Hacker News Reader

Comment by 11thEarlOfMar | Hacker News Reader

There are a several factors that conspire to make the shortage worse:

- These legacy chips fabbed at legacy nodes are made with legacy equipment. In many cases, that equipment is a) fully depreciated and b) no longer available. Yes, most chips can be made with new equipment, but the profits with a new tool set are not going to justify the expense. You wouldn't buy a new $8 million CVD tool to make a 180 nm microcontroller that might sell for $2.

- Many chips are fabbed at 1 place only. The chip company owns the IP and has a defacto monopoly on the features of that chip, or even class of chips. If that fab is shut down, no more are being produced and customers have to buy up what brokers can scrounge. The more warehouses get picked out, the higher the prices go. In one extreme example, a chip made at Infineon's Kulim fab in Malaysia, that sold for $0.47, was purchased for $100 @ by (ironically) a chip equipment company.

- Hoarding has made the issue far worse. Companies are instructing buyers to order more than 1 year ahead, and seeing lead times exceed 1 year as well. The more these lead times extend, the farther into the future buyers extend their horizon and lead times just keep rolling.

- It may actually be the case that organic demand growth is not the main problem. More likely, it was the successful (and therefore vulnerable) just-in-time supply chain that collapsed as soon as COVID interrupted the choreographed flow of material.

View on HN

This is worrying really. Back when I used to work as an engineer consuming parts it wasn’t unusual for some parts to disappear entirely with no notice. This usually turned out to be some legacy piece of equipment that broke at the manufacturer for which there was no replacement. There was no replacement because the whole manufacturer had been limping along with 1980s equipment.

One of the parts was devastatingly expensive to substitute. It was a very low noise JFET which was being used as a high impedance amplifier buffer. We had to evaluate ten different parts from other manufacturers then hand select a suitable part with the right curve and noise figure parameters.

To get 10 usable parts meant we had to buy 10,000 parts in, buy $40k worth of kit (HP FTW) and spend a month doing it. The BOM cost went to $6000 a FET from $50. Fortunately being in defence sector they just burned the cash while we redesigned the entire front end of the device which took months of qualification and test.

Automating this test process actually started my software career.

Edit: also I’ve got to stab Six Sigma and other process methodologies in the face here. They pushed elimination of waste and metrics that drove that to the front of business success. This meant eventually relying entirely on just in time manufacturing and supply chains. Coming from a defence background this was a disaster from the old days of buying in stock for the entire lifetime of the devices in the field.

I also had defense background and the the idea of buying enough components for lifetime of the device comes up regularly as part of the project obsolescence process. There are teams that would look at the supply of materials, and if one was going obsolete/limited availability, we would buy up as many as we need for the project plus safety margin and just store them.

Agreed on Six Sigma. I also worked at a major defense contractor for awhile and Six Sigma was a blight on us all. Not only did it cause the problems you listed but we also wasted so many hours from required "quality meetings" every week.

I picked up a book on it many years ago. The first chapter talked about needing to attack anyone that believe in six sigma, as in make an example out of them. Realized it was a cult mentality and avoided working for anyplace that used it.

Ah yes seen that as well. Don’t criticise the process or you will be eliminated.

I self eliminated and went and worked for a web design startup who’s process involved being drunk and winging it and it was more effective.

Oh god. Kill me with a spoon. I’d buried those meetings in a secret space in my mind along with all the other horrible things that happened in my life.

> One of the parts was devastatingly expensive to substitute. It was a very low noise JFET which was being used as a high impedance amplifier buffer. We had to evaluate ten different parts from other manufacturers then hand select a suitable part with the right curve and noise figure parameters.

This is relatable even to hobbyist electronics tinkerers, at least to those dabbling in low-noise measurements.

Yes. At least you can get the same noise figure measurement gear for a couple of orders less magnitude expense now though :)

There is a formal name for what is going on with the hoarding and an eventual collapse of demand. It is called the "bullwhip effect" and it will be brutal:

https://searcherp.techtarget.com/definition/bullwhip-effect

While a lot of people are saying the main issue is the "lean manufacturing" or "just-in-time manufacturing" supply chains, they do not understand manufacturing supply chains. These types of solutions (JIT or Lean) are designed to prevent the bullwhip effect and ensure that suppliers have predictable and constant demand, rather than feast or famine.

Without just-in-time manufacturing, there is significant waste and it is even harder for suppliers to predict what they should be making.

