Hacker News Reader

University labs (with the right funding) can totally do this, it's just not cheap. My university sold all its fab hardware to another university the year before I was able to take a VLSI class which at the time, had a practical lab. *

> As I read more about the dark art of IC fabrication

I want to push back on this being a "dark art" - there is no magic in engineering (nb4, any sufficiently advanced technology etc etc). It's a skillset that requires education, experience, and expertise on par with anything we do in other areas of engineering. The stakes are just a little higher than software because you're dealing with the physical world and physical things have tangible costs and/or danger.

The thing that may trip people up is that IC fabrication is one of those things that doesn't really have a hobbyist tier. Anything beyond a toy requires multiple people and support staff in addition to gear and raw materials that are hard to get as any old civilian - in addition to the clean room facilities. Like the reason my university closed their lab was partly because the grad/PhD students and professors had moved on, and partly because it was becoming more difficult to source wafers for research institutions that they could actually use (everyone got hired by labs in industry, where they were making their own wafers or buying them wholesale afaict).

* iirc only the penultimate project got taped out and fabbed with terrible yields due to time contraints

> I want to push back on this being a "dark art" - there is no magic in engineering (nb4, any sufficiently advanced technology etc etc). It's a skillset that requires education, experience, and expertise on par with anything we do in other areas of engineering. The stakes are just a little higher than software because you're dealing with the physical world and physical things have tangible costs and/or danger.

I think "engineering" in software generally means optimizing a path to a targeted set of behaviors so that the piles of garbage underneath don't end up blocking their execution for eternity.

Our starting point is therefore different. You ought to somehow be working around all the physical piles of dust and patchwork of fires that must be constantly igniting inside your laser machinery. I picture it something like the mad surgeon in Minority Report, creating a small transient sterile environment to do illegal eye surgery in a room full of filth.

In that light your "art" looks "dark."

It seems you have convinced me that IC fabrication is a dark art, despite your intentions.

I don't really know much about ic manufacturing.

Are you sure university labs are really able to to this? If so how come only a few companies like tsmc and that one Dutch company are able to manufacture microchips? Or are those two completely different things and I'm just confusing myself?

University students in Poland, under russian occupation no less, managed to clone and manufacture Intel 8080 using 6um Uni lab in 1982. Writeup in Polish http://retrokolekcja.pl/MCY7880.php

In 1983 cult Polish science education TV program SONDA documented design and manufacturing of first batches in a humorous lets bake a cake fashion. Paper plotters, light pens, developing/rinsing dies by hand, electron microscope debugging, the whole nine yards!

part 1 https://www.youtube.com/watch?v=AJGp7keIA_o

part 2 https://www.youtube.com/watch?v=KHl6m93Hay0

part 3 https://www.youtube.com/watch?v=rcOTwkT-PDU

> Are you sure university labs are really able to to this?

Yes, I know of multiple universities that have labs for small scale IC production. In fact anywhere doing research in the field will have some ability to build these things, or access to the industrial labs nearby. Even in industry, there are small scale labs that are used to develop the processes before they get built out at scale.

> If so how come only a few companies like tsmc and that one Dutch company are able to manufacture microchips?

There are thousands of chip manufacturers worldwide. TSMC is just the largest/most cutting edge. ASML is the company that makes special tools for IC manufacturing (however, researchers can/do experiment with the things that ASML is doing on smaller scales).

But keep in mind - no researcher at a university is trying to manufacture millions of 3nm CPUs for next year's iPhone. Just as an example, today we have GaN switches in our 100+W USB-C chargers that fit in your pocket. That directly came from university and industry research in small scale labs into high bandgap semiconductors, which was developed by fabbing real circuits and testing them.

The trillion-dollar-hard part is doing it profitably at scale. Drop that constraint and nearly any feature size is "only" million-dollar-hard (maybe 10M or 100M to run a R&D shop).

You can poke and prod anything into place with e-beams and FIBs and manually dipping wafers in baths and ovens and such. 1% yield, hour long write times, and all sorts of R&D jank are perfectly fine for checking functionality of your fancy ultra-FET design or making a ring oscillator to simulate integration. Did a grain of dust land on the wafer and ruin 100 of them? No prob, use the other 300, just try not to let it happen again. But integrating a billion transistors, coordinating them to do a billion calculations per second, QAing them to work for a billion seconds with 0 errors, and manufacturing them to profitably sell at $100 a pop? No jank allowed, no small scale antics allowed, and your budget now requires all the zeros it can find and more besides.

That's at the very highest end. As the element size gets larger there are more fabs capable of doing the work. The equipment gets slightly more standardized, etc., although ASML (the Dutch company) is still the big dog in the equipment space.

