How to build a $20B semiconductor fab
construction-physics.com
construction-physics.com
The topic for the last several days was on the CHIPS And Science Act and the new Semiconductor Fabs being built by TSMC and Intel in Phoenx, AZ.
It will be several years before the plants already being constructed will go into production but there is a whole ecosystem of current construction, education of the future workforce that will need to be hired in the future. Not to mention all the ancilliary companies that are needed to support these gigantic plants in the area.
The dollars from CHIPS Act are not only bringing in the manufacturing plants but will be essential to bring this lost capability back to the US in the scale needed both from an economic and national security perspective.
It was great listening to the show and the impact the CHIPS Act on people's lives already happening now and in the future.
For anyone interested the links to the specific shows are available as podcasts here.
There certainly is a lot of development there, but it’s not like a factory town or anything.
There’s an outdoor recreational shooting facility across the street. I can only assume that is a huge culture shock for anyone coming over from Taiwan.
The reason there is a chip fab activity in Phoenix goes back to 1949 [0] when Motorola built a lab there. In 1952, they started making semiconductors and eventually chips.
[0] https://go.gale.com/ps/i.do?id=GALE%7CA79561583&sid=googleSc...
[1] https://go.gale.com/ps/i.do?id=GALE%7CA79561583&sid=googleSc...
Tucson is more well known for optics, but more focused on space science than on use in silicon fabs.
> With Taiwanese transplants moving to north Phoenix in droves, Arizona officials — and a local baker — are working behind the scenes to make them feel welcome
BTW, about this topic, I always recall "A View To Kill" James Bond's (1985) movie [1]. The top hit in the soundtrack from Duran Duran [2] is also recommended and playing in radios even today. Seems like Intel passed the torch long time ago but don't forget to read the mantra book: "Only the Paranoid Survive" [3].
[1] https://en.wikipedia.org/wiki/A_View_to_a_Kill
[2] https://open.spotify.com/track/6I4snLrVOrJsLdd43isc27?si=7ba...
It is why outsourcing produced much more devastation than was promised and why onshoring will create much more work than expected.
The end result currently will likely be stagflation since like always politiciand do the wrong thing even when doing the right thing.
The issue without outsourcing is that the benefits are widespread (lower prices!) but the drawbacks are concentrated (factory town is now a hellhole). And our political system is incapable of redistributing correctly even though the net effect is highly positive.
The seminal study on the topic is the "China shock" paper from Autor et Al.:
... and outsourcing. The benefits are concentrated: profits captured by the companies doing the outsourcing. Sure, they may sometimes trickle down to the consumer, but the costs - the distributed drawbacks - are inferior quality of goods, elimination of local jobs, high ecological footprint, abusive business practices, lack of effective customer support. And the extra magic here is, it spreads direct responsibility over national borders, so it's near-impossible to hold anyone to account.
More recently, the drawbacks have been far more global in nature.
It seems like Intel is skipping ASML EUV entirely.
If that's the case, I'm trying to understand how Intel ever gets decent yields at 7nm to 5nm.
It's definitely not coming from High-NA, which seems like a short-term distraction.
any pertinent examples? r.g. schools for workers families ?
Just for the construction work alone, they mentioned that the pipefitters local union membership has doubled since 2020. Refinery level complexity on the specialized piping needs for the plants.
Special training programs geared towards the semiconductor industry being offered in the local Community and Trade schools training people to be the skilled and semi-skilled workforce for these companies.
People who were teachers now making four times the income working on the construction project.
Targeted investment! Toward future generations of chips, rather than of people.
BTW, having spent 35 years in the semiconductor industry, this is the best 'layman's' description of what goes into chip fabrication I have seen. A perfect blend of description and illustration without over simplifying.
I’d also highly recommend the book Chip Wars by Chris Miller for anyone looking for a deeper dive into the history and current dilemmas of the semiconductor industry
https://youtube.com/watch?v=IS5ycm7VfXg
It’s a thoroughly interesting video, but I’m a bit disappointed he never took his idea any further than he did. I’d really love to see something like a 6502 being made at home.
[0] https://atomicsemi.com/about/#:~:text=general%20fabrication%...
From DEC, to AMD, to ARM and Broadcom, his own firm, hops over to Apple which then buys his old firm, heads back to ARM and then over to Tesla and finally one last stop at Intel before going into startup land again.
Worked on the K7/K8, MIPs for networking, did the A4 and A5 for Apple, on the Zen/K12, and the Tesla TPU.
So that fairy is completely useless in this context.
Can you do it, given the right tools, training, and patience? Yes.
Will they be any good? No.
Will they be cost-competitive? No.
Getting chips made on a shuttle run for an old node is very affordable. There's really no need for it.
(According to one MPW supplier, 10mm2 of 340nm, up to 10 dies, costs 6400 euros, and it's unlikely 340nm is achievable in a garage anyway)
I believe the way Sam Zeloof circumvents the enormous amount of capital needed for a chip fab by relying on modern technology to create 1970’s technology. He simply mounts a cheap digital projector onto a cheap microscope - they didn’t have that advantage in the 70s, and thus it cost millions to start a chip fab. My point is that it could conceivably be doable for an individual to create old computing technology with the advantages of living in the modern world. I certainly don’t have the drive to do it, but I wish someone did.
If you did it often and didn't count your own labor costs, then maybe the average cost would be less, but that's an incredibly specific situation.
> I believe the way Sam Zeloof circumvents the enormous amount of capital needed for a chip fab by relying on modern technology to create 1970’s technology
Yes, exactly.