This current bullwhip is going to be horrible on suppliers after the supply catches up with demand and then the demand collapses as so many have bought years in advance.

This is a fatal mistake that the Automotive (and many, many others) who live, breath and preach JIT/6sigma/lean have made. The methodology simply doesn't work when the lead time for design-to-delivery is 18 months and you need your orders in and volumes committed in order to hope to see them on the market next year.

Fabs do not work on jit/lean order principle - if you up your order volume, you can't simply scale it up on demand the next month - even if your design is done and taped out. Fabs aren't sitting around idly waiting for orders nor are they overprovisioned "just in case".

An example of this would be AMD and TSMC. AMD books capacity 3 years in advance for which TSMC designs the entire fabrication facility around, orders the lithography machines (which have a 12 month lead time and can't scale up on a dime) and assembles and staffs those same facilities to produce the wafers. It's not a general-purpose fab as most think. There are rumors that Apple booked m1 fab capacity (on TSMC 5nm) in 2018. For the chip inside a car for example, it is built on much older process, and when 5 of your biggest customers "JIT" reduce their order volume for the next 3 months you simply put those staff elsewhere (probably on the 6x higher margin cutting edge fab you just got booked and prepaid for) since anyone else who might want that process and capacity has already has their orders being produced.

Due to the sheer capital requirements (and consequent commitments required) and multi-year forecasting involved in simply getting what you think you'll need, let alone having extra, it's not hard to see why a small increase in global demand would cause a dramatic shortage. JIT works in Automotive supply chain where you can request that your bearings get made 10 days before they're delivered to the factory, but not when there is an 18 month latency on "scale up" request/response.

It's applying JIT principles to inherently non-JIT supply chains.

You can apply the same principles, but you have to adjust the time coefficients to fit the realities of the supplier.

I guess if you account T = 18 months, JIT can still apply.

This is correct. Every type of suppler has different adjustment time based on their own systems. Tooling & upfront design is actually common in a lot of industries, chips just have extreme tooling & upfront design requirements.

Trust is the thing that keeps the entire global economy going. Sure, it is attenuated at many points and has to be back-stopped at others, but all division of labor is expression of trust in the work and promises of others.

Now, the whole world is trying to get TSMC to invest in other locales in advance of Chinese invasion of Taiwan at the same time that NATO is funneling arms and personnel to Ukraine ahead of Russian invasion. Work isn't getting done on expected timelines because of unprecedented disruption caused by responses to covid.

Meanwhile, down at the bottom of the pile, when airlines cancel my flights for unprecedented problems, I wait 3 hours on hold because they don't have enough people working support.

All this to say that, yes, the bullwhip is going to be horrible, but so is not being able to perform your function in the economy because your required inputs are unobtainable. There's no graceful path for a system with so much volatility. Hopefully we can all stomach the creative destruction.

> These types of solutions (JIT or Lean) are designed to prevent the bullwhip effect and ensure that suppliers have predictable and constant demand, rather than feast or famine.

If that were true, they would have ensured the chip manufacturers had constant demand, rather than a famine period followed by a huge feast.

The truth is that just in time is being used as a cost minimization procedure, by offloading the demand risks into the suppliers and increasing any feast or famine event. That maximizes profits as long as there isn't an emergency, what seems to be the favorite administration goal of the current age.

If there is an emergency, all the large companies can simply go to the government for free money, why would they care about doing anything but absolute profit maximization.

One way to deal with the bullwhip effect is to build robust futures markets around the underlying components. This improves the market two ways. 1) It allows both suppliers and demanders to hedge their exposure in the futures market, which gives them more confidence to plan the rest of their business without uncertainty. 2) It enables professional speculators to engage in price discovery and telegraph credible signals to the rest of the market about the long-term trajectory of supply and demand.

>> One way to deal with the bullwhip effect is to build robust futures markets around the underlying components.

These chips are not fungible commodities. In the automotive space chips are selected based on a number of things including feature set - how many CAN interfaces does it have, how many GPIO, how many PWM, what kind of brown-out handling, what voltage range, thermal rating, etc... Software is then written where a portion of the code is specific to that chip. There are cases where an OEM get really good pricing on a bigger (overkill) chip and helps a supplier get those at good prices too, but usually they just beat you over the head for the lowest cost. Even if everyone adopted a few larger parts and used just those, they're still usually made by a single chip company.

