TSMC’s 3-Nm Push Faces Tool Struggles | Hacker News Reader

TSMC’s 3-Nm Push Faces Tool Struggles

eetimes.com

65 pointscubefox70 comments

This is a small thing, but it's a bit of a pet peeve -- capitalization matters in SI units. A capital N is a newton, the unit of force. In the EE Times headline the N is correctly lowercased to spell "nm" -- nanometers.

See also: "mHz" being used to mean "megahertz" when it actually spells "millihertz". It's rarely ambiguous, but it's still wrong.

HN "fixes" capitalization upon submission... but if the user goes back and edits the title ex-post, it respects those capitalization decisions.

OP probably doesn't know that or did not notice the capitalization

It seems I can't edit it anymore now.

It's a little more problematic than usual in this case because a Nm would then be a newton-meter, which is a joule, which is not only common in general but relevant to electronics.

Joules are about as likely to come up in this thread as coulombs

"Nm" is the unit of measurement of torque.

3nm, 5nm, and so on are all just marketing terms completely disconnected from silicon geometry anyway. I get that it might be slightly confusing to see it misspelled, but it's still not that important.

This is something I’ve noticed for a while now: All of the major semiconductor manufacturers use the same set of tooling. In particular, they all use EUV lithography machines from ASML. So how can any one company maintain any sort of sustained lead? As a first order approximately, all of them should have the same level of fabrication capabilities. If TSMC tries to push too fast, they’ll run into major challenges. And if they figure something out, their competitors should match them in short order.

Semiconductor tooling is of a level of complexity that there's a huge amount of skill, institutional knowledge, process, and some art in getting them doing what you want at yields that pay off and on fast schedules. Lead is maintained by having all of that. They're also extremely expensive ($10b+ to get to leading edge node), which is a huge barrier to entry and just about requires nation-state help, as is the case in Taiwan, Korea, and the US.

That's like saying "all carpenters use the same hammers, how can one be better than the other?"—there's a lot more than goes into chip manufacturing than just the lithography machines. Not to mention that the lithography machines themselves are highly customizable and can technically work with different optical systems (not just the Carl Zeiss ones) and with different wafers (not just silicon ones), so there's a lot of room for experimentation.

But we don’t live in a world where we expect carpenters to double the complexity of their output every 2 years. At some point, the theoretic limits of the tooling becomes a thing.

I’m pretty sure you are confused about the optical system. All of the lithography machines have the same setup inside because only one company makes them (ASML). The next big upgrade involves upgrading the optics, but no one has that upgrade yet. We know, because again only company makes them and they haven’t produced them yet.

> All of the lithography machines have the same setup inside because only one company makes them (ASML).

I've definitely read papers where universities fiddle with the internals of ASML machines, including the optics, to try out different things. I'd imagine R&D departments at Intel, Nvidia, etc. probably do the same, but I'm no expert.

https://youtu.be/-EhDlXx3okU?t=203

No, that's not how it works. Dialing in a process is extremely difficult, expensive, and iterative, and a fab has to do a lot of dialing in to make a billion chips per year that each have a billion transistors which will do a billion calculations per second for a billion seconds with zero errors. Give or take a few orders of magnitude here or there. It's a long tail that has to be chopped off -- so long that the number of players capable of chopping it off has gone from three digits to 3.

> It's a long tail that has to be chopped off

Those ASML litho machines are astoundingly complex, with more parts and a more complex supply chain than a Boeing or Airbus commercial airliner.

So many parts that each part and subassembly must have an outrageously long mean time between failure, something like 16 years, in order for the entire machine to be sufficiently reliable.

No wonder only one company is able to make them.

Just a reminder you are on HN.

I am far from an expert, mostly parroting what I've heard from actual experts, but I think it gets back perhaps to a similar question of why doesn't ASML 'just' start a fabrication business since they already have the tools, and at cost no less?

The answer to that seemingly no brainer question is the exact same reason that TSMC have an edge over other fabs. There's a massive, incredible amount of complexity in the fabrication process, in particular doing it with economical yields - it's a completely separate business really.

