107 comments

[ 1.7 ms ] story [ 156 ms ] thread
I'm still seeing lead times til the end of the year for parts I've been wanting to buy for two years. Don't know how long stock will last with so much pent up demand.
tl;dr:

    ... the most problematic shortages will begin to ease in the third or
    fourth quarter of 2021, though it could take much of 2022 for the
    resulting chips to work their way through the supply chain to products.
    The supply relief will not be coming from the big, national investments
    in the works right now by South Korea, the United States, and Europe but
    from older chip fabs and foundries running processes far from the cutting
    edge and on comparatively small silicon wafers.

    ... Despite the auto industry’s desperation, there’s no great rush to
    build new 200-mm fabs. ... More than 40 companies will increase capacity
    by more than 750,000 wafers-per-month from the beginning of 2020 to the
    end of 2022. The long-term trend to the end of 2024 is for a 17 percent
    increase in capacity for 200-mm facilities. Spending on equipment for
    these fabs is set to rise to $4.6 billion in 2021 after crossing the
    $3-billion mark in 2020 for the first time in years, SEMI says. But
    then spending will drop back to $4 billion in 2022. In comparison,
    spending to equip 300-mm fabs is expected to hit $78-billion in 2021.
The only thing I can derive from this with confidence is that the chip shortage will continue until 2024.
Shortage for automative-grade chips on ancient processes will continue. It is exacerbated by the difficulty of mid-cycle changes.
Could these designs just be reworked for modern fabs?
It would likely require redesigning all of the related circuits where the chips are to be integrated. that’s a lot of work and would require more new chips.
Tesla did it at the beginning of the shortage, so it’s possible.
Tesla also just shipped cars with missing parts!
Tesla also does not build vehicles like other mfgs. It’s far more a computer on wheels than any mfg.

Saying “Tesla redesigned away from the chip shortage” is massively dismissing the challenges that they faced vs the other companies.

Not without a pretty high redesign cost and those products usually have really low margins.
From what I've heard, the main reason they use old processes is the wider voltage tolerance/lower sensitivity compared to more modern processes. I'm not competent at electronics though, could be totally wrong!
Should be golden times right now for algorithm and compiler makers. Get more out of the silicon we have and can make. The world doesn’t care if it’s 27nm or 9nm for most use cases. Needs customers to bust the hype cycles or else we remain at mercy of the chip makers and the other hype cycles.

There isn’t actually a chip shortage methinks - more like a shortage of intelligence and common sense.

Meanwhile electron continues to ship
you know there are integrated circuits which aren't CPUs or microcontrollers.

it's tough to program your way out of needing a regulator or amplifier or something.

No it's fine. We'll just throw the raw data stream at an AI.
I apologize, as an AI language model I can’t catch the raw data stream as I have no hands.
Input: 230V AC ; Output: As a large logic module produced at OpenIC, I am not allowed to assist you with dangerous bzzz (smoke rises)
The worst parts of the chip shortage are not affecting the newer modern lines.

It is the older lines, making voltage regulators and power management chips, making it difficult to put the whole solution together.

While some of these chips could be redesigned to use new processes, or older designs with larger format fabs brought up to accommodate demand, both of these situations are short-term and thus no one wants to invest.

(comment deleted)
Microcontrollers are often under-utilized, but that's just because their task is simple. The problem is that current designs use a lot of them: One or more per module. This could probably be reduced by a redesign, but then they could probably also shift to more available chips.
Then you likely need more different chips across your line. More different production lines which can stop you. If you standardize on fewer different parts you can buffer some fluctuation in the supply chain.
Some of the recent AI model advances, such as "Llama.cpp 30B runs with only 6GB of RAM now" are a pretty clear indicator to me that we might already have sufficient hardware for the AGI takeoff sitting under everyones' desks right now.
Consider that many of these chips are in things like automobile ECUs, aircraft parts, industrial controllers, etc. Changing them requires pretty rigorous testing for timing issues, operating temp ranges, considerable design and test to meet regulatory requirements, etc. Different CPU probably also cascades to needed changes for parts around the CPU, software, and so on.

