[nwiswell]

Mostly these machines are assembled from a combination of proprietary parts manufactured using regular machine tools by contract manufacturers, and from sub-units purchased from specialist suppliers.

For example: proprietary parts including optics, wafer chucks, process chambers, structural panels, etc are manufactured to spec by contractors. Pumps, valves, motion control boxes, wafer robots, loadports, servers, etc are all bought from suppliers.

There's not really any magic. The materials used for the parts are sometimes exotic (e.g. single crystal silicon, single crystal quartz, yttria coating, PEEK, sapphire, etc), but the contract manufacturers use the same old boring machine tools that everyone does.

[cactusplant7374]

> There's not really any magic.

I've read that it would take China at least 10 years to get to ASML's current level. It might not be magic but there must be a lot of unknowns when starting from scratch.

[nwiswell]

Yes, the design of those proprietary parts is very technical, but the manufacturing not so much. That's part of what makes the IP situation so tenuous and is a big part of why the FBI has aggressively prosecuted individuals for corporate espionage against WFE firms in the past.

[cactusplant7374]

The tolerances are very tight. Manufacturing must be difficult.

[bradknowles]

I worked a six month contract at ASML years ago. One of the most fascinating things was touring around the spaces with an experienced team leader, who explained a lot of their processes. Note that the below comes from my experience there prior to 2006. Nothing that I'm going to explain here is company proprietary information. Even if I had that, I wouldn't expose it.

Each top machine costs hundreds of millions of dollars to build. They are built by a dedicated team of people who only work on that one machine, and they take it from soup to nuts, including the installation in the field. They spend months taking measurements at the site before they ever start building the machine in the first place. Their placement of the machine and the measurements is very precise -- if the machine has to be moved by a centimeter in any direction from the original measurements taken, they have to basically start over from scratch.

The machine is completely built in clean room facilities at the HQ in Eindhoven, taken all the way to the point where it is used to do good size batches of test runs of known chip designs. Those clean rooms operate way beyond the cleanliness levels of a normal chip lithography clean room. They are some of the cleanest clean rooms in the world.

Once completed, the machine gets completely disassembled and shipped to the customer, along with the team who built it. Then they rebuild it on site. Once rebuilt, it will take months of running in and tweaking before it's able to be considered fully operational. That process alone can take over a year, if the team runs into more problems than typical.

The machines are so sensitive that a truck driving down the road ten miles away can affect the output and may cause the machine to be knocked out of calibration.

From the time of placing the order to the time when the machine is considered fully operational, multiple years have passed.

And that was way back before 2006, when my wife and I moved back here to the US. I had seen the door behind which all EUV design was done, but I never got a chance to see inside. It was locked up like Cheyenne Mountain. And yes, I did previously have a TOP SECRET/SCI clearance, and I've been briefed on exactly what the ANMCC and Cheyenne Mountain look like.

I have no idea how much more complex and sensitive the processes used by ASML have gotten since. I do believe the top machines now cost over a billion dollars, each.

[nwiswell]

> I do believe the top machines now cost over a billion dollars, each.

This isn't accurate, the top end is no more than $200M (and large customers will negotiate down from there).

> The machines are so sensitive that a truck driving down the road ten miles away can affect the output and may cause the machine to be knocked out of calibration.

This is obviously untrue, fabs are busy industrial centers.

The assembly-partial disassembly-reassembly process you laid out is accurate, however, and is the case for all major wafer fabrication equipment.

[seper8]

Wrong again

https://www.reuters.com/technology/ intel-orders-asml-machine-still-drawing-board-chipmakers-look-an-edge-2022-01-19/

[nwiswell]

Not really any harder than making an airplane.

For older generations of tools, there are a lot of bootleg replacement parts floating around on secondary markets. When tools reach their end of life at leading-edge fabs, or when those fabs are shuttered, the tools are often sold off to trailing edge fabs in permissive jurisdictions with shoestring budgets.

These fabs are not averse to buying whatever works cheapest, and the OEM WFE companies don't really have any way to enforce IP.

[seper8]

>Not really any harder than making an airplane. Except the airplane still works if something is misaligned by 10NM...