But even running a small-scale fab spitting out 7400 series chips and 555's is still pretty serious business; you need chemical engineers and material scientists as well as electrical engineers and software engineers (and multidisciplinary versions of those people) to keep things running at all. And nobody can do this stuff out of college -- everyone has extensive apprenticeships and practical experience working in other fabs because so much of the process is knowhow rather than technical specifications.

There is a wide gap between TSMC's cutting edge processes and what a university lab would produce. The features on the microchip go from a couple nanometers (TMSC cutting edge) to tens of micrometers (1000-10000x larger). Large size means less transistors, but million instead of billions still is plenty for large complex chips, just not cutting edge.

Yes, my alma mater has a nanofabrication lab on campus: https://www.rit.edu/facilities/semiconductor-nanofabrication...

They are even able to work with external clients to sell the chips they make.

ASML, that one Dutch company, is the only manufacturer of EUV photolithography machines, which are required to produce the cutting-edge of chips. There are plenty of chips that aren't cutting-edge, though, and plenty of reason to produce them in both academic and commercial settings.

Semiconductor Fabs have come a long way. The Shockley Semiconductor Laboratory opened for business in a small commercial lot in Mountain View in 1956. https://www.researchgate.net/figure/Shockley-Semiconductor-L...

There was lots of older or used equipment Universities could buy before Fabs started being millions of square feet with hundreds of million dollar pieces of equipment.

Yes indeed. My university had a clean room and research-scale fab equipment right next door to some of the lecture halls.

https://nanofab.usc.edu/

The key here is research scale. Larger process nodes, minimal automation, and smaller yields. Which is just fine, because the idea is to prototype new ideas rather than produce millions of chips.

that dutch company makes machines involved in creating the ICs, not the ICs themselves.

not even that, they really make the machines that make the patterns that are used to develop the electronic circuits on the ICs

I’m not sure that’s accurate: ASML don’t make masks (i.e. the patterns), they make the EUV photolithography machines that are used in conjunction with the masks.

The physical masks themselves are usually made by Hoya, and the technology to actually etch the masks is made by Veevo.

That’s like saying why can’t Toyota made a car that competes with Koenigsegg. One is on the absolute bleeding edge of everything and the other sells more cars than anyone else.

TSMC (and AMSL) are the bleeding edge of semiconductor manufacturing. There's a long tail of other semiconductor manufacturers that don't operate at that bleeding edge.

> oh so messy and built by trial and error

Not only built by trial and error, but also continuously adapted in near real time to deal with new sources of error.

The most complicated aspects of semiconductor manufacturing utilize statistical process control to determine the best course of action by relying on large sample sizes. You probably couldn't start up a modern manufacturing line without already having a manufacturing line due to this. Finding viable "hyperparameters" for a photo tool makes training an LLM look like a tutorial. Bootstrapping all of this required direct human involvement with ever-so-careful incremental offloading of these concerns to automation over a period of decades.

> Finding viable "hyperparameters" for a photo tool makes training an LLM look like a tutorial.

There's generally an unstated (and occasionally explicit, as in this case) reverence from software people for the kind of mythical engineering that goes on in fabs. In reality, if you've had any direct experience with the manufacturing process—and I'm talking about current- or next-gen processes for the most sophisticated mass market devices like those going into flagship smartphones, mining ASICs, GPUs, and critical applications like use in EVs—you know that a bunch of it is in the hands of folks whose most desirable asset in a prospective worker is that they'll accept low pay to eventually get the necessary work done to the prevailing standard best described as "adequate".

Valley types especially, but even other software folks would be really surprised by how much of what goes on in fabs is basically the sort of thing that you would expect to see from people plucked from amateur hour. I've posted about this before on HN. Where improvement to existing chipmaker operations is concerned, the fruit hangs so, so low.

Elon's biggest, dumbest misstep is not just buying Twitter; it's buying Twitter and not putting an equal or lesser amount of resources instead into gaining control over how his own (and others') chips are made—doing the same thing for the industry that he did with SpaceX for aerospace.

Again, because it cannot be emphasized enough: what passes for acceptable in fab operations is bonkers.

This sounds quite interesting.

Can you link to some of your previous comments on this subject?

Here's one: <https://news.ycombinator.com/item?id=31065392>

Training an LLM is a rudimentary exercise at this point, so maybe not the best example.

It's really expensive or difficult to have a one off object made though, and that's where 3D printing thrived. It fulfills that rapid prototyping itch.

People don't even really etch their own pcbs anymore, it's so fast and cheap, let alone spend $10k+ to manufacture a six cent item (maybe!), so there never was enough motivation for a diy movement to make ICs and other nanofabbed stuff

Especially when the goal is not the 6 cent item but rather the thing the six cent item makes possible.

