1. TSMC were spending ~$30B per year on capital expenditure already.
2. Since leading edge node is forever increasing in cost, the next 3 years, i.e 3nm in 2022 and 2nm in 2024 are expected to be higher. ~$100B investment aligned with their initial plan / trend / target anyway.
3. So really the major news is stopping ( or to be precise, delaying ) price reduction. Which is unusual but understandable given the current demand situation. And No, it is not TSMC's fault. You should ask how every single Fabless Semiconductor company has failed with their demand, supply chain analysis and projection. Although one could argue it is not their fault either, since their client gave wrong projection. The only company that is not affected is possibly Apple.
4. There has been mounting pressure from investors, politicians and MSM media on Intel IDM 2.0, and supply problems. This announcement feels more like addressing those concerns.
5. Remember both Samsung and Intel are expanding capacity. And even GF and many other smaller players. It took DRAM and NAND three years
to catch up with demand ( and then over supply ). Which in hindsight is pretty damn impressive. Although during the ~2016 - 2019 everyone felt awful.
> The only company that is not affected is possibly Apple.
This is an interesting point. Apple is pretty hard nosed on their predictions and capacity reservations, to the point where they could be funding part of this expansion at TSMC (as they have done for Hon Hai, among others, for decades).
I continue to be astonished by Apple’s ability to manage their supply chain so tightly (which requires not just an iron fist but very very intense internal process). They seem to be the only ones who can manage to do that.
And it’s not like other big companies are lazy; the part that really amazes me is that their process hasn’t leached out to other companies, as so much else does in the Valley.
I just saw something about this. The reason Apple hasn't been impacted is because their sales have been dropping for the past few years on iPhones and other devices. That trend hasn't changed during the pandemic. While many of the other players predicted drops during the pandemic, and initially it did, that demand dramatically increased, especially in the desktop market (AMD, Nvidia, etc). It is hard to fault companies from predicting the massive increase in a market that has been decreasing for decades(?) now. There is limited capacity at TSMC and Samsung. They leave buffer, but that was quickly consumed because almost everyone needed more capacity. It takes potentially years to increase capacity (build a new fab, etc), they have no room to allow any companies to adjust their capacity reservations.
So at the end of the day, it isn't that Apple was necessarily better...their demand just didn't change and continued the downward trend in certain product areas, giving them breathing room.
It does look like apple's unit sales (of phones at least which are the lion's share of product) have been declining (conveniently they stopped reporting unit sales long ago) but I don't think that really applies here. Apple's share of the total semiconductor market is far from dominant.
What's really interesting is how well they couple demand to shipment. One is that they appear to have low levels of unsold inventory, both on the incoming (BOM) side and output (unsold manufactured output, such as phones), reflecting some incredible discipline. They also manage to insure themselves against supply shock (because such a tight tolerance for over purchase and overproduction makes your supply chain more brittle) by making big moves in their supply chain like financing their suppliers, doing manufacturing R&D on their suppliers' technology and then supplying that tech to the suppliers, taking big positions in commodity markets (e.g. famously in DRAM a few years ago) and other such things you can do when you have so much cash on the balance sheet.
Better use of capital and better supplier flexibility (why consume your suppliers’ capacity on stuff you don’t need when you could switch them to the next gen thing you need?).
You could think of it as avoiding bufferbloat in the physical domain.
Also refreshes are real annoying/unprofitable when you have millions of old tech sitting in warehouses. Apple has pretty good science on the sneak peak / release / fulfill of new tech. You can basically get day1 or at least week1 release of any new product.
One thing to remember is that the current CEO Tim Cook started in the supply chain management side of the company...
And did that job for Steve Jobs during the growth of the iphone.
It would be expected that Apple would have extended insight into supply chain and capacity, seeing as the head of the company was/is the one who did that job.
It's also a strength on the demand prediction side which is typically done by marketing not the manufacturing side of the business.
Yes, I know Cook came from Compaq for his supply chain chops, but there are amazing supply chain folks elsewhere in the industry. On this dimension Apple is in a class by itself. Perhaps that fixation does come from Cook, as supply chain wizards typically don't rise that high in the corporate chain.
Apple is experiencing supply chain issues, and there is a shortage of laptop devices right now if you try to make bulk orders (>1000 units). I have no context on their phones.