Then, for any entity that successfully moves from one supplier with one process size to a different one...they are creating more demand there, which could precipitate shortages now in that space. Then, there's peripheral effects. Like perhaps some of the suppliers are liking the new per-unit margins during the shortage, and won't just let it all go when demand/supply normalizes.

Right? What’s more, we are now using warehouses of GPUs to sort out LLM completions, a lot of which is Python code, the slowest of languages. I mean, it seems like we’ve finally admitted even Python is too hard, so we should instead ask Deep Mind to spend a millennia computing the answers to our questions. And then when it tells us “42”, we need an even bigger computer to understand it, because we’ve outsourced understanding to the computers.

We are just coming off the highs of using GPUs to generate mountains of useless money that was used to effect massive scams. But even all that is still going on.

And free market capitalism is supposed to result in efficient resource allocation? What are we even doing, trying to accelerate the heat death of the universe? Entropy not increasing fast enough for us?

Like… maybe we wouldn’t need so much silicon if we were using it a little better. Might as well just store everything in linked lists. Who even cares about data structures anymore?

LLMs are about 0.5% Python. The hot path is all C/C++, or native code running on the GPU. Python is almost just a configuration language to connect it all together.
What I'm saying is that the completions are Python code, which is then run through a Python interpreter. I keep reading stories about how this is the future of development, meaning the number of transistors to get 1+1 executing on a computer is astronomical today.

Used to be that you could program a computer with everything in your head. Then you had to use the Internet, a global network of information, as a resource because it got so complicated. Now it's so bad that you also need an artificial intelligence that can interpret the information on the Internet. All so we can, what... serve more ads? Invade more privacy? Wage more wars? Earn more profits? What are we even doing with this technology?

> What I'm saying is that the completions are Python code, which is then run through a Python interpreter.

Even that is incorrect, or, to be precise, correct that it is indeed Python - however it is that part of Python which is actually written in C/C++ and highly optimized.

There is a reason why Python is used for data science, ML and similar. It is both convenient to use and fast, thanks to numpy and other libraries.

I realize that the point of your post is something else entirely, but no need to pick on Python. It's actually quite performant in areas where this is needed.

This is only true if Python is acting as a useful glue between things. The only reason it's a useful glue is because it's ostensibly easier to write. If we're writing the Python code with the LLM, then Python is obsolete. A better thing to do would be to just emit a binary executable that directly calls whatever machine code.

But no, here we are with the worst of all worlds. Maybe that day will come, but the long-term, industry-wide trend is more silicon needed to accomplish roughly the same end goal. We're playing the same games, writing the same excel docs, browsing the same webpages, but now it takes more energy, more silicon, more transistors... call me skeptical we're going to add all this LLM tech on top of the already towering tech stack, and it will result in a reduction of needed silicon because we've simplified the whole thing considerably.

> It is both convenient to use and fast, thanks to numpy and other libraries. I realize that the point of your post is something else entirely, but no need to pick on Python.

My point though, was that the fast and convenient is only interesting when it comes together as a package. Other languages are faster, even granting that some Python packages are written in native code. But if the new interface to Python is an LLM, then "Python" as a language doesn't have a place anymore; it'll be more convenient to emit code in other languages.

In Australia I keep seeing "Chip Shortage" signs on windows of restaurants, but its a shortage of potatoes leading to higher prices or more defects on your hot chips. Relief is expected in a month or so https://www.abc.net.au/news/rural/2023-03-20/hot-chips-short... but may repeat for another year. Like the other chips you never thought it would happen to such a fundamental product and then it did.
Lets call them spuds (and the silicon things: ICs) to disambiguate. The fried potato disks can be called crisps.
Japan is currently suffering from an egg shortage, thanks to the avian flu epidemic.
In America we are suffering from an egg shortage due to the greed of egg suppliers

https://www.cnbc.com/2023/01/23/high-egg-prices-due-to-a-col...

(comment deleted)
The article you linked to suggests otherwise:

> However, food economists are skeptical an inquiry would uncover wrongdoing.