In fact it might be the latter that's the least copyable part. For sure once competitors see that something is possible, some valuable information has been exposed, but doing that economically at scale is another matter entirely - and that's where there's a business to compete with. That can't simply be reverse-engineered or copied because the myriad details of the process are incredibly closely guarded IP!

Even then getting ASML equipment at cost only saves a few hundred million. They'd still be left with almost $20 billion in other expenses like the hundreds of millions in robotics and material handling equipment (the most precise in the world), hundreds of millions in HVAC and electrical equipment (heavily specialized to fabs), hundreds of millions in chemical cleaning (ditto), hundreds of millions in structure dampening, hundreds of millions in doping and chemical vapor deposition equipment, all the permitting and civil infrastructure to support the fab (water and power), and so on and on for a very long list.

The core competency of a semiconductor fab is getting all of that equipment to work together. It's more systems engineering (in the old school NASA sense) than anything else.

The tooling is only part of the equation...go read "Chip Wars" and it will give you a lot of insight into why TSMC is currently the leader. Your statements make a lot of assumptions and the landscape of semiconductor manufacturing is a lot more complex than you seem to realize.

I heard it's so complicated that they have built up a reservoir of institutional knowledge and in-house expertise that would take ten years to reproduce. I've also heard this as the reason that if China were to capture Taiwan and TSMC intact (however likely that is or not) then they still couldn't run it competently without those people, or without those ten years of more work. I don't know if it's true or not, that's just what I heard.

First, I'd say that sometimes companies don't use the same tooling even if they could. One of the things that gave TSMC an advantage was that they did push EUV while Intel didn't. Intel did make some investments in EUV, but pushed off the commercialization of it. So even when companies ca use the same tooling, sometimes people think something is too expensive and that they can get more out of something cheaper.

Beyond that, I'd say that we don't really want companies to have giant moats against competition. TSMC can keep an advantage as long as it keeps innovating. Having something a year before others can be important. Being able to produce better yields is important. Qualcomm and AMD wanted to move a bunch of their production to Samsung, but were disappointed in the yields despite Samsung having "the same process."

> As a first order approximately, all of them should have the same level of fabrication capabilities

Yes, but "approximately" is doing a lot of heavy lifting there. Each car manufacturer has "approximately" the same quality and value car. Windows/macOS/Linux are all "approximately" the same. One could easily reply that fabs are undifferentiated products and that is true to an extent, but there's a reason why Qualcomm and AMD went back to TSMC tail between their legs.

> if they figure something out, their competitors should match them in short order.

Again, I'd say "short order" is doing a lot of heavy lifting there. Are we talking a year? Two? Even when companies are using TSMC, they often can't get the latest process for a year or so after Apple because TSMC can't spin up capacity that fast - never mind a competitor that's trying to match TSMC.

I think in basically any industry, it's hard to keep a lead unless you continue executing well. Of course, one could argue that it's easy to keep a lead - just not over a couple other extremely well-run companies. Intel, Samsung, and TSMC have a huge sustained lead over basically everyone else. Does that not count? It's not like we're seeing lots of companies going around grabbing ASML EUV machines and spinning up even 7nm fabs because it's "easy". Even 5nm has been around for 2.5 years now. In fact, why did AMD wait 2 years to get 5nm if they could have simply bought an ASML machine and started printing the chips? Ok, I should walk back from hyperbole.

In some ways, I simply agree with you: they can't really have a sustained lead without continuing to put in work to stay 3-18 months ahead of the competition. However, I'd argue that a 6-12 month lead can be pretty big.

Intel is looking to introduce new stuff like RibbonFET (a successor to FinFET) and PowerVia backside power delivery so there's a lot more than "just use ASML EUV machines" and there is a lot of space for companies to innovate and create leads. But we also don't necessarily want companies to have leads that are too durable.

I guess: what is offering a durable lead to companies? Even companies that have well-differentiated products don't really survive well without continuous improvement. Some leads are more durable than others, but the investment and knowledge can be a durable lead.

