TSMC 2nm Update: Two Fabs in Construction, One Awaiting Government Approval
anandtech.com
anandtech.com
This is the first time TSMC has done two large feature set in one node generation. TSMC's N2 will first use gate-all-around (GAA), with an improved N2P coming with BackSide Power Delivery Network (BSPDN).
As with any new large node features, capacity will be constrained, so instead of the usual one additional Fab online per year expansion in leading node, they are bringing online 3 in the space of 2 years. This is also discounting that we dont know the total Wafer output of each Fab. But if we look at the recent Q4 report it does seems HPC ( or likely AI ) has infinite appetite for wafer capacity. So I wont be surprised if it is indeed a substantial increase in wafer capacity on leading node.
We will know ( or we can infer it ) once they start doing 2nm revenue reporting. TSMC tends to be very transparent with these sort of things. Something I hope Intel IFS will copy as well.
It takes ~3 years to stand up a fab and has only been 5-6 since EUV was proved out with first mass production.
TL;DR: TSMC bet on a particular manufacturing method (EUV), and it paid off. The company that makes the machines for that method is absurdly backordered. Also, people way overrate how ahead TSMC was/is, there was a halycon moment where Apple went ARM on desktop _and_ had exclusivity on TSMC's best node improvement in years. People were doing Apple's to oranges, ascribing the improvements from the die shrink to Apple's genius / TSMC's manufacturing. For verification dig into Apple community's in-depth reaction to recent M3 release, you'll find the take is generally def. not worth the upgrade if you have M2.
AMD hasn't stagnated, they've made some serious progress in the last years, and they've long ago gone fabless - they spun off GlobalFoundries in 2009, and mostly deal with TSMC. The biggest problem AMD has when compared to anything ARM is the instruction set - x86 carries a crapton of baggage with its 40-ish years of history and backwards compatibility, while ARM is relatively clean.
Intel stagnated because they messed up big with their node shrinking and for inner-political reasons couldn't say they go with TSMC as well so they were stuck with their old nodes.
That's why, I guess, they're looking to standardize x86s.
Intel in particular was finding ways to get older stuff to work for newer nodes but it also had drawbacks. And Intel had dividends to pay.
Now Intel is ordering the newest equipment, but hasn’t build up the experience with it. Yet.
Speaking of politics TSMC, is in a unique geopolitical position. Their government sees being the chip foundry for the whole world as an existential issue. China would like to take over Taiwan for historical reasons, and they can’t possibly build a large enough army to defend themselves against a really motivated full-strength attack. So instead they have put themselves in a position where any attack would be incredibly disruptive to the global economy and anger everybody.
The silicon industry is famous for very long roadmaps, very long investment payback times, cyclic demand, and winner-take-all competition. It is brutal. TSMC, serving everyone and with a friendly government at their back, is in a good position to out-endure everyone.
That’s the business part. The science and engineering is over my head, and I think sort of hard to talk about. Each node is a new engineering marvel basically.
The slower the progress, the easier it is for competitors to catch up "close enough". Currently being two years behind the bleeding edge makes a significant difference, but not as much as ten years ago, and probably more than a few years from now.
Though maybe I’m just out of the loop? who are the accessible close enough competition that you have in mind?
The entirety of the western world has made it clear, with cold hard cash and sheer political will, that they want out of the singular egg basket that is Taiwan.
Noone wants to be in the path of the inevitable freight train that is One China(tm).
Alternatively if we catch a unicorn and contact some aliens maybe they’ll tell us how to make CPUs out of the horn.
If the field actually stagnates for good, yeah, they’d have to pivot I guess. We’ll see if/when it happens.
Makes me wonder. At what point will Apple try to build/buy their own foundry?
Recall that famous article about the "largest hedge fund nobody has heard about" -- Apples hedgies that were a couple of people in Reno Nevada?
Surely they have been looking at the math forever - I wouldnt be surprised if they have already pulled the trigger on something.
However, they are masters of supply chain - and were investing in companies that are inputs to their supply chain, and rig the system for their own margins.
They would buy up a companies capacity for a component, and invest in them, and basically own that node in the chain.
With Ruthless brilliance.
As a customer, if TSMC falters, they can switch to Samsung or Intel (if Intel foundary materializes).
As a vertical foundary, it's very hard to switch if it falters. AMD managed to spin out their foundry which was becoming non-competitive and switch to TSMC, Intel has a lost decade because of the pipeline bubble caused when the Intel 10nm foundary arrived late and with poor yields.