I think in time chips should be fungible commodities. A lot of components, such as DRAM, NAND are fungible. As we move towards more open source tech, like RISC-V, for microcontrollers, maybe many of them become completely fungible as well. Especially if there isn't an infinite variety of them -- I guess the completely custom RISC-V chips won't be fungible, but if one can use an off-the-shelf standardized module, those could be.

> off-the-shelf standardized module

This doesn't exist now for the same reasons it probably wouldn't work in the future: You can save significant amounts of money if you cut the pieces out of the "standard" that you don't actually need. Chips are cheap to make, in bulk. It's almost entirely a fixed cost/area. That's the magic. So, it's almost always cheaper to make a custom chip than to pay for some bloated standard containing large sections that will never see power, especially when chips are something like 40% of the BOM, for modern cars.

>... especially when chips are something like 40% of the BOM, for modern cars.

No knowledge of the automotive industry, but that sounds impossibly high.

Sorry, that is high! Somewhere around 20% for low power electronics (not just chips), according to https://www.statista.com/statistics/1248308/share-of-automot...

Aren't these sort of year-out agreements basically futures contracts? Especially the ones that don't allow you to back out. I know at least in the business I'm involved in, the deals are tradeable in the same way futures contracts are.

2024 cars are going to be so cheap, imagine if that's the cause/effect that made EV cars finally affordable.

I think that the components being hoarded have little effect on the final car price, but are critical to the car working -- e.g. required but not sufficient. These are like $0.50 microcontrollers or similar.

I think there isn't a bullwhip effect on batteries, which is the main cost associated with TVs.

Used car prices probably will collapse, though.

Used car lenders will collapse.

The 72 month loan for a $50,000, 2019 F-250 will be completely underwater, and the car owners may decide to literally send the cars “underwater” to get out of the loans.

This doesn't sound right? I thought the value of the collateral had no bearing on the loan, except as a condition for the lender to call the loan if the collateral loses too much value?

If a person owes $40k on an asset worth $30k, why would they keep paying?

Based on what happened in places like California with mortgages, they will default one way or another.

Defaulting has a cost though. If you default on your car loan, how likely is it that you're getting a loan for the next car? What kind of interest rate are you going to get?

Depends. You can get a mortgage with a modest premium a year or two out of Chapter 13 bankruptcy.

In Europe at least, they'd still be on the hook for the rest.

This seems trite, but the reason EVs aren’t affordable now is that not enough are being produced to satisfy demand. The actual cell cost (not price) is crazy low, like under $80/kWh for the LFP type cells in the base Model 3. But the price goes higher because demand is so incredibly high, even though Tesla is making 1 million EVs per year and growing rapidly year on year.

Used EVs used to be pretty dang affordable already (I got a Leaf and a Volt for under $10k each) before all used cars became more expensive and the price of gasoline increased, increasing the demand for EVs.

It’s just about scaling up production.

China is turning out so many EV cars, dealers are buying back old one's to keep the price up. This is what BMW did (and possibly other German car manufacturers), in the late 80's/90's.

Another thing I see China doing, is pegging their currency to the dollar, now when the UK did this to the stronger German Deutsche Mark, the UK under Lord Norman Lamont crashed out of the Exchange Rate Mechanism and the UK saw interest rates hit 12/13% so I wonder if this chip shortage has more to it other than investment reasons only?

the other side of this cyclical effect is new ideas emerge to take advantage of cheap supply and drive new demand

Ultimately, I think in the automotive sector in particular the management made a conservative decision that ultimately may have been fatal to them. They broke the JIT pipeline because they assumed, wrongly, that the Trump admin would not bail them out.

The lockdowns happened and some lever got pulled. They dusted off the “We love you America, come back and buy a car someday” commercial’s, cancelled everything, and opened the bottle of Scotch.

> Many chips are fabbed at 1 place only. The chip company owns the IP and has a defacto monopoly on the features of that chip, or even class of chips. If that fab is shut down, no more are being produced and customers have to buy up what brokers can scrounge. The more warehouses get picked out, the higher the prices go. In one extreme example, a chip made at Infineon's Kulim fab in Malaysia, that sold for $0.47, was purchased for $100 @ by (ironically) a chip equipment company.

Perhaps this is an opportunity for FOSS silicon? Everybody understands that currently FOSS tooling/PDK's/etc isn't up to the task of creating a cutting edge CPU on a leading edge fab process, which is the usual retort when somebody brings up FOSS silicon.