There aren't many companies that can compete with 100 billion dollar capex demands. While the loans are freely available, the development risk remains. TSMC has a proven track record pushing into the latest node - intel does not. Samsung has been a little bit behind for a while while keeping pace.
Following this trend - 2nm will have something on the order of 200 billion in capex outlays. Given that being first to market has a premium, the loan will be high risk. If you borrow thinking you'll be first and are late by 4 years then you and your bank are going to have a problem.
On the bright side, the new processes will have a longer life. When your 135nm process is replaced by 14nm within a less than a decade, there's a huge gap there and the old process isn't worth much. If you go from 14nm to 2nm, you can still use 14nm for a lot of price sensitive customers.
The density increase is similar between those two jumps, isn’t it? Density significantly drives the price of chips since you can use it to create more chips per wafer. Maybe developing chips at 2 nm will be much more difficult due to quantum effects?
Before you start reaching quantum effects, you'll run into crosstalk problems where electrons on 1 circuit path can jump through the silicon substrate to other paths, effectively causing a short. There is a limit to how much you can do with silicon as your substrate simply because of the size of the atoms.
The law of diminishing returns. Shrinking brings lower and lower benefits because the jumps may be hefty in relative values (10% smaller) but in absolute values (nm) they don't bring the same benefit as they did 15-20 years ago.
On the other hand, even in practical terms the older nodes are more than adequate for most customers these days simply because they are far closer to a modern node than the equivalent would have been 20 years ago (for the reason stated above).
Every single fabless semiconductor designer failed to predict demand because their customers lied to them. The car industry pulled a really dick move last year, and then pulled an even more dick move in the opposite direction, and now everyone is screwed. Back in early 2020 the car manufacturers decided that demand for cars would go down, and because they have a religious aversion to keeping any stock, cancelled lots of orders with their vendors, screwing said vendors over. Their vendors could not afford their fab slots because automotive is such a big part of their revenue, so they cancelled their fab slots. Those slots were happily sold on to entertainment and computing customers, who anticipated a jump in demand due to people staying home more. So far so normal.
However, the car industry got it badly wrong - people, afraid of public transit, started buying more cars rather than less. The car industry, being screwed due to their just-in-time religion of zero stock, was faced with their production lines stopping so they called up all their vendors, and asked for those orders back, and some more on top. The vendors then tried to get their fab slots back, and were told to come back next year. Some of them ended up buying other fabless IC designers out of their slots, causing the problem to spread. Others cancelled their existing orders to other customers, and auctioned off their existing inventory to increasingly desperate car manufacturers at a 6x to 8x premium. Anyone who was not prepared to pay that or didn't act fast enough was screwed. From that point on, a bunch of companies that depend on those lines of microcontrollers had to rapidly redesign their product to use another device, taking even more devices off the market with unplanned demand. The users of those devices then had to move to others, causing even more availability cascades. This is how two nasty moves by the car industry caused global market disruption in a number of industries that depend on electronics. This is not a normal "demand has increased, and industry can't keep up" event, it's elephants dancing and trampling everyone else underneath.
This is further aggravated by the top three automotive semiconductor suppliers (NXP, Renesas, Infineon) having their facilities destroyed in two unrelated disastrous events - a fire at Renesas' wafer processing plant, and Texas freezing over, destroying NXP's and Infineon's fabs through cleanroom contamination and process interruption. Those events took out months' worth of production, and destroyed product that had already been sold before manufacture. This would have been recoverable in a normal market, because distributor stock could hold a couple months, but in this case it was game over for non-automotive customers as all distributor stock was already gone by then.
I see this in my work every day now - customers coming to me for help with redesigning products to use a different microcontroller, or help with sourcing parts from unusual sources because their normal channels are gone. I've been in this industry a long time and never seen anything like this before. This is not a failure of supply chain analysis and projection on the part of the fabless semicon vendors. This is their biggest customers fucking their vendors over not once but twice by lying to them about their own demand.
Now I know why my business, that has nothing to do with semiconductors are booming too.
I own a store that sells parts and tools to attach things to each other (originally we sold nuts and bolts but those are unprofitable without gigantic volumes).
In the last months suddenly we started to get an unusually high amount of orders from factories intending to use our products in manufacturing, while until then all we ever got was orders for replacements parts and maintenance.