I would whip out the credit card if I could make 555 timers at home for fun for $1,000.

Not sure if I put a second mortgage on my house to have a chance at maybe making one if I didn't screw up too much.

Nobody ever created a reliable self-contained foolproof PCB etching procedure. That's why nobody etches their own PCBs.

If there was a box that received supplies and outputted usable PCBs with minimum external mess, a lot of people that currently buy boards would use it instead.

(And well, PCB manufacturing is basically the same process as chip fabrication, without the miniaturization. If nobody managed to create a "PCB printer", why do people keep hopping for a "chip printer"?)

Etching your own PCBs has been a common electronics hobbyist activity for 50 years or more. Un-etched one-layer and two-layer PCBs were a standard stock item at every Radio Shack. Local electronics stores stock ferric chloride etchant.

My grandfather taught me how to do it 40 years ago.

Its true, I'm not even that old (late 30s) and used to etch tons of PCBs for ham radio projects.

> Nobody ever created a reliable self-contained foolproof PCB etching procedure

LPKF did, it's just expensive.

> That's why nobody etches their own PCBs.

Get into exotic laminates and ultra-high-speed performance with very tight tolerances and you'll see some prototyping done in-house.

Sure, that's the initial niche of 3D printing. But now people want to be able to repeatedly print something else again and again without having to do some sort of maintenance on the machine.

Now, you're getting those massive print farms that are able to change what they produce on the next print.

you're spot on!

clearly a big part of why all these tech has been so succesful is also how it's all about investing a lot up front, but eventually being able to mass produce in a ridiculous scale, few industries have such a ratio (possibly pharma?)

so it's all about making chips by the hundreds of thousands. it requires a very different approach from any tech intended to make chips by the handful

Thin-film transistor circuits can probably approach more of what you are envisioning than silicon integrated circuits. There are even organic semiconductor versions of TFTs that use lower temperatures and liquid chemistry for layer deposition.

there's no chance of DIY silicon fabs taking off, but the industry becoming more accessible to hobbyists is way more plausible

imho, the deeper problem is that there are just very few situations where you need a custom chip that can't be covered by existing options or FPGAs, and vanishingly few people have the expertise to get anything interesting done even if they had cheap access to fabs

(check out tiny tapeout, though!)

Sure, if one assumes people will focus on traditional CMOS design, and high-volume rapid fab processes.

Most people can't even make FPGA that work properly, and YOLO the simulation given metastability is beyond their users understanding.

>few people have the expertise to get anything interesting done even if they had cheap access to fabs

Chicken or egg problem... a walled garden simply gets fewer visitors, and people with expensive toys tend not to share. =3

> but surely we could get to ~10um.

Well, why not 100um then? It's still way better than discrete components.

I’m convinced this is the way to go. Rather than imitating commercial fab techniques, let’s find something that works without the toxic chemistry or vacuum chambers, even if it’s janky at first. 3D printers were janky at first too.

If you could fit even 9 logic gates into 1 square mm with a construction mechanism that scaled up to 2cm by 2cm you could build a rather capable 8-bit CPU.

I do wonder if taking this approach would work better with a novel construction method. Lithography and nasty chemicals are easier for resolution, but nasty chemicals.

On the other hand there will always be someone standing by to tell you an FPGA could have done that.

And someone next to them saying there's a chip that already does that, with someone standing next to them holding a commercial product that does the thing, and someone next to them tell everybody it's all a waste of time.

Someone's going to judge what you do regardless. As long as you're not hurting someone else, go build what you want to build, others be damned.

Totally agree. As creator of a lot of projects where people have asked "Why would you even do that". I wholly encourage build and be damned.

I've done

    Pacman in DCPU-16 asm
    A HTML/JS desktop environment
    8-bit AVR assembler in JavaScript
    An 8-bit fantasy console with web IDE.

and whatever the hell https://c50.fingswotidun.com/ is

plus a bunch of other things that really make people wonder if I have any sort of life plan at all.

100 um is 4mil, which is the resolution one can get from the cheapest PCB offerings. (e.g. for single digit $ from JLCPCB).

> 100 um is 4mil

Just checking, "4mil"? When I see (or hear) "mil" I assume millimeters, which clearly isn't right here, but I don't know if this is autocorrupt or if this is shorthand for something else I've never seen called this before, say "1e-4 meters"?

USAmericans will do anything to avoid using the metric system. To them, "mil" means "milli-inch".

A mil in manufacturing is 0.001 inches(in this case 100um =3.937mil).

mil is an unfortunate shorthand for one-thousandth of an inch widely used in the electronics industry.

mil in this context is 1/1000 inch

If anyone is interested in future small fabs: https://atomicsemi.com/ Looks very promising. Founder is an interesting person.