[nwiswell]

That's adjusted dynamically during operation by interferometer controls, not in manufacturing. You'd get more than 10nm of misalignment just due to a couple degrees of temperature change.

[markus_zhang]

10 years is probably an underestimate given then current geopolitical condition.

[rjzzleep]

Just last year people were commenting here on HN that China is years away from making anything under 14nm and that even that would be hard for them. Then we heard about China being years away from hypersonic missiles and how they couldn't even make fast memory chips.

Then we were surprised not only have been making 7nm chips for a year already, they also make the fastest memory chips. Our response? Let's sanction them so they can't sell those memory chips.

Instead of believing in the fairy tale that we can somehow block the country with the highest number of STEM graduates out of technology, maybe what we ought to be doing is to invest a little less in the bullshit professions and a bit more in the hard tech, not just here but also in places like Japan where they had the potential to compete with ASML but dropped out because it wasn't profitable enough more than a decade ago.

[ksec]

>Just last year people were commenting here on HN that China is years away from making anything under 14nm and that even that would be hard for them.

Forget about the topic of China for a bit, 99.999% of HN comments on anything hardware are pretty much worthless. To the point I question how did software developers get so abstracted that they had little to zero basic understanding of hardware. When they should be considered in the same field.

If it is wasn't for the 0.00000001% of rare comments for those who actually work inside Intel, AMD, Lattice and ASML ( or some other SemiConductor Companies ) I would have skipped hardware topics on HN.

[bradknowles]

Note that China was actually one of the biggest customers for ASML, going back to 2006 and earlier.

The difference is that China didn't get the latest and greatest machines. Taiwan did, but not China.

But ASML definitely had sales people in China that were willing to do whatever it took to make the sale. Including giving hundreds of gigabytes of free pirate DVD rips to their prospective customers. Guess how long that took over the 2Mbps E-1 lines that the local offices in China had to get back to the HQ offices in Eindhoven, where those could then get out onto the public Internet? Guess who helped run that central mail system that was frequently flooded by all these DVD rips that were being sent by the sales people to their customers?

[tooltalk]

>> Note that China was actually one of the biggest customers for ASML, going back to 2006 and earlier.

Why would you make such an absurd comment about ASML's 2006 sales? or perhaps you mean to say 2016?

  2006
  Korea 1,085,497 13,730 662
  United States 931,971 740,036 24,262
  Taiwan 739,432 16,058 483
  Rest of Asia 470,915 937,107 1,282
  Europe 369,289 2,145,710 166,415
  Total 3,597,104 3,852,641 193,10

  2016
  Taiwan 2,140.3 2,815.9
  Korea 1,594.3 28.7
  United States 1,087.5 4,200.6
  China 758.2 2.6
  Singapore 245.6 0.8
  Japan 415.1 4.6
  Rest of Asia 26.7 2.8
  Netherlands 1.1 2,737.9
  EMEA 606.3 2.5

China's share in chips was a paltry low-single figure until few years ago -- less than 5% of global chip sales, according to SIA's report in 2022, and that also had a lot to do with foreign chip makers (eg, Samsung's Xi'an plant which opened in 2019 accounts for 40% of their entire DRAM production).

[bradknowles]

I know where the sales reps were, and where the customers were that they were selling to. Maybe those sales were to companies that weren't technically Chinese, even though they were operating in China. And so maybe the sales got reported for some other region.

China was huge business back then before 2006. Kept our E-1 lines quite flooded with all those pirated DVD images that they were sending by e-mail. But we did fix that problem. And I did get a nice little invited talk out of the work I did on their e-mail system.

And yes, I said 2006 and I meant 2006.

[markus_zhang]

If you properly sanction China it would be very difficult for them to make anything 28nm+. A lot of stuffs have to be imported from foreign countries so if US can convince JP/SK/EU to join the sanction it would be tough time for China.

Making hypersonic missiles or any military equipment is a different story as:

- You don't have to care about quality. You can make 100 chips, fail 90 and still have 10 to use, which is OK for military but would be a disaster for civilian;

- Worst case you can use espionage or diplomatic channels to find a small number of chips. You can't do that with mass market products.