Apple has a very hard time selling to business, and I think would not be able to credibly provide foundry services to others, so it only works if they can keep it on the leading edge.
As soon as they are tired of having money.
It isn’t just the cost of building Fabs, literally their entire supply chain is based on being Fabless, and changing that is ridiculously expensive. They have some amazingly skilled chip designers inhouse, but even so many aspects of the actual manufacturing side of the design process were not done at Apple, although they exert an outsized influence.
Likely never. As long as you understand how foundry works, economy of scale, long tail of older nodes, capacity planning and amortisation of R&D.
i.e Unless Apple is happy to Fab for their Android competitors and their Android Competitors are also somehow happy to buy from them the likely hood of that happening is practically zero.
I also asked because I remember in the early 2000s talking with people from Tower [1] and at that time tgey were super bullish for the model I think it does not work pretty well from them even with top-notch engineers.
Intel are catching up and will likely launch their 20A node before TSMC 2nm. Intel's 20A also includes gate all around and backside power delivery, so I don't think it's accurate to say Intel are 'stagnated' any more.
However, according to Intel's schedule, they've been at Intel 3 since Q3 2023 (and Intel 4 since Q3 2022). The first Intel 4 processors actually launched in December 2024, 17 months after Intel 4 on their roadmap. We've yet to see any Intel 3 processors. I'm not accusing Intel of lying. Their roadmap can be when they've achieved a milestone rather than having it at the scale to start shipping chips. To be realistic about Intel's roadmap, 20A processors are probably coming in December 2025 at the earliest. That is earlier than TSMC's 2nm with gate all around and backside power delivery, but then there's also the question of whether Intel will actually pull it off.
I do like that Intel is recommitted to its fab, but I'd also say that it's too soon for me to believe that Intel will overtake TSMC. Why? Intel 4 is behind TSMC's N5 process in transistor density and Intel has only shipped a single line of processors with it - and they've had to use TSMC's N5 and N6 for the graphics and IO tiles. Basically, Intel has started shipping Intel 4, but not for most of their processors - none of their desktop or datacenter processors are using it.
There is a possibility that Intel will pull it off and I agree that Intel isn't stagnating anymore, but I'm not sure I'd go as far as saying it's "likely" that Intel will overtake TSMC. Yes, TSMC is still struggling to get 3nm beyond Apple. Qualcomm's upcoming 2024 flagship chips are using TSMC N4P. But while Intel has been making progress way faster than it had for a decade, their progress hasn't been as stellar as their marketing of it. Intel 4 doesn't match 5nm transistor density, Intel 3 still seems to be a mystery, and Intel is shipping few Intel 4 processors. Does Intel 20A end up falling between TSMC's N4 and N3P? Does Intel ship one processor at very low volume with 20A in December 2025 and it's late 2026 or even 2027 before they've got all their processors there?
It just the kind of situation where there's a lot of nuance because it doesn't matter what a company has "achieved". It matters what parts you can buy and at what price. I can get Apple products with 3nm, but it looks like Android devices won't be getting 3nm until 2025. That's available, but not to most people. Even if Intel "achieves" 20A: at what volume, at what transistor density?
I think you meant December 2023
But to put it short, and one point that is not pointed out enough. TSMC has always been the leading edge foundry for nearly all SemiConductor company apart from Intel and AMD. If you want the most advance third party foundry, you basically only have TSMC to choose from. That has been the case since at least 2008. ( Or arguably a lot longer ). So for pretty much everyone in the industry nothing much has changed. TSMC mapped out their roadmap, speak to a few key client on their needs and capacity, plan their investment and ROI accordingly. The only thing that had changed is that they somehow not only are the leading edge on Foundry services, but also leading in terms All Silicon producer, where Intel used to be the king. So one way to think about it, is that TSMC has been executing like clock work for the past 20+ years. While Intel has multiple minor slip and one major slip in terms of node process, and also completely missed out on Smartphone revolution which provided an additional 2 Billion+ consumer electronics market using silicon chips, and they didn't get to enjoy the amortization over those potential 2 Billions unit volume. Even if they did like the 200M Apple Modem order, they completely failed to plan the CAPEX and OPEX spending and capacity planning. Which was later confirmed Intel's CEO Bob Swan. ( Not his fault. It was all on Brian Krzanich ). So to answer your question. "how they are advancing too fast while Intel (also AMD?) are "stagnated"?"
