[r-bryan]

This scenario is very reminiscent of the leadup to the Innovator's Dilemma. The dinosaurs grow taller and fatter seeking and thriving on the juicy foliage higher up the tree, while the scrappy little mammals claw away at its soft underbelly.

Nobody's going to fabricate semiconductors in their garage. But couldn't someone make a modest profit without first investing $2.0e10, like Intel? Or is it just that the profitability scale is such that if you had an extra $2.0e7 to invest, you're better off buying Intel stock?

[rossdavidh]

1) technically, someone is fabricating semiconductors in their garage, but as regards the world economy your point is still valid: https://hackaday.com/2021/06/29/garage-semiconductor-fab-get...

2) the equipment used in older fabs, is in many cases not even made any more. It's much like why the vinyl record manufacturing took a decade or more to ramp back up when the demand for vinyl records began going up instead of down. Nobody was making that equipment any more. The older fabs don't use the same equipment as the new fabs, except less of it; they use quite different equipment, and in some cases newer equipment actually wouldn't be able to make the older chips. I don't work in semiconductors, but I did from 1989 to 2004, and worked on several cases of trying to move old processes into newer fabs. It's not easy or quick, and in some cases you have to move the old fabs' equipment into the new fab.

[midjji]

Something is very weird with intel. Compared to fang pe around 30 or nvidia, amd around 100, intel has 10. Which combined with long term stable decently high profits is insane for an am tech company. Intel has been offered gigantic subsidies to build american fabs, and more recently just fabs period, but they arent, and at least one of these is approaching the point where they might get sued for not having even tried.

The answers I hear range from a deeply embarrassing and complete loss of competency on all levels, but primarily on the fab side, to an utterly incompetent and out of touch leadership. I favor the latter explanation as intel has been underperforming for nearly two decades now. My best guess is that intel has entered its cash cow phase, and will simply phase out product after product as they become obsolete, ending up as a half dead zombie barely supporting a few rare products in a few decades. Problem is, even if thats what is happening, their stock is surprisingly cheap. So it seems analysts seem to know something important I am completely missing.

Not to say they are the only company that fucked up, amd has been offered the same subsidies for building more fabs, and we arent talking 2020, we are talking for more than a decade, and they also havent for very unclear reasons. Amd also fucked up and completely missed machinelearning due to being unwilling to make a convenient interface or even just halfassed support for cuda. But they did that while nvidia went, ah fuck it, well let amd get every major console, and basically only make gpus with graphics ports as a side project.

Amd and nvidia is havent acted optimally and both have failed to exploit what seems like low hanging fruit, but they are highly valued in a way which matches market demand (which has been increasing quickly since 2014, not just 2020), and their improving technology. Intel however, does not even appear to have tried. Performance improvements have been incremental at best, and some "features" like the built in gpus are so terrible they mostly degrade performance.

[bsder]

> I favor the latter explanation as intel has been underperforming for nearly two decades now.

Remember "Only the paranoid survive"?

Well, people steeped in Intel took that to heart. So, when they retired or died, a lot of them had no heir apparent and their knowledge went with them.

[sliken]

No argument on the fab side, but that's easy to outsource. Samsung, TSMC, and Global Foundries will happily make Intel chips.

Alder lake is a good example, from what I can tell it's a market leading core, generally faster than the AMD competition, and a good basis to make future products. Future Intel GPUs seem likely to compete on price/perf, at least when any GPU from the last 3 generations seems to sell out instantly.

[anamax]

> Samsung, TSMC, and Global Foundries will happily make Intel chips.

And how long will it take to "port" those designs to another fab?

It's not a matter of "make install."

Yes, everything comes together with tools, but every one of those tools and its inputs is verging on full-custom and those designs were tweaked with the characteristics of those tools in mind.

The folks who know how to do that work are currently working on new products (either for Intel or someone else), so are you willing to slip those products?

[sliken]

Quite a bit faster than building a fab for a leading process and making a new product on that fab.

Alder lake seems plenty competitive (even market leading) for today, AMD has nothing big (like zen4) due anytime soon. Moving the next gen to some other fab in time to replace alder lake looks quite feasible.

[anamax]

Yes, Intel can do new designs for other fabs.

The question is when it makes sense to port old designs to other fabs.

FWIW, I'm not convinced that it takes less time to bring up a new design on a new fab than it takes to bring up an old design on a new fab. (The exception being a new fab that is identical to the old fab.)

