It's a bit scary to think how much depends on the expertise of this company and their ability do produce those machines.
I recently got recommended the video "Imaging at ASML" [0] and it absolutely blew my mind. If you decide to watch it, don't ditch it before you see the animations of the light waves and the problems that need to be solved.
> expertise of this company and their ability do produce those machines
The core IP for EUV Lithography is still owned by the US DoE (LLNL, LBL, and Sandia), and they licensed it to Applied Materials a couple years ago. They even released an EUV Lithography etcher last year that doesn't compete with ASML's IP but complements and simplifies the process.
Applied Materials and ASML don't directly compete in the EUV Lithography space, but in a couple years there will be less dependence on ASML alone.
If you think this is cool, you should see the system responsible for moving the wafers between the tools.
In the industry, the robotic system that moves the lots through the factory is what makes all of the economics viable. If a human has to move wafers between tools manually, none of us would be using the devices we have today.
The AMHS is where all the magic comes together. It's really easy to get laser focused on one dimension, tool or metric, but that makes it hard to understand the entire game. Semiconductor mfg is all about a system of systems taken to the extreme and the robot that ties all the physical tools together is doing a ton of heavy lifting with regard to integration.
I also thought that seems implausibly large. My guess, based on some quick research, is that a large fraction of that is procedurally generated, not human-written. Here's one example:
> Formal verification tools allow those teams to be confident that their solution is complete and error-free, and the code itself is then automatically generated... With such a model-driven engineering (MDE) approach, software developers at ASML recently replaced half a million lines of code that had been built the conventional way.
as another sibling points out, this seems too high.
Their official website speaks about millions of lines of code [1]. They also published an article (2017) where they proudly say, they got rid of half a million lines of code [2]. A comment on HN from someone who might have insider info says 20M lines of C [3].
Would be great to get the sources from the video. Maybe they talked about compiled assembly?
I’ve talked to a fair share of insiders myself and general consensus is that the hardware basically sells itself, and customers approach the software stack as something they’re unfortunately stuck with.
ASML developed new software products over the years but it didn’t pan out and at this point customers are wary of anything being sold and would rather keep 20 year old inefficient but proven to work systems around. Think racks of control systems next to DUV machines that are officially years EOL but ASML is forced to sell today because it’s all a customer is willing to buy into.
I have 0 idea about the new software they failed to make, but let me guess:
-- Machine has to be connected to internet at all times for no functional reason whatsoever
-- Account has to be created for every employee to login to "cloud"
-- Software had "smart" in the name
-- Monthly subscription, the more accounts the higher the cost
-- Completely bugged and broken web interface at all times. Make sure the style and behavior of UX components is unlike anything else and broken.
-- Hosted through Azure with the most elaborate and complex infrastructure setup possible. At least 100 mb upload per served webpage and 1 GB of RAM useage at the client side.
-- Machine serves the website/API from the smallest possible microcontroller for no reason.
-- Webapp / client runs in Microsoft Edge only.
-- Tracking sent to ASML servers for "no reason"
-- Everybody at ASML working on the cloud software had to work on-site in Eindhoven for no reason whatsoever and was paid as little as possible by forming renumeration fixing cartels with other companies in the area.
Most customer fabs run airgapped because they contain some of the most valuable IP on the planet. So pretty much none of the above applies.
It’s mostly legacy hardware and code combined with ASML support contracts. ASML support is provided on-request remote access to the systems, or someone flies out if the customer is particularly paranoid.
ASML would kill for any tracking or diagnostics to be regularly send back to Eindhoven for R&D purposes, but it mostly doesn’t happen because machine diagnostics data is intertwined with the chip IP.
As Throwaway stated all fabs are offline. The office portions are online but moving data in and out of the fab is pure sneaker net.
You realize how much of a complete joke airport security is once you have gone through the security gauntlet to get into a fab.
Getting a laptop into the complex is a difficult. USB drives, cell phones, etc. are carefully controlled. Making a direct connection to a tool in the fab with anything is a major PITA.
and they can still get viruses and its a shit show. Fortunately most of the stuff runs on not windows or incredibly out of date versios of windows (3.1?)
This equates roughly to 1k developers writing net 1000 lines per day, every day for 5 years.
While it's unlikely, the numbers aren't actually as far off as I expected. Maybe they have 10k developers. Maybe the systems have been worked on for 20 years, maybe devs are only doing 100 lines a day. Lots of combinations come to the right order of magnitude here.
However, most engineers in teams of 1k-10k aren't this productive (not necessarily their fault), and much development is rewrites, so I suspect that the actual amount is about an order of magnitude lower.
What might make up the numbers in order for this to be the "headline" summary number, could be that only net lines were counted, or that they use some sort of monorepo where they have e.g. all of Python committed in, all of GCC/LLVM or other toolchains, etc. Add a few toolchains like that and you'll quickly make up the numbers I'm sure.
ASML have a defacto monopoly. These machines being this complicated is a giant plus for them.
edit: To give an example from another industry, Dave from EEVBlog has a nice teardown for a DLP projector from the upper price segment.
https://www.youtube.com/watch?v=O3eGqjV9yG0
You can argue that an important design criteria building this was often the pricetag of the end product.
I wish videos were to the point. I really wanted to know how the wafer is moved. But it goes on and on about history and what the machines are like and light sources.
True, but the comparison with old technology makes the video more interesting. The replacement for the hydraulic actuators and also the twin scanning system are motivated by the historical development.
Details were missing and it was surprising that 120 wafers per hour boils down to 1 every two minutes. I got further confused when an earlier x/y positioning was related to a later vertical positioning in context of magnetic levitation.
Its worth noting...there are sub systems upon subsystems at play here. I'll bet there are probably around 100 discrete computers within a modern lithography tool. It is more like (well...literally is) a network as opposed to a single computer.
Asianometry is really great. Some of the points in the DNA sequencing videos, an area I know well unlike chip fabbing, had very slight inaccuracies or maybe just incorrect focus, but they are so much better than other sources that I trust all his videos. The detail was fantastic and my harshest complaints were minor.
the brain like computer one was also quite off. however neuroscience is both vastly more complex and less accessible to the layman than photolithography, so i'm all too happy to let the Gell-Mann amnesia wash over me and enjoy some decent infotainment
I used to work for SemiTool (now Applied Materials) in high school and it's amazing how big each machine is. I was tasked with disassembling a SUV sized machine that was used to coat the wafer in copper. Only one of hundreds used in a fab line.
I'm distantly remembering old video about someone shouting at hard disk rack. Those machines are probably acting like a giant microphones when just moving air around is enough to have a too big effect on position measurements.
[Engineer monitors a EUV machine]
[Distant farting sound, machine instantly flashes yellow light for a brief moment]
Carl [shouts back]: Bob, for fsck sake, told you to stop with the beans!
[Machine throws red light and discards current wafer].
There is a paper from 10 years ago where someone used one for making 450nm lithography
https://www.semanticscholar.org/paper/High-resolution%2C-low...