But commodifying a sub-$1 chip? Yeah, why not?

It's the tooling you're making that chip with, not the chip design.

You can buy used equipment from a few nodes back for basically scrap value, but building that machine today would cost more than building it when it was new.

You'd need to bootstrap a whole "old nodes" industry that does nothing but build obsolete machines to build obsolete chips once all the second/third/forth hand old node machines are gone.

In short: we've climbed the ladder and set it on fire behind us.

I think in retrospect you can use the former eastern bloc as a metric for how this can go. They built a whole semiconductor industry from scratch and were never financially competitive. Parts were significantly more expensive and generally less reliable than the western made ones. You’d be doing the same as them but with less budget and fewer customers.

The risk really is entirely that we single source parts.

A few years back I designed something using only generic multi sourced parts and it was large and inferior. So there’s trade offs everywhere.

If you, for instance, only 3D printed plastic parts you could find in stores, you'd also find that your prints were generally significantly more expensive and less reliable. The upside is the ability to get anything you want at any time, as long as it's sufficiently far from the cutting edge, whether production setup for it exists or not.

The point here would not be to replace traditional manufacturing but to augment it by acting as a backstop for total supply failure.

Perhaps (re)bootstrapping the old nodes industry is exactly what's needed?

So traditionally, the chip industry worked such that that the high-end products moved first to a new process node, and less demanding equipment was fabbed on equipment a couple of generations old before that equipment was thrown on the scrap heap. Rinse and repeat, everybody was moving towards smaller and smaller nodes, just that some could trail the leaders by a few years.

But it seems these days the world is different, in that the cost per transistor has actually been going up for the leading edge processes, and power consumption has remained roughly flat. So for anything that doesn't need the latest gen process, there's really no upside to moving. Depending on who you ask, it seems the cheapest cost per transistor is somewhere in the 20-40nm ballpark. If so, it would imply there would be a long-time demand for these process nodes, and there is a need for an industry to supply tooling for these nodes, once the existing crop of second-hand equipment (left over when the leading edge users moved on to smaller nodes) is in need of replacement.

> Depending on who you ask, it seems the cheapest cost per transistor is somewhere in the 20-40nm ballpark.

While correct, this is only true because of all the old infrastructure sitting out there that has been fully depreciated but still works fine. It used to be the latest and greatest thing and paid for itself back then. Nowadays it is just sitting on the asset sheet of a bunch of companies who are all going "if it still makes money, keep running it!".

The paradox of this all is what you identified: "if so, it would imply there would be a long-time demand for these process nodes". No one is going to invest in developing old nodes because the ROI isn't there. The demand is, but there probably isn't much cost differential between building a nearly leading edge fab and a 60nm fab.

If you could sell a turnkey 60nm fab right now that was delivered and setup quickly the demand would be insane. But no one can because fabs have never really been turnkey. When you factor in the cost to develop all of this, the capital you'd have tied up, etc. the ROI suddenly fails to compete with interest bearing bonds I'd guess.

>cost per transistor has actually been going up for the leading edge processes

Isn't this just because of overall chip shortage? We might see very different scenario if TSMC, Samsung and Intel overinvest now.

Also, you could say that cost per transistor increased for older processes as well...

>power consumption has remained roughly flat

Any source for this? Anything I've saw contradict that.

The cost per transistor is going up because the latest nodes require too many process steps. An older node might require 90 steps, the lastest nodes require 900 steps. Every step requires a "tool" and every step has a chance of yield loss. The transistor is getting harder to make faster then it is getting smaller.

The new EUV steppers with soft x-ray wavelength can do in one shot what took 4 shots with 193nm ArF tools. And for each shot you have to apply, bake, develop inspect and strip the resist. This was supposed to help but the EUV steppers are so expensive it is kind of a wash price wise.

> Isn't this just because of overall chip shortage?

AFAIU, no. The basic problem, again AFAIU and I'm not an expert on this topic so take with a grain of salt, is that while the marginal cost of each transistor might still be less than at the previous generation, the R&D and capital costs have grown so immense that once you amortize those costs over the stuff produced for the lifetime of that process generation you end up with a higher per transistor cost compared to the previous generation.

> Any source for this? Anything I've saw contradict that.