Old lithographic technology is so crude that you can even use modern high resolution laserjets to print masks (10000 dpi is less than 3 microns).
Even so, 1970s-era CVD, PVD, and plasma etch is still quite complicated, and CMP is impossible (it hadn't even been invented yet). So the devices you can create are significantly integration-constrained.
This isn't hypothetical, I've done it -- in grad school we would send out (I believe) 30000 dpi print jobs on transparent polyester film, and then adhere those to glass blanks to create cheap masks for MEMS fabrication. We had an old Canon i-line lithographic aligner that accepted the glass blanks.
I think the print jobs cost us about $100 each.
Here's the first Google result for a vendor (I don't remember who we used). There's a price list on their page and it looks like they have capability up to 50,800 dpi.
https://hacker-fab.gitbook.io/hacker-fab-space/fab-toolkit/p...
This yes can be priceless in some circumstances.
Industrially (by which I mean how it was done circa 1970), silicon oxide and silicon nitride was etched using a buffered HF solution known as BOE (buffered oxide etchant). The buffer was typically ammonium fluoride; because of the presence of the buffer, the concentration of fluorine ions in solution stays constant even as some of the fluorine attacks the substrate to form e.g. hexafluorosilicic acid. Since the concentration of fluorine stays constant, so does the etch rate.
If you just pull some rust cleaner off the shelf at home depot, the etch rate will crash as the concentration of fluorine ions decreases. That's compounded by the fact that the HF concentration isn't very high in the first place.
As a result it would be very difficult to determine how long your wafer should remain in the etch bath. Underetching could easily cause "opens" in the circuits from unremoved insulator, and overetching and/or undercut can destroy the patterns you're trying to produce. Either way it can ruin the chip.
Ammonium fluoride definitely isn’t as easily accessible as rust cleaner, but you could buy it for a somewhat cheap price on Amazon.
Thus only a very small number of companies (currently TSMC, Samsung, and Intel) attempt to operate leading-edge nodes, and the industry has shifted to a “fabless” model where companies like Apple and Nvidia design their chips but have them manufactured by “foundries” like TSMC. By pooling the orders of many different chip companies, the foundries can achieve the scale necessary to afford cutting edge fabs.
There's a good reason why the latest gpt-4 competitor is meta or Google.
> (There is a Moore’s Second Law, also known as Rock’s Law, which posits that the cost of a semiconductor fab doubles every four years.)
If this were to hold, then in under 30 years a single fab would cost more than three trillion dollars, which itself implies a hard upper bound on node improvements by way of economic considerations.
30 years ago, Microsoft matketcap was around the same 20bi. Today it's 3T.
It would totally make sense for it to happen again.
How much?
Personally I only believe one of those things. I also believe the point about additional cost of development could have been made without the value judgement.
"I think the Asian engineers tend to be more methodical, they tend to be more studious, more orderly engineers; whereas, the U.S. engineers tend to be more innovative, but they tend not to be as methodical and orderly as the Asian engineers. ......
On the other hand, I also think that one group can accompany another group, you know. You can have an innovative group but not so methodical and so on, and then you can have a methodical group maybe not very innovative accompanying each other. "
https://archive.computerhistory.org/resources/access/text/20...
Is Intel buying any EUV machines from ASML. Not asking about High-NA machines, but regular EUV machines?
I can't find the answer online although I do see that the New Ireland Fab seems to be in the regular EUV range.
The sheer complexity and costs means there's no real possibility of a competitor. Nor would a billion dollar fab want to experiment with a new company.
Fab 34 in Ireland got its first ASML EUV machine two years ago.
I don't believe High NA EUV machines will end up in this generation of fabs. The first one was shipped to Intel's dev fab a few months back.
Intel 14A will use high-NA EUV.
Step 2. Build the $20b semiconductor fab
A tolerance of 1/8 millimeter (5 mils) can certainly be achieved by CNC machining, but not always by the cheapest tools.
You mean that it is easy to achieve much smaller tolerances than this by CNC machining?
Googling for "CNC machining tolerances" finds at first hit a guide that says "For CNC machining, the standard tolerance limit is set around +/-.005” (0.127 mm)."
The next hits show the same 5 mils standard tolerance, but they also mention the availability of better tolerances, like 2 mils or even less, but at higher costs and subject to various constraints on the features to which they are applicable.
So it appears that the tolerance written in the article is indeed typical, as claimed.
> Similarly, a fab will use very large amounts of ultrapure water for wafer cleaning and CMP, along with the regular water for things like chillers for process cooling. A large fab can use millions of gallons of ultrapure water a day, as much as a town of 50,000 people, and producing it requires its own specialized plant.
This is wild. All of this water so companies can create chips that will ultimately be used to …
pump out “advanced” chat bots.
I really hope all of this sacrifice is worth it in the end. Climate change is accelerating the loss of drinkable water around the planet.
If the best we could do is a slightly better chat bot, then we are doomed.
The big scary sounding number at the input has a big number at the output.
Not to mention that up to 98 percent of the water is reused on site. It just gets cleaned and goes back into facility. It’s a big loop, not an input disappearing into a parallel universe.
This is always so funny to me. Oh, something that can make enough chips for millions/billions of devices also uses as much water as a small town? That sounds perfectly reasonable. Don't build it in the middle of the desert, I guess, but otherwise it's not a problem.
Ignoring Alphafold, moderna partnering with openai, the new class of antibiotics, etc etc. There's a lot more to ai than chat bots. That's a disingenuous reduction.
I agree that it's a bold bet though, burning how many ever billions a year on the hope that ai can help us solve medicine and fusion and climate change.