chip != chip. TSMC manufactures hundreds of products, check out slide 4 of the last quarterly results .
"AI" (on 5nm or below) is currently only a smaller share of total capacity. 7nm was only around 50% utilization, compared to 110% a year ago.
The recent chip shortage was mostly from 28nm or above. Think automotive, and the like.
The inventory correction happening now is from a strong reduction in smartphone business, and HPC is also down thanks to PC sales declining 10-15% this year.
And there is nothing surprising about this, TSMC has been guiding this inventory correction for more than a year now.
As a reminder to others, 28 nm was the last planar process; there's a discontinuity in cost moving below 28 nm to FinFET processes. For applications which don't need the last bit of density of power efficiency, 28 nm is likely to stay a sweet spot for a long, long time to come.
As an example, the TI Hercules is an old but by no means obsolete safety processor. It's one of the best ways to get single-core lockstep capabilities for new designs today. The original TMS570LS parts are still built on a 130 nm process, but the "newer" (not new, but also newest) TMS570LC parts are on 65 nm.
These days most of my work hours are focused on the Aurix TC3xx, which is a 40 nm (tweaked 45 nm) part. This is a multi-core safety processor with an obviously higher transistor count than the Hercules (which is why it was selected for application), matching its smaller process.
Similarly in FPGA land, the Lattice MachXO is on an 130 nm process, while the newer MachXO2 migrated to 65 nm. I know Lattice does newer parts on 28 nm (and I'm sure there's 45 nm out there), but I haven't run into them.
Certainly for more density-heavy applications newer processes (28 nm, often) have come to the forefront, and we're at the point that we're seeing cost cuts going from 90 nm to 28 nm for equivalent functionality. But a lot of these designs are pin-out limited (the silicon area is dominated by getting signals on and off chip, not by the total area of transistors), so cost very much doesn't scale with transistor density; the cost per transistor has gone down, but the cost per unit area has gone up.
One thing that is driving new development is the lack of 8in capacity and the move to 12in (300mm) due to equipment availability. Most of the installed capacity of 80nm and larger was 8in and you can only really build new 12in fabs.
600Mhz and 128MB of DDR2 RAM is plenty for many applications. It's pretty much the lowest end processor you can get that will still run a Web server / GUI for easy network dongle. Even if it's on an ancient ARMv5 it still gets the job done today.
Car and automotive is even lower end in practice. Every button is a uC whose only job is to convert button clicks into LINbus or CAN bus commands for the central computer.
No reason to go below 180nm for a lot of these chips... Well aside from cost. The older 90nm or 180nm nodes are IIRC getting more expensive as the old wafers are drying up (200mm vs modern 300mm).
A lot of cost optimized chips are 40nm to my knowledge, and others are 28nm now. TIs MSPM0 line comes to mind.
We know winter is coming. The signs are all there, the desperate juicing and enshittification- growth is running out and nobody wants to blink first.
As for the "revolution in AI", TSMC isn't going to spontaneously manufacture chips that nVidia and AMD haven't ordered, so your "something doesn't add up" should be directed to them. The same goes for Apple.
I stopped reading after that.
And as someone who knew about TSMC and ASML before 99.99999% of internet even start reporting news on TSMC. Most if not all of TSMC news are just plain BS.
There is rarely a shortage that isn’t followed by an oversupply, so in fact it would be weird if it was different this time.
> Samsung Electronics plans to cut back memory chip production as its operating profit in the first quarter of 2023 is expected to plummet about 96% from the previous year.
The global macroeconomic slowdown, memory chip oversupply and sluggish demand have hurt its profit, the world’s largest memory chip maker said in its preliminary earnings release on Friday.
I think it took Samsung most of a year of production cuts to clear out excess inventory.
Def possible I just didn't read it well but I ended my review of it still confused about whether they were worried about lack of or increase in demand
I missed the news, any links on the subject?
The enshittification of Ars is the thing that "doesnt add up for you."
Middle? hasn't the hype died already?
I think we are likely seeing this to some degree.
if you're concerned about Reuters credibility (you should be, for unrelated reasons), check this particular journalist's history.
Edit: Two sources confirming stuff is not just how journalism works, but also intelligence services.
Mon, Sep 18th, Taiwan shares tumble to end below 16,700 points as TSMC hit hard