Just search the web for articles about the breakdown of Dennard scaling, plenty of stuff to read on the topic. Even wikipedia mentions it https://en.wikipedia.org/wiki/Dennard_scaling

> In short: we've climbed the ladder and set it on fire behind us.

This is a really good description of what JIT does when no one considers economic reflexivity. The time duration for problems to surface is long enough that no one can "do the right thing" and last past next quarter's or next year's financial performance assessment (whether internally or by financial analysts). So everyone participates in a game theoretic sensible fashion, and gets burned either an SKU at a time as single source or limited source manufacturers burn out their legacy equipment, or all at once with black swans like this pandemic.

I'm certain no one who modeled JIT in business school research has worked out if it was more cash efficient to put up with and pay the piper when (not if) the disruptions happen, or to JIT only specific items that meet supply robustness criteria. The purchasing people mania for single sourcing plays into this dynamic. Supply chain management is enormously complex, but I fear they're still way behind the eight ball as they find out it gets fractally more complex than they had bargained for when they set out to cost rationalize it all.

That's true, but given the original design collateral it would be easier to design a version that could be manufactured on present-day processes by re-doing the layout and timing than by starting from scratch.

Still not free though. I think people underestimate how niche all the skills are here.

> In short: we've climbed the ladder and set it on fire behind us.

I guess that's fine as long as we keep climbing, right?

100% on the hoarding thing. Everyone I know of is making larger buys than ever, enough for years or production if not a products lifetime worth. This has t have an impact on the available supplies.

nice to see companies having behaviour as stupid, short-sighted and damaging as the people hoarding toilet paper

Just like toilet paper, hoarding is just game theory in action. If you don't take part in the madness, you're at a disadvantage as the "sucker".

For toilet paper, the cost of that disadvantage is having to wash your backside with the shower head and the benefit is the warm fuzzies of doing the right thing. For a business with no chips, it could be an existential threat. Warm fuzzies don't pay salaries.

The only way to avoid it is to have every company in the world to agree to not hoard, and harshly punish "defectors" somehow. Or have the supplier strictly ration supply, which doesn't work if the supply is coming from existing stock in the hands of distributors unless they also cooperate (or collude, depending on how you look at it).

None of that is remotely possible in the global industrial system. JIT is often blamed, but to avoid it, trillions of dollars worth of material would need to be hoarded constantly, which is its own problem: storage costs money, logistics and space, and its a substantial risk if you end up not needing it or the goods expire or become obsolete on a shelf and overproduction is a waste of resources. Going bust because you hoarded 5 million components that you never actually needed and couldn't offload is still going bust.

For most of the world, bidets are a thing an it is a curious thing as to why the toliet paper industry is so powerful in the west.

I prefer the bidet, and seeing normal reams of toliet paper cost now $30 for the "mega" value pack, I'm amused at the production of something, the trucking of something, the retailing of something that is ultimately discarded . The whole thinking and execution of it is backwards.

---

It is remotely possible, the governments can set immediate import/export restrictions and also guide manufacturing as they did during Covid for incubators and such. While it won't be effective, it can either help the stock/supply rebound, or throw into more fear...

> For most of the world, bidets are a thing an it is a curious thing as to why the toliet paper industry is so powerful in the west.

At least in America, my Grandmother's generation (great depression) grew up with outhouses. When homes were renovated to add an indoor bathroom, they were done on the cheap.

Heck pre-pandemic the idea of a bidet was crazy, I had to work hard to find one that was reliable and at a good price. Now they are super available and advertised all over the place, but when I went to get one installed 10 years ago it was not easy just finding a reliable, cold water, model! Warm water models were obscenely priced in the US.

It was likely a chicken/egg thing. Consumers didn't know about them, so no one made them for the US market, which meant the few that did exist cost an arm and a leg, so no one bothered to buy them so no one bothered to mass produce them.

On top of that, when I inquired about getting an outlet installed behind my toilet for an electric bidet, my plumber and electrician thought I was crazy!

Same question in 2021, "sure we do that all the time".

Same here, they all come from the Far East, though they have at least started fitting the right plugs rather than a deathtrap adaptor.

Pretty sure mine didn't make them much money since they didn't include a pair of stainless steel mounting brackets, and they FedExed them to me. From Korea. In a box 100 times the necessary volume.

Here in the US, Home Depot now sells an affordable heated bidet, under $200 IIRC. Doesn't have all the features of nicer models, but it gets the job done and it is available all over the place.

Bidets have gone mainstream in the US!