> “I don’t think we’ve seen anything that makes us think that there’s something there other than normal economics happening right now,” said Amy Smith, vice president at Advanced Economic Solutions.

> “I don’t think we’ve seen anything that makes us think that there’s something there other than normal economics happening right now,”

When profits are suddenly up 700% this claim no longer holds water.

> High egg prices send profits at largest US producer soaring more than 700%

https://www.cnn.com/2023/03/29/business/egg-profits-cal-main...

You need sustained higher profit margins to make that claim.

In the short term, I expect profit margins to increase for goods experiencing supply / demand shocks due to market participants trying to figure out which way supply and demand curves are moving, and in shortage scenarios, buyers willing to pay more than necessary to secure the goods.

Sorry, but a company that reports it has had zero avian flu outbreaks in it's facilities and still jacks up it's prices so high that profits increase by 700%... that's textbook price gouging.
So buyers willing and able to pay more for something should not have the ability to bid more to guarantee they can obtain it?

Note that the buyers have options to consume other foods for their nutritional needs.

Buyers didn't have a choice, since suppliers that didn't have to cull their flocks chose to price gouge.
The alternative to price gouging is empty shelves with no option to buy at any price.

The high prices work to incentivize competitors to enter the market and increase the supply. If only the government would stop letting competitors merge together...

> The alternative to price gouging is empty shelves with no option to buy at any price.

No, the alternative to price gouging is to maintain a normal profit margin since your flocks didn't have to be culled and your costs have not gone up.

If anything, the flocks that had to be culled would have caused a sudden chicken feed surplus and their feed cost should have fallen.

Cost of goods sold only sets a minimum price.

If there are fewer eggs, then there are fewer eggs. The only thing left to discuss is who has the right to obtain them, and who will not. Usually, society accepts that people with more money get more access to things than people with less money.

> Cost of goods sold only sets a minimum price.

Colluding to price gouge consumers when costs have not increased is illegal behavior.

Who is colluding? And price increases in the US are not restricted to only when cost of goods sold increases. Any seller can change the price they want to sell at at any time, unless there is an emergency such as natural disaster where price gouging laws come into effect.
The issue being highlighted is that they suspect suppliers are colluding the drive prices higher. In a collusive monopoly the number of competitors artificially decreases.
Isn't price gouging where possible normal economics?
A market research firm for suppliers doesn’t see issue with what their paying customers are doing? That’s not convincing, even if it provides a second opinion in a well balanced article. In the past most journalism showed perspectives from both sides, so all articles would suggest otherwise for the set of all things otherwise.
Taiwan and other asian countries / cities too.
This article is old:

> This post was corrected on 30 June to clarify historic 200-mm fab equipment spending.

> This article appears in the August 2021 print issue as “How and When the Chip Shortage Will End.”

This article needs (2021).

Only the top few paragraphs are an update, the remainder of the content is "29 June 2021".

TSMC has been badgering their customers to move to 28nm for the past year. They've even been building out more 193nm capacity usable down to N12 because they can do it without waiting on ASML (Nikon can provide 193nm immersion lithography equipment) since ASML is where all the EUV advanced node supply chain holdup actually is. For everything that's not EUV probably 2024 is when wafer starts stop being insane.

The rest of us waiting for high end processes to expand in availability and drop in price? ASML basically has an infinite waitlist at this point.

Is it possible that ASMLs supply chains get broken, and they stop being able to deliver new machines indefinitely? I worry we might be at Peak IC.
The last few years ASML's problem was not supply chains, it was a very high demand. Production was ramped up successfully, but not enough to meet that demand. With regards to Peak IC: atoms aren't getting smaller, Feynmans' "Plenty of room at the bottom" might indeed run it's end.
If you chronically can’t meet demand for your critical patented technology, should you be forced to license on terms that allow demand to be met?
The issue isn't patent licensing. If it was, we'd see something competitive coming out of China.
Not for something like ASML which is bleeding edge manufacturing tech that only one company in the world has figured out. It’s very very difficult to replicate that research unless you steal it and replicate all of the surrounding work. Stealing is the easy part and I’m sure China has done that and is attempting replication. But there’s a lot more involved to it I think that’s tricky, troubleshooting and solving bugs is harder when you haven’t internalized how things work (you’re basically rediscovering all the things ASML has learned producing and integrating their tech into a finished product).
That's exactly what I mean. Patents requires publishing so that it could be replicated.
Knowing the underlying principles is only a starting point. The hard work is going from that to production.