Heck, look at TSMC's efforts to build a US-based plant. It isn't "just ship them a machine to print chips." They're struggling to make it work despite the fact that there's literally nothing to "figure out" in terms of the technology. They're already producing 5nm chips. Why can't they just start doing that in the US? It's just ASML machines and knowledge they already have! If TSMC can't figure out how to make 5nm chips in the US easily, why would it be easy for another company to be able to produce 5nm chips at a different place just because they could similarly buy ASML equipment and "figure out" whatever TSMC has figured out?

There's a lot more than just being able to order from the same parts bin.

> First, I'd say that sometimes companies don't use the same tooling even if they could. One of the things that gave TSMC an advantage was that they did push EUV while Intel didn't. Intel did make some investments in EUV, but pushed off the commercialization of it. So even when companies ca use the same tooling, sometimes people think something is too expensive and that they can get more out of something cheaper.

They're all on EUV now.

> In some ways, I simply agree with you: they can't really have a sustained lead without continuing to put in work to stay 3-18 months ahead of the competition. However, I'd argue that a 6-12 month lead can be pretty big.

That's my point. That's about as big of a lead you can get. There's nothing one company can do that can't be copied by another.

> Intel is looking to introduce new stuff like RibbonFET (a successor to FinFET) and PowerVia backside power delivery so there's a lot more than "just use ASML EUV machines" and there is a lot of space for companies to innovate and create leads. But we also don't necessarily want companies to have leads that are too durable.

They're all investing in GAAFETs. Their roadmaps look very similar.

The leading fabs have a closer relationship with TSMC that gets them working machines sooner and deeper knowledge transfer from staff. Plus what others have written.

>In particular, they all use EUV lithography machines from ASML. So how can any one company maintain any sort of sustained lead? As a first order approximately

Literally all modern Fabs uses DUV from ASML before EUV becomes a buzzword. So why did TSMC, or before that Intel has the lead? Why not Samsung Foundry, or Global Foundry? It is not first order approximation, it is a wrong assumption to begin with.

We actually had a thread about it not long ago [1]. The common analogy would be you buy the same 3D printer and expect all the outcome from everyone to be the the same ( not roughly the same, but "the same"). But that is not how it works. And any TWINSCAN is millions times more complex.

( Some of these discussions and tones are really getting quite tiring )

[1] https://news.ycombinator.com/item?id=35665076

In case one is using this info to make purchasing decisions, 3nm is a big improvement from today's leading edge 5nm process in speed, density, and power consumption [1]. We've been on 5nm variants for about 4 years now, and though each year brought gains, they were more incremental.

In Mac terms, this means M2->M3 will be bigger step up than M1->M2. Apple also delayed their next-gen raytracing GPU from last year, so it's likely that will come in as well. However with the difficulties of manufacturing, and Apple prioritizing iPhones, it's not clear whether we'll get any M3 Macs this year.

[1]: https://www.tsmc.com/english/dedicatedFoundry/technology/log...

Transistor density doesn't really matter by itself. What matters is that density influences speed, power consumption, and price. Historically all these have improved exponentially with exponentially increasing transistor density [1]. Yet:

> Power and Performance were the first two metrics to “fall off” the Moore’s Law curve. Transistor power reductions slowed nearly 20 years ago—each successive node generally reduced active and standby (leakage) current, but not at the 2X reduction per node. Transistor performance gains slowed shortly after; again, each node is generally faster, but not by 2X.

Moreover:

> Where Moore’s Law has slowed or even reversed, however, is the cost per component. Doubling density used to cut the cost per transistor in half, which was the primary enabler of the electronics revolution. Moore’s Law meant we’d get twice the number of transistors (which were faster and less power-hungry) for the same cost.

> As the chart (Fig. 2) shows, this began slowing at the 40-nm node, and virtually stopped when the industry moved to FinFETs. New nodes enable “close-to-Moore” chip-size scaling, but this is nearly completely offset by exploding wafer-fabrication costs. Various analyses show fabricated wafer costs increasing 3-5X over the 28-nm node, with 5-nm wafer costs approaching $18K each.

There is also the fact that SRAM scaling has almost stopped with TSMC's N3 [2].