What is really needed is expose the JIT pipeline to upstream. We are buying 10,000 parts but need 100 per week, figure out what works with your production pipeline to deliver at that rate or above.

Issue is that people will skew the numbers reported upstream. This way they can affect what their 'fair share' would be. This tends to devolve into everyone mis-reporting numbers in the same direction, causing the reported numbers to be useless upstream.

The other way to avoid it is to have the government step in and impose rationing, as done in times of war... at least, back when war meant the common American citizen was actually expected to sacrifice for the greater effort.

If the semiconductor shortage continues to worsen, I think we'll start to see calls for this type of intervention. One step that could be taken immediately is to prohibit sales from the distribution channel to overseas brokers, which are currently resulting in stupidity like $800 FPGAs that used to cost $20. But it's also true that this sort of market manipulation has downsides of its own.

I honestly don't know what the right thing to do here is. But I do know the industry had best unfuck itself, and that right soon, or this situation is going to have major consequences in everyday life.

> ... is going to have major consequences in everyday life.

is going to ? Already has.

Used car prices start climbing Jul '20.

https://www.cargurus.com/Cars/price-trends/

Car prices going 5-10% up are very far from _major_ consequences.

You reading the same page that the GP linked? It says YoY car prices are up 36%... which is enormous.

That's got a huge ripple effect across the market, as best as I can tell, because it pushes people to buy older/lower cost cars across the board. Either that or they continue driving their old/potentially unsafe car even longer.

Apologies. The index below does not correspond in any way to chart, which confused me.

The thing with toiler paper is that we never actually ran out of it. Only the last part of the chain did (the shops). Toilet paper just takes a lot of space and sells for relatively little so shops in general do not want to stock much of it.

Basically even during the worst of the panic it only ran out for the duration it takes for a truck to be driven to the manufacturers warehouse, load it, drive to the shop and unload it.

With semiconductors we are actually running out of stock/capacity at the very beginning of the chain instead of last mile delivery.

While you could use newspaper as a substitute for toilet a paper some company has nothing as as a replacement. New part means costly re-certification of the whole product. No parts means one is out of business. So buying everything that is available on the market looks like very good decision from business perspective.

If you're that dependent on components that there is no replacement for, then perhaps buy enough stock to last you through a redesign and recertification is in order. Having a contingency plan for such scenarios seems like a no brainer.

There are sooo many no brainers, that bureaucratic monstrosities also known as Big Corps just can’t get right. I work for one and more than half of the daily problems are home made and were simple no brainers elsewhere.

I got all my electronic parts on time for my private development and first small manufacturing run. It is indeed no brainer.

It's because they value short term gain over long term gain and stability. I see it all the time in software development too.

I'm pretty sure that whatever business you work for has no planned replacement for their local power company ceasing to provide electricity.

This doesn't mean they should go out and build their own gas peaker plant.

We live in a society, and no man is an island. You can't control everything you depend on. Sure, you can control some things you depend on, but that's a long, and expensive rabbit hole. Hoarding twenty years of canned food and seventy thousand rounds of ammuniton is all well and good, unless doing so puts you in a situation where you can't pay rent.

Nice straw man. But I'll go with it for no other reason that I hope no one believes in that absurdity as an argument against planning for component supply issues. Issues that businesses have experienced time and time again since we began mass producing transistors.

That is because it's a utility that source power from several producers. The risk of no power what so ever for a prolonged time is not something to even be considered. Should that happen, society as we know it is probably at risk and my job is no longer important at all. Oh, and someone actually do plan for intermittent power delivery failure including having two different outside sources for power, a battery backup and diesel generators, data centers come to mind.

Don't big businesses often have Diesel generators to keep power to essential devices going through outages?

I don't mean a two-day power interruption, I mean a two-year power interruption.

And if they intend to feed the generator in perpetuity, I'll happily shift the goalposts to some other critical supplier not providing their services.

I don’t know that massive datacenters are truly a good investment for my employer, but at least we have several zones in two regions in case of power/network/natural disaster issues.

You’d better tell the city of New Orleans that.

The best case would be somebody keeping a big stockpile and selling from it at a premium. “Anti-gouging” laws prevent that for toilet paper, so each of us has to hoard it instead.

Hoarding is just client side caching. Once the initial shock wears off it’s actually a benefit and prevents future shocks. Keeping a stock of critical components on hand keeps business from needing to place emergency orders.