Since we are a store, not a manufacturer, our prices aren't lowest as possible... so we are very confused about why Toyota/Honda and others for example, called us wanting parts, instead of calling our supplier, since we know they have their contact anyway (I won't say who it was but for example one time a manufacturer asked us if a product would help them, and asked us to design something for their production line... we did, then they ordered the product from our supplier and never paid us anything for all "free" engineering work we did for them).
So now I can guess what happened: car industry cancelled non-semiconductor orders too, their slots got sold, and now they want it back... so my store that tends to have higher stock than others keep getting new clients willing to pay through the nose to have parts because our own supplier doesn't have them in stock and can't deliver any in short term...
The thing about a car is that BOMs are monstrously large.
If you have 1000 ICs in a car, and one of them is missing, that's a $100,000 car you can't sell.
Car industry haphazardly buying out last stocks of ICs will not help them work around that "one missing chip" problem, and production lines are potentially stuck for many more months.
The panic was undue, well, or best say of no use. The are screwed, but its of no use for them to hope for some desperate moves improving the situation now if they can't assure 100% availability of all, and every component on their BOM.
I have few buddies who went to work on an ECU for MTU/Siemens. They ran exactly into that when their companies went for a complete redesign of their ECU to run on consumer STM32. They had hopes of that such old 180nm-130nm CMOS chips easily tolerating around 130C°, they did tests, it worked fine, and then they ran into undocumented high temperature protection kicking in on a slightly newer chip revision, but they already bough few millions of them, and other ICs for a new design in inventories.
Dude, I ended up buying like 10 or 15 blue pills a year or two ago with genuine STM32 chips for like $2/each. If they are clones, they have all the peripherals and complete ram/flash from my testing.
You write at the beginning and end of your post that the car industry lied to their vendors, but then you also write they got it badly wrong, which means they were incorrect about their projections of demand. Surely it can't be both, and it sounds like the car industry did not lie, but simply were wrong about their predictions for the future.
Or am I misunderstanding?
Nevertheless, thanks for providing context for the whole situation.
The car industry got their demand planning wrong, so they lied to their vendors that they wouldn't need that inventory. They then turned around and went "hey actually screw that we want the inventory after all", making their previous promise that they wouldn't buy it a lie.
I would not characterize that as a lie, which typically involves an intent to deceive.
Rather, the only people with intent to deceive might be the chip vendors. They either sold chip capacity that the automakers still had the legal rights to, or they reneged on sales to non automakers to resume supplying to automakers.
No, the chip vendors had their orders cancelled by ford/gm/etc., so they cancelled their orders with the fabs, who then sold the fab time to other chip vendors for different products.
Why would you build chips you know you can't sell?
Also, people talk about how big of an impact car manufacturing is to the economy and while that is true, they aren't chip manufacturers biggest customers. Apple spends more per year on semiconductors than the ENTIRE auto industry. To top that off the automotive ICs aren't high margin stuff, so if I was a fab or chip vendor, I would be focused on higher margin stuff.
The blame for this is SOLELY on the auto manufacturers.
The most surprising thing about this story is the apparently massive number of people who didn't have cars, relied entirely on public transport and then decided to buy cars? How certain are you of that analysis? Surely that is a rather rare case, as the typical public-transit-only person tends to live in a city where parking space is at a premium. Whole tower blocks cannot easily switch from transit to cars overnight just because they prefer it.
Or is this the mass exodus from American cities I keep reading about? They're buying cars because they're moving from NY to Florida, that sort of thing?
It could be both. I imagine public transit is something people would try to avoid in a pandemic. When I lived in NYC I always thought about getting a car but it never made sense to spend that money. Bring in COVID and that might have done it.
My running theory is that manufacturers and retailers continue to underprice their goods,likely because they expect to eventually return to normal. This ends up effectively hiding inflation.
Many buyers have more discretionary cash than usual due to reduced spending from lockdowns. Demand is therefore increased on the things people can and want to buy. Then resellers/scalpers see the margin and buy up inventory to arbitrage. The scarcity compounds.
Just in time delivery has its downfalls when vendors have issues, but this is a known trade-off to that model. If a company keeps inventory on hand, then they have the worry of owning excess parts that they can never use and is a sink cost later. So these companies can lose money with either, it just depends on the circumstances that cause it.
I'd be interested to learn more about the order cancellation you were saying that fabs were doing to non car manufacturers. Shouldn't contracts prevent that sort of behavior (without proper compensation)? And if they did that, they are burning bridges that would make people less likely to do business with them in the future. Maybe the car chip business is enough money to warrant such moves, but seems potentially sort sighted depending on exactly all what you said happened.