[riskneutral]

I don't think that you can build a new semiconductor factory for $20 million (2e7), especially not when there is a chip shortage happening and you need factory equipment containing such chips.

[bluGill]

There is the volume vs cost problem. You could put a fab together for under a million - but to do so your would trade labor for automation. A simple chip that a large fab can make for a profit (amortized) selling for 5 cents each would cost you more than $50,000 each to make. If you only need one chip that price is worth it, but most users of a chip are thinking a lot more and it typically turns out that getting a large fab to do your part is not only cheaper, but faster. Still you can do the highly manual process of making a chip if that is what you want to pay for.

[jfoutz]

It's a little weird to me. I see in the comments 50 cent chips going for 75 dollars. The old process chips _seem_ diy able - http://sam.zeloof.xyz/first-ic/ Obviously it would take me, like, a year or two to get anything to work. But I'd guess there are folks out there that could spin up far far faster.

I guess I don't understand why there isn't much, or any, labor intensive fabs combining on line, with a plan to automate away parts of the line as money comes in.

I suppose, the chemicals are so toxic, and the skills are so rare you can't get any production for months, even at the small scale. and perhaps even at the crazy high prices it's still not worth taking on the risk.

I know I don't know what I'm talking about. But golly it does _seem_ like I could make chips in a couple of years, and there's bound to be thousands of people out there that could do it in weeks. It seems like the kind of thing you could scale out with more labor and training, and improve process with automation as you go.

edit

Most trying to understand. Is it too toxic? are the skills so rare? Is the risk that high? Is that what's keeping low end chips from using a more expensive process even though they command higher prices - it's not enough to offset?

[11thEarlOfMar]

The $0.50 for $75 is a short term situation that will resolve once the fab's pipelines get filled again. We buy >1,000,000 per year and had no trouble getting them until the fab shut down. We'll be fine soon (looks like they've opened production again).

The primary costs of starting up a fab are 1. facilities, 2. equipment. Facility per square foot is likely the most expensive you'll find due to the appurtenant systems required for air flow and purity, a variety of gas handling systems, ridiculous electrical power systems, possibly gas plant systems, water purification and then treatment systems for both air and fluids exiting the plant. Some truly scary toxins are used and produced and have to be contained or captured and neutralized.

In terms of equipment, you'll need at least one of each of 40 types. Lowest cost is about $500,000. Highest is likely $25,000,000 (at the geometry you're talking about). Average in the area of $3,000,000 - $5,000,000. You'll need a cadre of engineers and scientists to keep it all running, and 2-3 years from breaking ground to producing product. If you did buy one of each tool type, you'd have most of them sitting around doing nothing most of the time, since the process times vary dramatically. In reality, you're looking at $500,000,000 minimum to build a fab that can produce $0.47 chips. Seriously, you don't want to do that.

[jfoutz]

Thank you. That makes a lot of sense.

I had some misconceptions around how versitile auto manufacturers are. I'd sort of assumed they'd be very good at turning a low volume, low quality process into a high volume, high quality process. I'd also sort of assumed they'd have a ton of resources on hand, like space and power, the and the ability to build a lot of the required infrastructure.

I suppose they are far far more specialized and reallocating resources like that is the kind of thing you'd have to do right away, make a big commitment, that's basically pointless.

[bluGill]

They are experts in iron, and processes, but they still would have a lot to learn, and it is unlikely they could ever do better than TMSC who isn't stupid about manufacturing. Maybe they could find something, but so can the others.

The commitment just to play is very large, it probably isn't worth it. Even if you had proof of this would happen in 2015 and so enough time to build I doubt any would build s fab (though they would buy chips to store)

[bluGill]

Sure old process can be done,but the labor is high enough that you need to sell for more than 75. For simple stuff you still need well over 1000 each to make it worth it.. that is simple stuff. I know a few years back the government was paying around 60000 for Pentium level complexity chips (this was for a classified program so I doubt I was given full details)

[11thEarlOfMar]

State of the art fabs retail for US$10 Billion. TSMC, Samsung, Intel & China can afford to build them.

(for context, Freedom Tower cost US$2 Billion.)

[baq]

I think in 2022 it’s more like 15B. 10 sounds very early-2020.

[amelius]

Perhaps you could do it at a 1000nm node, at very low volumes.

[analognoise]

EBeam lithography will get you sub 10nm, but it's one at a time.

[amelius]

This is close to how they make the mask-sets, very slow.