Sounds simple, costs millions.

A nation state like China is developed enough that it doesn’t care about millions or even billions for something as strategic as this. They’re looking to potentially invade Taiwan to take over TSMC to get their hands on bleeding edge tech which is going to be significant more difficult and expensive.
East Germany tried the same strategy during the Cold war
Companies know this and a) intentionally make the parents vague (also helps with enforcement) and b) keep a huge amount back as trade secrets. That’s why you have industrial espionage.
Patents don't provide enough information to replicate an invention generally. Especially for complex industrial processes.
The principal goes that no, your prices rise which throttles demand until the two meet, then with the extra money you've charged, you build out new production capacity. That's the idea of free market anyway.
That assumes supply chains still exist that can provide all of the necessary components.
Nah, since US has banned China there's definitely less of a demand.
The irony is there may already be a glut of 28nm capacity. But that embedded chip you're waiting for isn't fabbed at 28nm, so everybody loses.
It's starting to happen that high-end embedded is running at 28 nm. Mainstream microcontrollers are still at 40-90, but their next revisions may be aimed at 28.
If you can figure out some clever innovative way around this to make the shortage irrelevant, you'd make a killing right now.

I know this sounds obvious but I hadn't thought of it before

Not possible, computing demand is skyrocketing. We’ll be in demand shock for the foreseeable future honestly
Brilliant ideas often come from people too dumb to know something's impossible
Just make a huge grid of 4 bit input/4 bit output LUTs, clock them in alternating phases like a chess/checkerboard... you'll never face a race condition. You can route around bad LUTs, and there are a lot of other advantages to this otherwise colossal waste of transistors.

I've had that dumb/good? idea stuck in my brain since reading George Gilder back in 1981.

You're describing systolic array processing, circa 1979. Not a new idea. Awfully inefficient in power and space usage. "Those that fail to learn from history are doomed to repeat it."
Just as neural networks were well known to be incapable of handling real world problems, which was part of the AI winter.

Until they just worked thanks to backprop and deep nets

Isn't this just an FPGA with more disadvantages?
It is an FPGA without routing issues due to inhomogeneities. Routing in a bitgrid is almost trivial.
The chip shortage is here to stay. Those margins are too good to let go. Why keep inventory when you can keep a wait-list.
Thats a dangerous game to play too lomg when China will happily provide older nodes capacity and make people switch to different lines of product, buyers might not be coming back...
"Your margin is my opportunity" -Jeff Bezos
Because somebody else then has an incentive to move into the market, increase production, and get the prices down. If that doesn't happen, then the market economy we supposedly live in is broken. And that would be a bigger problem than chip shortages.
> If that doesn't happen, then the market economy we supposedly live in is broken.

Well imagine that

It isn't broken, its just that the market economy doesn't really behave like free market fan boys would like to believe.
We should not call shortage what is just prioritization.
Why does the prioritization happen?
Because they can. This industry cycles between both states fairly often. Fabs lower output, then ramp up all the time.
Semiconductor is the new Oil. In the future expect the supply, demand and prices to be controlled by whoever controls the raw materials, manufacturing, supply chain and IP. The war bugle was already blown after the pandemic, and the war has started with the west blocking China's access to ASML's technology. And recently Japan's Prime Minister proposed a joint semiconductor development strategy with India, and China and Russia declared their own own strategy.

Hope to see more Oil like use of semiconductor technology for strategic geopolitical purposes. So the supply problem is not going to end, but rather fluctuate, just like oil.

Oil has the Petrodollar, what’s the currency for the semiconductors in the future?
Skilled Workers
> Oil has the Petrodollar

China is pushing to break that link. With Saudi Arabia, Iran, Russia and Venezuela on its side it may very well succeed in that.

