TSMC tells vendors to delay chip equipment deliveries(reuters.com)
reuters.com
TSMC tells vendors to delay chip equipment deliveries
https://www.reuters.com/technology/tsmc-tells-vendors-delay-chip-equipment-deliveries-sources-2023-09-15/
53 comments
> Six months ago it was a global chip shortage that had supply chains backed up around the planet.
chip != chip. TSMC manufactures hundreds of products, check out slide 4 of the last quarterly results [1].
"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.
[1] https://investor.tsmc.com/english/encrypt/files/encrypt_file...
chip != chip. TSMC manufactures hundreds of products, check out slide 4 of the last quarterly results [1].
"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.
[1] https://investor.tsmc.com/english/encrypt/files/encrypt_file...
> The recent chip shortage was mostly from 28nm or above.
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 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.
With 3nm being the last of the FinFETS will it end up in a similar situation in a decade? Will it be used for a lot of applications like 28nm is now for a long time?
It comes down to what the cost delta ends up being between FinFETs and GAAFETs, and if that's where the discontinuity is. I suspect that, even though EUV is "solved" now, the solutions are so expensive and sketchy that the DUV/EUV break is more likely to be important, with the final DUV processes being very long-lived. Unclear where this will end up exactly; I'd expect to see mature 7 nm DUV processes, but it's possible that the economics will either favor maturation of DUV + one or two EUV layers, or that a pure DUV 10 nm process with less multi-patterning ends up maturing so much cheaper that that wins out.
I hear there are developments of half-node reductions in 28nm processes to get down to "22nm" mostly with wiring density. But your point stands, it's the last planar node for mixed signal.
Yep, there will be continuing improvements to 28 nm as a long-lived node; 45 nm was a similarly long-lived node (as are 90 nm and 180 nm, still seeing plenty of use), and today's 40 nm is a heck of a lot more polished than 45 nm was at release, while still allowing trivial design porting.
Could you mention where these chips from older mature nodes are used and what they look like(in terms of form factor / development environment), for context? I understand some aspects of chips being used for automobiles , but aren't most compute oriented chips based off ARM designs and such which will be on the newer nodes?
Microchip just launched the SAM9X60 like three years ago.
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.
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.
Radio chips used in iot devices. Bluetooth, wifi, zwave, zigbee etc. usually rely on the lager node size. Designs are produced and sometimes supported for decades.
Most of the chips I interact with on a daily basis are 45 nm, some are 28 nm. Currently working automotive, same in aerospace. A lot of this is because embedded flash generally trails several years behind general availability of a node; so if you're designing a µC or similar with embedded flash, you're starting a few years behind; and then you're guaranteeing a 10 or 15 year supply chain, so on average a production part is on a ~10 year old node (but development activities are biased towards the front of this, since you often need 7 years of remaining supply chain guarantees after start of production). Another factor is that it simply takes time to validate nodes. Automotive nodes trail general use nodes; SEU data takes a few years to gather. Even for non-embedded-flash parts, this can cause a similar delay.
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.
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.
Embedded Sensing, Most Automotive, DisplayDrivers are usually mixed signal solutions that don't benefit from process shrinks as much as others. For a long time costs continued falling on legacy (55/90/180) processes, but that has stopped and during Pandemic, and prices shot way up. Some of it is transistor noise scaling (for ADCs/DACs), some is lack of Embedded Flash (as mentioned, though TSMC is now sampling at 28nm), and some die are already so small they're pad limited. In any case a process shrink has significant design costs (and testing/requalification) so if you're not going to make it up with reduced prices (customer) and higher margins (supplier) it's unlikely to happen.
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.
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.
TSMC is diversified, they are basically an index. Everything from cars to medical devices to household electronics to AI uses their chips. This is why Nvidia can soar while TSMC barely moves.
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.
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.
Apple buying every N3B wafer isn't enough to match the orders they placed. They had originally anticipated other customers on N3B too. But only Apple has an N3B design ready for manufacture. Intel - the only other major N3B customer - plans to use it for Arrow Lake S which is a late 2024 product. Every other major customer decided to wait for N3E because N3B is too expensive. TSMC therefore wants to push out the acquisition of additional machines until next year when N3E is ready. It's also the reason TSMC pulled in N3E a quarter or two.