FTR I'm just a hobbyist, and I've been hoarding chips and boards too (though probably not in quantities that affect anything)

Paging a semiconductor engineer. Isn't it not necessarily possible to do die shrinks and get drop-in compatibility, particularly for Q rated automotive chips? Moreover, some of these chips eat 5 volts by design and I don't think 28nm or 14nm or 7nm will provide enough resistance for that. Plus, some have pmic or internal power mosfets that are specific to various processes.

It's almost never possible to do a function-compatible die shrink for anything that needs to work over a large voltage and/or temperature range. Many automotive chips operate at 80V or more, 5V and 12V parts are ubiquitous, and indeed lots need bulk area to be able to switch large currents. Making them on small processes is not only an enormous design effort where it's possible at all, it also doesn't make sense because there's still way more capacity on larger processes than on smaller ones, and the only way small processes win is by packing so many dies on a wafer that they can get away with their terrible per-wafer cost and awfully bad defect rate.

Thanks for the informed reply.

Think of it this way. You could probably design something logic-compatible, voltage-compatible, speed-compatible... but a chip is fundamentally an analog device with sensitivity. You can't change the mask, let alone the process, and still have the same device. Something will be different. Certification is the process of saying the difference is OK.

Right right. I see a lot of pmic parts, newer MCUs, 3.3V and 5V tolerant parts, as well as various FPGAs out of stock. All of these are fabbed on >40nm or more, sometimes much more.

Not going to ram 5V through a 7nm finFET.

I came here to say exactly the same thing. I'll go one step further... not only are the equipment/fabs fully depreciated, but they are 8" rather than 12" and so none of the new equipment/processes will work with them. It's a huge fight to even get anything older than 90-110nm in 12". That is why the relatively high volume, but very margin constrained, display industry is stuck at those nodes. It doesn't make sense to build and bring up a new fab with new equipment at margins below 30%, but the entire industry is built on that cost structure.

Further, a lot of the designs for older processes have been iteratively improved (yield, performance, etc), but the older models (simulations) for them are insufficient to design new chips without explicit reuse or iteration. They are effectively stuck on that node at that foundry, unless they want to incur significant (opportunity) costs.

The other issue that I ran into with printers last year is that the OEMs have to be careful not the sell the future.

Everybody wanted small all-in-one devices last year, and nobody wanted big office printers. Companies like HP, Epson, etc won’t invest too much capital in production because that demand will implode later. Those companies probably sold 3-5 years of demand in a year for some SKUs.

It similar to the issue in the ports where everyone acts I’m their perceived best interest, which often screws over everyone.

My guess is that we’ll see a price collapse in the next 12-18 months for some commodities as the hoarding orders start getting fulfilled and companies are stuck with warehouses and balance sheets full of inventory.

You mention JIT, and lead times greater than a year. Are these 2 things not at odds with each other? Are you saying the vulnerable JIT supply chain when interrupted led to hoarding? The hoarding causing lead times of up to year which is what JIT was supposed to minimise in the first place?.

Once the chip shortage happened, partially due to geopolitical reasons, that changed everything. The West/US will never fully rely on a single point of failure again no matter how hard the HBS MBAs and Chicago thinking push it to trim and be "efficient". Some industries are too important for other industries and leverage on that over those areas is too risky and costly now.

I'd pay double right now for GPUs directly from the source, not from some sketchy third party.

Right now our EV/auto, space and even AR/XR industries as well as gaming and everything that requires chips, we are at the mercy of an external market that has a slant against the West. It will take some years to get out but we'll never not expect that in the future again. If costs go up costs go up, but availability should never be allowed to be used as leverage again, that is too risky and too costly long term.

Availability that is reliable is always more important than efficiency or cost, because right now lack of availability is costing lots of extra time that has the potential to lose entire industries, that is not acceptable in any way.

Very little margin and too much optimization/efficiency is bad for resilience. Couple that with private equity backed near entire market leverage monopolies/duopolies/oligopolies that control necessary supply and you have trouble.

HBS is even realizing too much optimization/efficiency is a bad thing. The slack/margin is squeezed out and with that, an ability to change vectors quickly. It is the large company/startup agility difference with the added weight of physical/expensive manufacturing.

The High Price of Efficiency, Our Obsession with Efficiency Is Destroying Our Resilience [1]

> Superefficient businesses create the potential for social disorder.

