has 5 nm chip, people will complain that this chip is same as 2022, i am not buying this phone, it is re branded 14.
So Apple has to release new chip with new node every year to sell devices.

 

[deconstruct60]

So to summarize: Your suspicion is there will be plenty of capacity for anything that Apple wants to put on N3B during the first half (or so) of 2023, but things could get tight for N3E during the 2H 2023 (because it will have many customers in addition to Apple).

'In addition to Apple' I'm somewhat skeptical of. I won't be surprised if Apple doesn't use N3E at all. That they just jump straight to N3P later. If they already have N3B designs done in mid 2022 , then it doesn't make much sense to flip them for N3E. ( Only makes sense if N3P is going to be far, far too late for 2024 A-series run. ). If Apple keeps all of the A17 and M3-generation stufff on N3B that will soak up lots of capacity. Kind of expensive but Apple isn't selling their SoCs super cheap either.

Didn't follow this sentence: "The slow baking N3B stuff will soak up a disproportionate EUV machines and and everyone who wanted to jump over N3B will soak up any slack." Do you meant that those customers who can't get enough N3E in 2H 2023 might substitute N3B?

Processing N3B > 20 layeras takes longer than N3E sub 20 layers. Those extra layers soak up EUV machine time. it is going to take longer to do more layers. So the more N3B done the more the capacity is used up if only have a fixed set of EUV machines ( which TSMC pragmatically has.)

The folks getting N3E largely want lower costs and faster wafer delivery. They are not gonig to switch to N3B. They are not design library compatible so have to do a whole substatntive redesign. I strongly suspect apple was one of the primary folks asking for the feature set of N3 (now N3B) in the first place so they probably have much larger sunk costs into making it work so they will just go a whole broad generation on it.

Other players looked at N3 and said 'heck no... staying on n4 or N5" for an extra year. they are not switching to N3B, but they also are not going to ship bulk volume until much later into 2023. Vendors not evenly distributing themselves out over the options is what causes a log jam later in the year. Those vendors waited so they are now a larger group "in a hurry" so tighter scramble of N3E wafers as they ramp up and N3B probably isn't utterly collapsing either . [ In other words , there are a bigger group of vendors who probably time shifted out of Q3 '22 - Q1 '23 off into Q3 '23 -Q1 '23. Everyone at Disney World wants to ride Space Mountain in the same two hours there is a capacity problem . ]

 

[tipoo]

Apple’s orders are so huge that it can literally make and break companies. Good ending ones include Corning, while bad ending ones include that company with sapphire glass.

I still hope they're looking at sapphire iPhone screens some day. I always get the DLC black steel Apple Watch because the combo of DLC on the body and Sapphire on the screen always leaves those ones flawless after years, where the Aluminum and Glass are scratch prone. Would be a good feature to start on an Ultra one year and move it down through the Pros.