> Shouldn't contracts prevent that sort of behavior (without proper compensation)?
That depends on how big the customer/order is, no? If (as a manufacturer) you could get away with not having it in a contract, why include it?
Similarly, instead of actually cancelling they could also "lie". Say that there are capacity issues, etc. That the capacity issue is mostly because they resold the existing manufacturing capacity, is left out.
As with all things, an optimization (e.g. agile, JIT/lean manufacturing) is predicated on certain assumptions. In this instance, "we exist in an ecosystem where our suppliers can rapidly respond to volatile demand from us."
If the assumptions fail, then it's not a failure of the optimization, but an incorrect application of it. Maybe we don't currently live in a world in which elastic semiconductor availability at scale can be assumed.
When I first studied about JIT in college, I asked my professor: Doesn't this manufacturing methodology lack fault-tolerance and isn't it deeply susceptible to natural calamities or economic failures ? He replied that it was proven, adaptable and reliable. Other suppliers would come in if some suppliers failed.
JIT (aka lean manufacturing) today has now permeated pretty much most of the global supply chain. Human civilisation is far more fragile at the moment than in the last few. hundred years.
The true argument should have been that it's more efficient (for everyone involved). Which would have invited the honest debate between the relative merits of resilience vs efficiency.
That reminds me of Taleb's idea of comparing certain investment strategies to "picking up pennies in front of a steam roller" (or insuring mortgage backed securities). It's easy money until it isn't. But when it isn't, it is really bad.
>Human civilisation is far more fragile at the moment than in the last few. hundred years.
It was less fragile when there were famines all the time?
I don't understand how people can repeat this rhetoric when it seems obvious to me that however fragile it is, it's less so than in all of history. Even if things rapidly get much, much worse, it wouldn't change my opinion.
How do you think we would determine which is the correct perspective?
There's a distinction between technology level and time disposition, I think.
Compared to a self-sufficient small-scale agrarian society (say, 10th century Europe), what would have caused famine in their time would not for us.
At the same time, we allocate our time differently than they did -- few of us actually farm for ourselves.
If we allocated time more similarly + applied current technology, it'd be pretty hard for people to starve (between improved long term food storage, GMO crop yields, and environment mitigation).
Side note: the always educational Bret Devereaux lays out a solid argument for why famines were the result of an underdeveloped monetary and trade system, that led to fragile choices being optimal for individual farmers. [0]
Nearly all industries have moved to Just-In-Time to reduce costs, at the cost of making the entire economy more fragile. We saw this highlighted in the pandemic as suddenly there were no reserves of hospital capacity due to JIT optimization.
I think "lying" is pretty harsh. The pandemic caused many, many people and companies to have trouble predicting future demand. To me, it's a miracle that the American economy has kept going as strong as it has. I know it's very important for car manufacturers to predict future demand, and I know it's an extreme frustration for many people whose logistics are screwed up, but really, it shouldn't be a surprise that the car manufacturers mispredicted the impact of a once-in-a-lifetime pandemic.
You left out the AKM fire in October, which caused some essential production to get offloaded to Renesas. I think it may actually have gone to the same Naka plant that had the more recent fire.
> because they have a religious aversion to keeping any stock
In the US people love to keep huge numbers of cars in stock. Apparently most people buy from stock there, and almost nobody does JIT, for some reason. So this isn’t the case everywhere.
If you're talking about dealers, they're very different from the manufacturers. Pretty much every car has someone who's bought it already (whether that be an end consumer or a dealer) by the time it rolls off the line.
Top comment. Really appreciate this analysis. I think there is a tangential question to be asked: are consumers really demanding "software defined cars" in the first place?
Even that is not exactly the same. Cells, or logic gates per mm° is too rather ambiguous, as different implementations of same logic families do things differently, and the metric gets even wronger for different logic families.
The best I believe will be a whole cell library metric using some lowest common denominator blocks like registers, adders, bus pieces etc
I'm going to be that person and say it doesn't matter. Either you're a consumer who really only cares about a finished product's performance or you're a chip designer who has the time and money to research past the nm number.
All us armchair semiconductor fabrication experts can have a good old time arguing ove who's 2nm is better. The real world will continuing doing it's thing.