[1] https://www.electronicdesign.com/industrial-automation/artic...

[2] https://fuse.wikichip.org/news/7343/iedm-2022-did-we-just-wi...

It'll be interesting to see how this plays out over the next year or two.

A lot depends on how well TSMC, Samsung, and Intel execute. Intel's roadmap has had Intel 4 happening in 2023, but it's possible that Meteor Lake will be pushed back a bit and it's looking like Intel won't be making high-end Meteor Lake processors. Intel also has RibbonFET and 20A on its roadmap for 2024/2025. If TSMC can't meet Apple's 3nm demand in 2023, it seems likely that only Apple will have 3nm processors in late 2023. Given that it was 2 years between Apple shipping 5nm and AMD shipping 5nm, it's not out of the question that AMD wouldn't get 3nm until late 2025 - and TSMC wasn't having trouble meeting Apple's 5nm demand in 2020. That leaves a ton of time for Intel to close the gap.

If TSMC is having so much yield trouble with 3nm, it seems reasonable to think that TSMC might not be able to keep up its advantage. Of course, one could also argue that Intel's roadmap might simply be bluster. That's why I think it will be interesting to see how this will play out. TSMC has done very well over the past decade while Intel has faltered. If Intel is able to get RibbonFET and 20A out in the next couple years, Intel might regain a nice lead.

If Intel can get its 20A out in 2024, it seems like that would put them solidly in the lead again and put a damper on both AMD (generally) and data-centers moving to ARM. AMD has seen a huge rise in part due to Intel's fab issues offering them a huge lead. ARM processors have also seen an opening via Intel's fab weakness. If Intel gets back to having superior fabrication, we might get to a market where Intel takes back the gains that ARM and AMD have seen in some areas.

But a lot of this depends on how well companies execute over the next couple years. Roadmaps are one thing. Processors in people's hands are another.

We are getting closer and closer to physical limitations of current microchip technology. Even TSMC starting to slip schedules is not really surprising.

When they are really getting close to physical limits, generational improvements will also be smaller. Then it is probably easier for others to get close TSMC. Currently being (e.g.) two years behind TSMC means being also substantially behind in performance. But when improvements slow down, such a two-year lag may not be so substantial anymore in terms of performance.

Don’t forget the slowing down of deflation in technology prices. Getting the same magnitude of improvement costs more and more and more R&D money.

Can someone help me understand something about these "roadmaps", the roadmaps that go out 5+ years and define the path of transistor shrinkage?

What I don't understand is... what's preventing a chip fab from leapfrogging a transistor size? Why must there be a gradual process, planned year by year?

If one manufacturer has even a modest lead, can they use this lead to invest in two generations down the line, to keep their lead cemented?

Does the technology that supports X nm transistors become an input to building X-1 nm transistors? Or is each new generation like starting all over again?

In the 70s and 80s, these leapfrog events would sometimes happen. One fab would find a way to get a huge lead and others would go bust. A chip would have 10 or so masks and a wafer might take a week to go through the fab before you could fully test it.

Nowadays there are 80+ masks and a wafer takes 4+ months to go through the fab. There may be 100k process parameters to tune, and if any one of those gets too far off your yield goes to zero. Finding process improvements is more difficult with these slow iterations and everything is so tiny now it is a miracle the chips work at all. Gradual improvement is all that is left.

It's funny how when Jony Ive and Steve Jobs were at the helm, Apple users always said performance wasn't everything and what truly matters is the complete package. Nowadays, it has flipped. The company's competitive advantage is their best in class chips. Every year I try high-end Windows laptops and Android phones and they are great to use but their performance aren't up to par and I always go back to Apple.

In the mobile market though, perf can directly translate into battery life.

Yeah, and it's important in the laptop market too.

The top line performance of the M chips is obviously fantastic per se, but probably more important for most use cases is that performance also being delivered reliably and dependably for a day (ish) on a single charge.

The combo of those two is a massively bigger win for most people than simply a little bit of extra incremental perf Vs the last gen, or whatever.

If you ask me user-facing performance was always a priority. iPhones always feel snappier than Android to me. Android phones often have a noticeable lag when scrolling, the apps take longer to open, the screen freezes more often.