> A superefficient dominant model elevates the risk of catastrophic failure.

> If a system is highly efficient, odds are that efficient players will game it.

Hopefully that same mistake is not made in the future. It will take time to build up diversification of market leverage in terms of chips for availability. Hopefully we have learned our lesson about too much concentration, with that comes leverage and sometimes a "gaming" of the market.

This chip shortage, and all the supply chain problems during the pandemic as well, will hopefully introduce more wisdom and knowledge into business institutions that just because things are ok while being overly super efficient, that is almost a bigger risk than higher prices/costs. Competition is a leverage reducer. Margin is a softer ride even if the profit margins aren't as big.

[1] https://hbr.org/2019/01/the-high-price-of-efficiency

History tells that "we will never ... again" almost never works; people forget what happened and start all over again, the cycle is infinite.

There's no organism called "humanity" - there's a bunch of individuals with indirect, lossy access to the past through the study of "history" - which is useless if said individuals don't study it.

This is why you see so many older and/or young-but-learned people telling you to read history - and yet, it's really hard to listen to them until you start getting personal, emotion-laden experiences that convince you to do so (such as this pandemic).

Wouldn’t it make a lot more sense to have ICs that aren’t space-constrained be modular in these applications?

In a phone or laptop, obviously it would take too much physical space to have such components be changeable, but for vehicles and machinery couldn’t you regulate that common boards/chips/memory be easily accessible and interchangeable to a standard? (Kind of like how you have a fuse panel for your car or a fuse box for your house).

I get that this wouldn’t solve this issue, but it would certainly help by making parts that are susceptible to supply shocks be easily recoverable and reusable.

A lot of parts are modular, if you'll accept "standard packages and pinouts" as "modular". And this helps a lot!

The problem is that the parts themselves aren't standardized. So often a very careful selection process followed by a requalification is required.

They’re pretty much doing this with MCUs and power at the moment apparently. Designs are moving to modular so you can respin a carrier board for the available stock rapidly rather than have to respin the entire product and stock.

I have no clue how things are built in this field, but can’t you put the same design of a 180nm one on a 7nm one?

Not really. You can't reliably make large features with a small process node, and because the smaller processes are dramatically slower and more expensive and have lower yield, unless you can counter all that with more dies per wafer you end up with a whole lot of defective parts at enormous expense. The bigger the die, the higher the chance of a defect landing on it, and making it useless. Also small feature processes have much lower capacity than larger feature processes, so even if you were able to switch there wouldn't be nearly enough wafer capacity. The fact it's 4-5 orders of magnitude more expensive doesn't help either.

> You can't reliably make large features with a small process node

Why is that?

Because (very very roughly) the way you make small features is by scaling down your projection area to a tiny size and then stepping it across the wafer. Your projection area boundaries are where errors are most likely, and the bigger your features the more such boundaries they cross. The details are quite involved, but this is the general problem.

But then the errors are only along the boundaries of a very large feature, which was originally designed for a less precise process.

Also, you could use the design files from before any compensation you apply (to compensate for lithography errors), and use those files to derive a new compensation.

No, the errors are inside the feature itself, since it doesn't fit on a single mask. You can't reuse design files on a new process. Everything will be wrong. You have different material properties, different gate capacitances, different parasitics. No matter what you do, you have to redesign to move to a new node. The problem with large features is, even if you do all that work, you still can't reliably make them with the smaller process, and all the benefits of the smaller process (smaller die size, faster switching, higher speeds, shorter interconnects) go away if you use large features.

Aren't the necessary changes relatively "local", though?

That is - timing closure is hard, and it's a "global" problem - if your design fails timing closure, you might have to re-do the entire layout and/or design itself.

However, if you change the gate capacitances, then I would expect that you'd just have to make a lot of small changes in a lot of different (but unrelated) places - which seems like, while tedious, would be immensely parallelizable with the number of engineers, and not risk any complete re-layouts like meeting timing closure would. So, a "redesign" more than a redesign.

What am I missing here?

You're missing that large feature sizes are often used because they need to handle particular voltages and temp ranges. If you change the material properties those characteristics are no longer viable with the original spacings and dimensions, so you get to relayout anyway.

Not a chance, while you can stretch out horizontally to make a 7nm look like a 180nm (conceptually, still has problems in reality), you can't do that vertically. For example, the gate oxide at 7nm works differently than at 180nm.

Reply on news.ycombinator.com