[kurthr]

I think you're comparing prototype low volume/yield production to 1000s of wafers per month at >90% yield. Those aren't comparable.

Intel could do 7nm (they called it 10) with DUV immersion equipment and double/quad patterning, but the yield was low 90s even for their highly tuned processes. That's why EUV is so important.

Don't believe all the "news" hype, it's just agit-prop entertainment.

[pphysch]

Underestimate? Unless Washington aggressively nukes mainland China, recent overperformance by a post-Zero-COVID China suggests 10y is an overestimate.

Existential stress is an excellent motivator for rapid innovation. Or did we learn nothing from the 20th century?

[yywwbbn]

> overperformance

What do you mean?

[pphysch]

Growth, high tech innovation, etc.

China has been on the verge of collapse since 1991, according to Western popular consciousness.

[yywwbbn]

There is no need for exaggeration. But.. Well clearly the last couple of years were not great for China compared to 10 or so prior to covid.

[tooltalk]

China doesn't really have any chip-making equipment supply-chain of their own; neither does Taiwan or South Korea. I think that 10 years to EUV timeline is a bit too ambitious considering that it required ASML about 20-30 years to develope and commercialize (with support of the industry, TSMC, Samsung, and Intel). Unless ASML is allowed to license their EUV to China and all 600+ suppliers who make high-precision lenses/machinaries, cutting-edge light-source, electrochemicals, etc, can export their supplies without any restriction, it's still a pipedream.

[dirtyid]

10 years optimistic but PRC in position to create indigenous semi supply chain. After elevating semi to first-level dicipline in 2018, PRC is spitting out 30k IC graduates per year. They're still about 200k short, ~520k/720k out of what IC talent white paper estimates PRC needed for semi industry. I wadger that's comparable to total direct semi industry talent globally, which is really effort from a handful of countries (NL,US,JP,DE) in hardware industry that historically got 2nd pick on talent due to how well software pays. Would take less than 20-30 years to catchup especially if industrial policy mitigates commercialization concerns. It won't be easy but feasible especially if PRC indigenization starts taking domestic shares and take revenue/R&D stream western incubants who has to boost subsidy to plug loss.

[tooltalk]

There is a lot of wishful thinking there.

[hashtag-til]

Oh come on… you just broke the myth I had on my mind for many years, that this company was some magical craftwork shop for mechanical parts.

Anyway, nice to get a bit more realistic view. Thanks!

Eager to hear more about the software, which I heard some crazy stuff about how it is engineered and tested.

[nwiswell]

Sorry :)

I should not really say too much, but the software tends to be kind of a trainwreck. WFE companies are definitely engineering-first organizations, but software is a second-order concern in general since it is a check-the-box requirement rather than a truly differentiating product feature, and companies like Lam and AMAT which are headquartered in the Valley have serious trouble competing for software talent. Moreover the software has a multi-decade pedigree with a concordant amount of technical debt. Pretty much everything is C++.

Anyway, there are fairly well-specified standards issued by SEMI like SECS-II and GEM that govern interoperability. In my (highly biased) opinion, the really interesting stuff is the real-time control systems that operate on millisecond timescales to make minute adjustments on the fly using data from onboard sensors to improve process outcomes.

[phkahler]

I commented last week about floating point issues in a simple low pass filter. Someone responded confirming similar issues working on nanometer resolution servos. I just figure he's from ASML or similar. It seems someone from everywhere is here on HN.

[markus_zhang]

Ah, multiple decade C++ codebase! My dream. Do they hire juniors in Canada?

[mgfist]

The magic must be in the orchestration then, I'd suppose. Otherwise China wouldn't be struggling so much.

[nwiswell]

Partly orchestration, but mostly in the design of those proprietary parts. There's decades of institutional knowledge and engineering iterations involved here. IP is a huge concern because if a Chinese competitor got their hands on the engineering drawings, there would not be a whole lot to stop them from producing copies.

Patents are viewed somewhat skeptically, since if they are not actually enforceable against a likely offender, they simply serve as a disclosure mechanism.