Even with higher performing chips on some Samsung phones, the iPhone always feels faster.

Most consumers don't care about how performant the chip is, but they do care about visible performance.

> If you ask me user-facing performance was always a priority.

Go compare a Mac in 1998 vs. a garbage-tier shopping mall Windows machine. Apple almost died for a reason.

Samsung just makes ungodly slow Android UIs. I have stock Android on my Pixel 5 and I haven't noticed any slowness.

That may be true but here is the issue. An iPhone is an iPhone is an iPhone for the most part, the same cannot be said for Android. I don't have a Pixel but I have a Samsung phone to test with that makes me want to pull my hair out but a similarly old iOS device doesn't. The older iOS device might be slower than a new one but it still outperforms the Samsung.

> the same cannot be said for Android. I don't have a Pixel

It's a little silly to criticize their argument, then. Android is not a uniform experience because it runs on everything. Your Raspberry Pi can boot Android with the right drivers. It certainly won't be your "iPhone is an iPhone is an iPhone" experience, but that's because the AOSP exists. It's reductive and a non-argument.

Now, you also don't need to go build a fresh AOSP image or install GrapheneOS to argue in good faith. That being said, if you have to make vendors look like the boogeyman to win an online argument about cell phones, you're likely building a case that works equally well against Apple.

> That being said, if you have to make vendors look like the boogeyman to win an online argument about cell phones, you're likely building a case that works equally well against Apple.

Samsung is the top/biggest Android manufacturer from everything I can find online and all the other non-Pixel phones are equally trashy (full of crapware, reskinned apps, etc). The Pixel is almost a rounding error in North America coming in at ~2% vs Samsung at ~30% and Apple at about ~54% [0]. I think it's perfectly fair to judge the OS based on the biggest install base.

[0] https://gs.statcounter.com/vendor-market-share/mobile/north-...

> I think it's perfectly fair to judge the OS based on the biggest install base.

And not the designer of the OS? I guess that's fair enough, but you should be mad at Samsung then. Google licenses Android to thousands of manufacturers, and all of them can modify it. That's a non-argument though, like I was saying. Again, you don't need a Pixel to argue in good faith here, you're just being mad at the entirely wrong party.

Ahh, I apologize, I misunderstood the point you were making. "Android" is not to blame for "Samsung"'s implementation of it, I agree. What I was trying to say originally (and I'll admit I think we got off-track) is, more or less, that Android is not 1 "thing", it's a bunch of things and some of them are not well done (Samsung/etc) and other are better (Google) but the inconsistency taints the whole OS for me.

I'm coming at this both as a user and as a developer. My iOS apps or web pages generally have a bug or they don't across all of iOS, the same is not true for Android. I pay for BrowserStack first and foremost so I can test my app on the exact make/model/OS of Android that a user has who reports an issue (or that comes from Sentry). There are differences in the iOS versions but I've rarely been bitten by that whereas I have been for Android a number of times (normally due to Samsung vs Google vs Other, even if they are all the "same" Android version).

> If you ask me user-facing performance was always a priority.

You're absolutely correct, every major Apple architectural change has been motivated in large part by performance.

For example, the MacUser cover introducing PowerPC was "FASTEST EVER! PowerPC Macs • Five Times Faster • Outperforms Pentium". Apple was effectively forced to move to Intel because IBM fabrication caused PowerPC to fall behind performance goals. And of course, performance continues to be a primary motivation behind the move to Apple Silicon.

Performance per watt, not pure performance.

That's an important part of the performance story, but Apple wouldn't have switched architectures without a clear leap in pure performance even if performance-per-watt was better. Apple waited to transition desktops/laptops to Apple Silicon until they could demonstrate a definitive performance advantage over Intel.

Responsiveness definitely matters. It's something that users perceive even if they can't name it specifically.

On that note, it might've just been my imagination, but I swear that OS X 10.5/10.6 had a different responsiveness profiles on PowerPC and Intel Macs. Somehow Core 2 Duo Macs that on paper and in benchmarks outstripped the single core PPC G5 Macs they replaced quite significantly on every metric felt noticeably less responsive with more frequent beachballing. Really weird.

