• Guest, The rules for the P & N subforum have been updated to prohibit "ad hominem" or personal attacks against other posters. See the full details in the post "Politics and News Rules & Guidelines."

TSMC investments and expansion 2021

Page 2 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Doug S

Senior member
Feb 8, 2020
525
716
96
This is really interesting, do we know what these decisions are based on? Like if 28/16 turned out to be more durable or have lower error rates than 20 or anything like that?
I think they decide well ahead of time - the customers who want to choose these processes for long life would need to know up front when deciding what nodes to target.
 

DisEnchantment

Senior member
Mar 3, 2017
757
1,801
136
Stopped in its tracks by rain/rivers/dams etc, not Samsung semiconductors.
No wonder they are open to building fabs elsewhere.
 

Doug S

Senior member
Feb 8, 2020
525
716
96

4 nm production starts in Q4 and Apple has bought the initial supply. Could be for 2022 Macs.

Apple is using N5P for the 2021 iPhone.
Interesting, I didn't expect them to use a half node. Maybe has to do with timing of when they want the Macs containing them to be introduced?
 

moinmoin

Platinum Member
Jun 1, 2017
2,369
2,941
106
Interesting, I didn't expect them to use a half node. Maybe has to do with timing of when they want the Macs containing them to be introduced?
My impression is that if TSMC is able to offer a new process node (half node or not) in time and at the quantity required for a launch, Apple will use it. I can only guess that N7+ failed at that, which is why first Apple (with A12 and A13) and then AMD (with Zen 2 and 3) made use of N7 twice which was uncharacteristical for both of them.
 
  • Like
Reactions: lightmanek

Doug S

Senior member
Feb 8, 2020
525
716
96
My impression is that if TSMC is able to offer a new process node (half node or not) in time and at the quantity required for a launch, Apple will use it. I can only guess that N7+ failed at that, which is why first Apple (with A12 and A13) and then AMD (with Zen 2 and 3) made use of N7 twice which was uncharacteristical for both of them.
Apple used N7P for A13, they didn't use N7 twice - just like the use of N5P for the A15 isn't using N5 twice. So basically they didn't need N7+ and perhaps also didn't want to take the risk of using the first EUV based process (for a few layers) at the time they had to make the call on what process to tape out for.

The use of N4 is interesting but Macs are an order of magnitude lower volume product than phones/tablets - especially if this is not a 4+4 design like M1 but something targeted at higher end Macs which have a volume that's less than TWO orders of magnitude smaller than the phone+tablet volume. It will be interesting to see what actually ends up getting made on N4. I'll bet it it not a 4+4 "M2" product unless that forms the basis for a "chiplet" design that scales up to the Mac Pro.

My money's on them not being confident of hitting the Mac Pro's planned launch of June 2022's WWDC using N3, so they took the best available process that they knew could hit that target. While I suppose the Mac userbase wouldn't care too much if it was announced in June but didn't ship until September, maybe Apple feels they've screwed the Mac Pro people too much already the last few years with launch delays and don't want it to make it happen again.
 

TheELF

Diamond Member
Dec 22, 2012
3,238
384
126
My impression is that if TSMC is able to offer a new process node (half node or not) in time and at the quantity required for a launch, Apple will use it. I can only guess that N7+ failed at that, which is why first Apple (with A12 and A13) and then AMD (with Zen 2 and 3) made use of N7 twice which was uncharacteristical for both of them.
N7+ would require a redesign of the chips and that would not be worth it since the increase in performance/density isn't that large.

Meanwhile, the N7+ process is expected to deliver a 20% increase in density, a 10% increase in performance, or alternatively a 15% power decrease. Although these improvements are slightly larger than what N7P offers, they also come with the cost of a new physical re-implementation and new EUV masks. The N7P process has already entered mass production in the last quarter (Q2 2019).
 

DrMrLordX

Lifer
Apr 27, 2000
17,031
6,001
136
@Racan

Intel has decent capacity. The problem they have is with process R&D. They spent a lot of time spinning their wheels on 10nm. If 10nm had worked better and had they switched more fabs over to that process, there wouldn't be as big of an issue - or potentially no issue at all. Building more fabs won't help Intel reinforce their IDM model. Fixing their nodes will.
 

EVI

Junior Member
Apr 20, 2020
4
11
41
TSMC to Spend $100 Billion Over Three Years to Grow Capacity - Bloomberg

ENVISIONED OUTLAYWHEN/WHAT
TSMC$100 billionOver three years to expand capacity
Intel$20 billionTo build two new fabs in Arizona
Samsung$116 billionOver a decade to expand foundry business

And Intel is going to catch up to TSMC?
This is a bit apples to oranges in comparing the numbers listed in this table. A more accurate representation is the following:

TSMC is spending $100B in capex over next three years, split something like $25B/$35B/$40B, all pure play foundry. Quick Edit: It looks like its not clear whether TSMC $100B is capex or capex plus R&D. Assuming some portion to R&D, the low end of TSMC capex would then be about $29B-$30B per year over the next three years.

Intel is spending $20B on two fabs for just its new foundry business, IFS. They are spending $20B in 2021 capex for its existing fab footprint. Future Intel capex will probably be high teens billions per year with the $20B in IFS spread over 3 years. I'd guess base-line capex will be $20B/$25B/$25B. This is all Logic and Foundry since Intel's NAND business will be at SK Hynix

Samsung's $116B foundry spending is not all capex as 55% is R&D expenses and 45% is capex for actual fabs. So Samsung foundry capex is $52B but that number is the guide for total aggregate spending from 2020-2030. So yearly capex from Samsung for Foundry is something closer to $5.2B per year over the next few years.

If anything, I'd say the read-through is that INTC has put forth a sizable spending gauntlet vs TSMC and TSMC is saying they will answer in kind. If I were Samsung, I'd be thinking about how much more spending needs to be done to remain competitive at leading edge foundry.
 
Last edited:

ASK THE COMMUNITY