> what’s the currency for the semiconductors

I see a multi-currency world. For semiconductor it won't be currency but access to raw material, supply chain and technology. Two examples of this are the blockage of supply of rare earths metal to Japan by China a few years ago during a diplomatic crisis, and very recently refusal by ASML, due to pressure from the US, to supply semiconductor equipments to China.

It will end when managers and pundits listen to economists and stop calling it a shortage
It will not, GPU market gets no break with going from mining apocalypse to the AI one heh.
I am not fully persuaded there is shortage. People could not get what they want is more about price not availability. That because manufacturers made too many things customers don’t need. Base on Q4 IC companies report, most of product driven companies, like Intel and AMD sales less than last year. However OEM like TMSC actually produces more and had more revenue.
I think the supply shortage was not at the high end but the much larger volume & lower priced parts.
Yeah, stuff like voltage regulators and fets and whatnot.
It hasn’t been possible to purchase a raspberry pi 4 from any seller listed on raapberrypi.com for months and counting now
Current price of raspberry pi 4 is killing it as a product.
Naturally, price and availability are intertwined. In this industry, availability is easy to see, because the manufacturers keep track of inventories on their own shelves and at their distributors. And there are tools for researching substitute parts, such as Octopart.

Even if you can find a substitute, it can take engineering effort to accommodate it, such as board changes and testing. Oh, and the people who can do this, have all moved to software engineering jobs that pay more, or they retired during the pandemic.

My tiny side business just went through this. The chip that I need became unavailable, as did all reasonable substitutes, at any price. I had to re-design my board in order to accommodate a different chip. Fortunately my labor is cheap and available, but dumping a whole bunch of that kind of work on a commercial business, all at once, can be painful.

My man… I have 100 week lead times on some parts and they haven’t changed for roughly 100 weeks.

There is a chip shortage.

Of course there is no shortage. You can buy all the parts from shady brokers at 10x price. My case it was 10x for PHY chip. ADCs were cheaper, only at 6-8x price.

Solution: buy another, obsolete PHY chip for only 2x price and strip ADCs for older products.

Just look at Mouser and Digikey, too many parts aren’t in stock and lead times with 1,5 years are ridiculous.

There's some reasonable demand for the lower density, more robust chips that were made in "paid off fabs".

Largely, I've read, those fabs were turned off -- they were in a legacy / sustaining engineering mode before covid. Lots of things, when you turn them off, for any reason, they're very difficult to turn back on; I suspect fabs are like this.

So -- the age of $0.50 parts made in a fab that was state of the art in 1997 is over. We must make new parts, on new processes, that are substantially different from those original parts, or we must make a 1997 fab. If you can't source all the 1997 machine tools to make the 1997 fab, that fab is likely a $200 billion dollars to build the first one and many tens of billions for each added 1997 fab. So those $0.50 parts are likely going to be $80 parts for the first 8 years of that "1997" fab's lifecycle.

Likely not going to happen.

The subtitle of the article is literally, "Fabs using older process nodes are the key."
Right -- that $0.15 chip is only possible to be made at the end of the lifecycle of a $10 billion dollar asset (the fab); that fab can only be made with a $500 billion dollar development to make all the supporting stuff.

You can't make that $0.15 chip on a new fab partially because the chips made on a super tiny process are different in a whole bunch of different ways (operating voltages, temperature / environmental tolerances, etc).

Maybe you can scrape together enough legacy parts from expired fabs to make some number of legacy "new" fabs, but even that's super tricky because just making the fab facility itself is hugely expensive and you can't fund it by selling the high margin parts that funded the 1997 fab from 1997 to 2004.

The designs need to change to allow using the higher density nodes (redesign entire ecosystems of parts / products) or you need to accept that the $0.15 part is now a $20 part for the next 6 years or longer.

But the entire article is not just about employing these old fabs, but about investing moderate amounts to increase the throughput of these old fabs. Compared to the large investments required for SOTA fabs.
(comment deleted)