Can they beat https://www.minimalfab.com/update/news_en/244 along with https://www.minimalfab.com/update/news_en/226 & https://github.com/mineda-support/ICPS2023_5 -> https://mole99.uber.space/2024/NAND_tutorial/Design%20of%20a... -> https://mole99.uber.space/2023/Minimal_Fab/About%20Minimal%2... ?

When there is a need (remote space colonies for example), they might need to develop a more robust process that would trade off size and speed of chips for ease of manufacturing.

OTOH, remote space colonies get zero-g manufacturing, along with free vaccum so hard that makes our best artificial vaccum systems seem like a Florida garden during a hurricane in comparison.

What they get to do may not help with DIY in a garage on Earth.

Zero-g adds nothing. You cannot even purify silicon in zero-g, you need some-g for impurities to go up and down.

The average distance between molecules of the atmosphere is 3.3nm, this is about 10 times of the typical atom diameter. You need 1/1000 of the standard atmosphere pressure to make this distance ten times bigger. Which will pretty much be the hard vacuum at the scale of the atom manipulation.

> You cannot even purify silicon in zero-g, you need some-g for impurities to go up and down.

Only if you're using the purification technique that was developed for use on Earth that takes advantage of Earth's gravity.

There's other ways to purify silicon. Off the top of my head and not because it's necessarily a good idea even in zero-g*, there's the Calutron: https://en.wikipedia.org/wiki/Calutron

> The average distance between molecules of the atmosphere is 3.3nm, this is about 10 times of the typical atom diameter. You need 1/1000 of the standard atmosphere pressure to make this distance ten times bigger. Which will pretty much be the hard vacuum at the scale of the atom manipulation.

That's famously how gases differ from liquids and solids, yes.

I'm more pointing towards it being easier to control the doping of the semiconductors when you don't need to worry about the presence of oxygen (or water vapour), and that this is a very very clean "clean room" that you get for free without having to filter out the dust** or pollen because there wasn't any in the first place.

* it might be cost effective or not, I'm making no claim either way because I don't care enough to try and engineer something like this and then compare it to the alternatives

** depending on where you go in space, of course; I'm just saying you can pick a place without any, you're not obliged to do this e.g. next to an asteroid.

FWIW, there _is_ work being done on the hobby front for IC fabrication "at home". We're still far from buying a miniature chip-fab-in-a-box product, but current technology makes yesterday's tech far more affordable. We're on our way.

Example: https://www.youtube.com/watch?v=RuVS7MsQk4Y

We still need this, no matter how hard it is. If we can't make our own computers, we're stuck with the computers made by big companies and those come pwned right off the factory. We'll never be truly free unless we can make our own free computers at home, just like how we can make our own free software at home.

I would be happy if electronics companies started offering more dense circuits printed on film instead of thick 1mm PCB. There's too much volume wasted on tracks, that could be reduced layering discrete components and there is film that can isolate the heat.

You can get flexible printed circuits (FPC) from vendors like JLCPCB and PCBWay, which are essentially what you describe.

https://jlcpcb.com/blog/flex-pcb-available-at-jlcpcb-from-sp...

https://www.pcbway.com/fpc-rigid-flex-pcb/flex-pcb.html

And in case folks reading this don't already know it, multi-layer rigid printed circuit boards are a common technology based on laminating together multiple very thin rigid layers with each layer carrying separate traces.

FPCs have poor electrical characteristics generally.

How big is a spec of dust? What size can you do if you don’t worry about dust? 1000nm? Smaller?

Dust isn't just a single size, it's a giant range[1], and no, even 1 μm process requires a cleanroom, let alone anything smaller.

[1] https://www.engineeringtoolbox.com/particle-sizes-d_934.html

Engineering problems can be solved with engineering solutions, e.g. better material science that's not toxic (PLA is common now but it was an engineering marvel).

As long as there's a problem and there's money to be made, these things you mentioned can be solved.

Another issue is the insane cost for test/fan equipment. Not even for semiconductor stuff but even material testing.

Vacuum chambers and vacuum pumps....metrology microscopes....

Cost to grow single crystal boules is insane.

Etc....

Teaching rocks to think is not for the faint of heart.

you could always have a go a programming a FPGA - it's still cool and fairly cheap now

eg these guys are putting retro games on to FPGA https://www.analogue.co/developer

metrix had a laser etching rapid pcb prototyping laser in 2013 in seattle. trace routes down to 1mil

There are multiple approaches to easily making basic single-sided PCBs at home, but the rest is hard: Multi-layer PCBs, vias, through-hole plating, solder mask... Those are all things that even hobbyists need, but generally require annoying chemicals and multiple manual stages.

All these things have been done by hobbyists before, but I suppose doing all of this for a single PCB just isn't attractive.