*I know this is hacker news and some of you might be real experts. I know nothing. Please don't feel disrespected.
You're basically right. Nobody doing actual chip design cares about process nm numbers. Process selection done right is a fairly tedious comparison of dynamic power consumption, static power consumption, achievable operating frequencies, wafer cost, SRAM density, IP availability, process maturity/yeild, fab capacity/ availability and about a dozen other factors.
There will be a large-ish excel spreadsheet somewhere to do the comparison.
The nm number is for press releases and non-technical investors.
Yes, that's a good analogy, although the nm number by itself isn't the full version e.g for 28nm alone, TSMC had 28HP, 28HPL, 28HPM, 28HPC and 28HPC+ variants, of which only HPM and HPC (but not HPC+!) are at all compatible.
28HPC+ and 28HP have very different performance, even if they're both "28nm".
And if you really want to specify a process you also need to know the metal stack (lots of options there), Vt selection(s) (most processes have 2-5 options) & high-voltage device support option.
Honestly, SRAM bits per square mm seems like a perfectly fine metric for me. While I recognize that the memory-to-logic density varies, sometimes significantly, between processes, (a) SRAM cells are optimized to within an inch of their lives ("within a nm of their lives"?), including often being revised in minor process updates, so accurately reflect the capabilities of the process; (b) many chips, including the ones I personally tend to be interested in (processors, etc) are SRAM-heavy, and get more so each generation; (c) it's a single number with some connection to reality, which is better than both a single number with limited connection to reality ("process"), or a whole suite of numbers that require deep interpretation (cell library metrics).
And it is exactly because SRAM gets so intensely optimised with each generation, I did not want to list it. SRAM density, and performance is not characteristic of how the rest of logic will perform no other logic component get squeezed so much.
Their 20B investment into their Arizona campus has me convinced. They're doing the right thing (if you live in the US), my 5900X is pretty unreliable and I never had issues with my Intel systems. The stars have aligned so I'm going to buy their chips from here on out.
Ever since installing a BIOS with AGESA 1.2.0.0 (and now 1.2.0.1) my machine locks up on multicore loads. It also will lock up while navigating in the BIOS. My original BIOS that supported Zen3 was 1.1.8.0 and I had no problems whatsoever, and I was tinkering with it then far more than now, since the build was new and I was excited for it. The problems started immediately upon the newer BIOS being installed, and they put a read/write lock in the 1.2.0.0 release so I can't flash back.
I normally don't update unless I need to, but they advertised so many fixes that I felt compelled. It's reproducible, I can lock it up immediately upon starting Prime95, and it'll occasionally crash running CBR20. I would hope over time, probably take a year or two, they'll get this sorted out.. but I may just order an i9-10850K and a Z490, and be done with this. I run bone stock UEFI BIOS settings, no PBO or other overclocking.
I run a small business off this machine, and it has to be reliable, performance doesn't matter if you have stability issues.
>You should ask how every single Fabless Semiconductor company has failed with their demand, supply chain analysis and projection.
That's an oversimplification. As lead times increase, overbooking increases and that gives rise to further extension of lead time. Delays also become more prevalent at high utilization because there's no capacity in reserve. It's very difficult to predict where demand will go in a year of totally abnormal market behavior and unpredictable helicopter money.
In other words, fabless chipmakers are asked to do the impossible and some of them, predictably, failed to accomplish that task. Imagine asking AMD in April of 2020 to predict how many PS5s will sell during Christmas.
2. Since leading edge node is forever increasing in cost, the next 3 years, i.e 3nm in 2022 and 2nm in 2024 are expected to be higher. ~$100B investment aligned with their initial plan / trend / target anyway.
3. So really the major news is stopping ( or to be precise, delaying ) price reduction. Which is unusual but understandable given the current demand situation. And No, it is not TSMC's fault. You should ask how every single Fabless Semiconductor company has failed with their demand, supply chain analysis and projection. Although one could argue it is not their fault either, since their client gave wrong projection. The only company that is not affected is possibly Apple.
4. There has been mounting pressure from investors, politicians and MSM media on Intel IDM 2.0, and supply problems. This announcement feels more like addressing those concerns.
5. Remember both Samsung and Intel are expanding capacity. And even GF and many other smaller players. It took DRAM and NAND three years to catch up with demand ( and then over supply ). Which in hindsight is pretty damn impressive. Although during the ~2016 - 2019 everyone felt awful.