It was't TSMC moving too fast. It was Intel completely fucked it up. I wrote a lot about how Intel should have done Foundry in 2011 and how they missed everything by 2014. But a lot of these were only obvious now. Some of my comments below from 2016 and 2020.
Edit: That doesn't mean I am negative on current Intel. I was so happy and cried when Pat Gelsinger got back to Intel as CEO. It is a tough battle. But if Pat dont make it work then I doubt anyone else could either.
Copying what I wrote in 2020 [1]
>Tl;DR, It is TSMC, not Apple.
I have been thinking starting a blog on the topic considering how often this question keeps popping up. From a very high level overview.
1. Apple is partnering with TSMC. TSMC is now the leader in leading edge semiconductor manufacturing. A title that used to belong to Intel.
2. TSMC is now a generation ahead of Intel, meaning those thermal efficiency you see comes from using a better node. Nothing much to do with ARM or x86.
3. Both ARM and TSMC has dramatically change the Industry, you can now buy Designs / Blueprints from ARM, ( Or any other IP vendors such as Img PowerVR ) and Fab ( meaning producing them ) them with a Foundry.
4. TSMC is a Pure Play Foundry, meaning the Foundry does not produce their own chip and sell in the market to compete with its customers. A Non Pure Play Example would be Samsung or Intel, where Samsung produce their own Mobile SoC Exynos, and Intel with their x86 Chip. If you were Qualcomm producing your chip in Samsung's Fab, you are directly competing with them.
5. Apple now has the volume, or economy of scale to produce CPU themselves. Apple makes more silicon in unit volume than Intel per year.
Another Comments I wrote in 2016 [2]
> These process, R&D, process optimization, materials science are all cover by their cost to TSMC / Samsung. Btw TSMC's 2016 Capex ( including R&D ) is now equal to Intel $9.5B. When you consider TSMC is Pure Play and Intel being IDM, this is very significant mile stone.
It's hard to directly answer "how consequential", but it seems like it could be as significant as the step to TSMC 16nm.
That's not entirely true. They might not decrease the minimum gate length, but generally the improvements made when bumping down the "nm" number has been a similar improvement as what making the gate length smaller would have given. So I don't find the numbering as misleading as others think.
It's not like any of us a designing gate cell libraries (well, I have, a very basic one, but I'm guessing most here haven't), and actually care about the gate length.
Using the minimum gate length as a naming scheme was always been a proxy for things chip designers actually care about: logic density, timing, power, etc.
Next smallest is picometers, 1000pm = 1nm
"TSMC 2nm Update: Two Fabs in Construction, Third Awaiting Government Approval"
Customers would probably want to have plants outside of Taiwan, but proactively approving a new plant and starting to build one even before you have customers gives more leverage to Taiwan. This is just as much about Taiwan's security as it's about good business.
Edit: they have massive economic value of course, but economic factors alone have rarely prevented conflicts.
So not all conflict is tied-in with economics sure but wouldn't loosing most TSMC production and china-based supply chains be a major shock? For all sides? And as such would be taken into account?
I'm just curious where'd does TSMC find free real estate in Taiwan for so many fabs. At this rate Taiwan will be all TSMC fabs.
TSMC is working to bring water consumption down, but yeah, the water consumption is still critical
It's fair also that some farming activity is not very compatible with high human density nearby, such as spraying and dust. So it's hard to intersperse high rises with production farmland.
> At this rate Taiwan will be all TSMC fabs.
It's almost in every big cites, but there are still some regions in Taiwan that don't welcome TSMC to set up factories.
Famous last words
> Today, terms such as “2 nanometer” and “3 nanometer” are widely used as shorthand for each new generation of chip, rather than a semiconductor's actual physical dimensions.
Its almost as if we live in 1984 where newspeak goes un-noticed
https://en.wikipedia.org/wiki/2_nm_process
https://www.pcgamesn.com/amd/tsmc-7nm-5nm-and-3nm-are-just-n...
Who reserved all the expected output I wonder.
I really don't understand why this keeps getting brought up. It's not even an interesting fact.
When the chip manufacturing industry started finding other ways to increase logic density, improve timing, reduce power consumption, etc, they just started bumping down the "nm" number to indicate the improved performance, in a similar way as reducing the transistor gate length would have done before. Simple as that.