It's not like the Windows laptops are great on the design part either, witness all the threads complaining about Windows 11...

Everyone will tout performance when they're ahead in performance but that's not why customers are really buying. If the iPhone was slower than Qualcomm people would still buy iPhones. Likewise with Macs.

There are simply diminishing returns from increased performance. Smartphones already feel pretty responsive for a while now, e.g. in terms of app start-up time.

I disagree, the whole package is still my top priority, it's just that Apple is winning in performance (aside from GPU/graphics) and UI/UX right now. I think Apple has stumbled on UI/UX in the past fews years but I still feel they are leagues ahead of the competition. Windows and Android still drive me crazy and feel substantially worse in UI/UX than anything Apple has put out.

my girlfriend recently had ads in the browser from microsoft urging her to use bing for a month to earn a few dollars in amazon credit. I was baffled how you can cripple the UX of your flagship product without any respect to it, and in such a cheap, tasteless way.

so install firefox + uBlock Origin for her?

The ads are in the browser interface itself

also I wasn't really talking as if there's not fix for it, but the sabotage of MS on the UX of their flagship product. It feels cheap and not pleasant to use if you're bombarded with ads and stupid cortana-stuff.

I really don't like the painting of business leaders as heroes, but here it feels a little bit appropriate: I bet steve jobs would have fired you on the spot for this. And the managers under him would have never dared to do this and still don't.

The people building these things aren't using them (at least not in the standard user configuration)... or they don't have the ability to stop these things from going out.

I thought Google had generally great UI, even if Apple might be a bit better.

I have plenty of complaints about Google but their UIs are fine.

How has it flipped? Apple is still constantly talking about how performance isn't everything and it's about the complete package: iOS, MacOS and the accessory and app ecosystem.

Are you comparing with similar cost devices?

TSMC is getting help from customer Nvidia in lithography.

The “cuLitho” software and hardware is moving expensive operations to Nvidia GPUs, which will help TSMC deploy inverse lithography and deeper learning, according to C.C. Wei.

This jumped out as interesting to me. How would this relationship work? Is it like, "hey Nvidia, send us a few of those shiny GPU's and some engineers and we'll knock a few bucks off your wafer price"? Or more like, there is a division at TSMC that uses GPU's for process improvements, and they choose their vendor independent of any customer relationships?

It cuts deeper. Morris Chang is a hero of Jen-Hsun Huang and relationship goes way deeper than customer-vendor relationship. https://www.youtube.com/watch?v=u-x7PdnvCyI

Probably some Physical Design engineers who looked at the problem, and realized that it was a complex geometry problem that could be sped up buy parallelized computing. So they called up Synopsys (probably not TSMC) and were like "can you port this to CUDA?" The reason NVIDIA is taking credit for it is because Synopsys isn't know for its ability to deliver great software, and the folks at NVIDIA realistically probably wrote it.

3 Newton-meter?

edit: This was a comment on the incorrect capitalization in the title. SI units are case sensitive 1 Nm != 1 nm. HN should really stop the autocapitalization of titles.

>At present, we believe N3 yields at TSMC for A17 and M3 processors are at around 55%

What does that "55%" refer to precisely? Wafer yield? Isn't 55 quite low?

you didn't read the whole line

> 55% [a healthy level at this stage in N3 development], and TSMC looks on schedule to boost yields by around 5+ points each quarter.

55% of the dies on each wafer are good. That does sound like low yield.

> Isn't 55 quite low?

When Samsung lost Qualcomm as a customer for 4nm Snapdragon 8 Gen 1 chips, Samsung's yield was said to be about 35%.

https://www.sammobile.com/news/samsung-foundry-4nm-yield-sna...

Seems the server went down real fast.

https://web.archive.org/web/20230426182003/https://www.eetim...

It seems to be back

Another story where the lede is buried. In just two years we have gone from a horrible chip shortage to a massive chip glut and this is reverberating through the industry.

Reply on news.ycombinator.com