The chip shortage was about common, low-cost chips made on older (i.e. larger), mature process nodes that TSMC does less business in.
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.
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.
https://en.wikipedia.org/wiki/Bullwhip_effect My naïve assumption is that it has to do with this?
That one knows no mercy.
>as the world's top contract chipmaker grows increasingly nervous about customer demand, two sources familiar with the matter said.
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.
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.
Nvidia‘s H100 chips may sell for a lot of money, but in terms of wafers, its just around 10% of volume compared to their gaming chips.
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.
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.
It's a typical boom/bust cycle, isn't it? Plus tech is going through a recession, there's less money tossed around, inflation is cray cray, people have less disposable income, etc.
I would imagine it has to do with the industry wide slump in demand that appeared earlier in the year.
> 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.
https://techcrunch.com/2023/04/06/samsung-cuts-memory-chip-p...
I think it took Samsung most of a year of production cuts to clear out excess inventory.
> 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.
https://techcrunch.com/2023/04/06/samsung-cuts-memory-chip-p...
I think it took Samsung most of a year of production cuts to clear out excess inventory.
I'm equally confused. I came here to say the same. One thing I'm not confident about is my understanding of this article. Is it, in fact, saying that they're slowing down production because they fear low consumer demand? I may have missed it in the article but I don't feel like it ever clearly states it like that.
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
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
Recession kicks in. Huawei able to do 7nm...but the bulk of chip sales are at 14nm+. China in 2022 end has successfully dominated that market near 80%. The pre-order for Huawei mate 60 already reaching 10m+ just a week ago and projected to reach 60m by 2024 end. Meanwhile China pretty much banned iPhone "subtlely" by forbiding their government servants using that (which can translate to around 100m people if you read between the line meaning ccp members). The euphoric news sensations are mostly CIA Uncle Sam propaganda. If you go into labor market or ask any HR the picture aren't as rosy as what you read on the media. Also, I have very reliable source China already at 3nm just yield is very low for mass market production. They should be able to reach that around Dec 2024 though TSMC probably will announce their 2nm. So 2025 is when China on par with TSMC making the war inevitable around 2026 as many fengshui masters calculated large chaotic event in that year.
LOL you get voted down because mentioning those 3 words of american 'security organisations'
The article just obfuscates the news and hides a key detail at the end. China is a huge market for everything TSMC makes. China just showed this month that they can manufacture a near current generation flagship SoC using 90%+ domestic supply chains. This makes a big revision on future outlooks in China and ripples through demand forecast and capacity planning.
Does this mean that China is buying a bunch of ASML EUV machines?
China won't be able to unless they develop the tech domestically.
Right that's what I thought, so how are they building a "near current generation flagship SoC"?
The near current generation chip being talked about is the one in the latest Huawei phones, made in collaboration with SMIC. I believe SMIC have older machines that can still be pushed to get to close to state of the art. It is newer machines that are under sanction and can't be sold to Chinese firms.
>China just showed this month that they can manufacture a near current generation flagship SoC using 90%+ domestic supply chains
I missed the news, any links on the subject?
I missed the news, any links on the subject?
>In the middle of a revolution in AI? Something doesn't add up here.
Middle? hasn't the hype died already?
Middle? hasn't the hype died already?
I can't speak for the mobile phone market, but computer makers tend to be very cyclical based on the business cycle.
I think we are likely seeing this to some degree.
I think we are likely seeing this to some degree.
I worked in semiconductor manufacturing, and I've been exposed to a lot of other industries. It's extremely volatile. Not quite like oil and gas production, but still enough to make your head spin.
" two sources familiar with the matter said." I don't really get how an article can be built off that. Isn't that completely insubstantial on the face of it? Are we supposed to just trust that some reuters journalist is sourcing this info from a credible primary source?
This would all be information that would be under NDA and certainly not to be discussed publicly. Industry insiders leaking information under condition of anonymity is the only reason we have any idea about what goes on behind the carefully curated PR from major corporations. Reputable journalists will confirm with multiple sources and only trust reputable sources. If you don't trust the publisher or journalist, check their previous releases to see how accurate their sources tend to be.
that's usually what journalists do.
if you're concerned about Reuters credibility (you should be, for unrelated reasons), check this particular journalist's history.
if you're concerned about Reuters credibility (you should be, for unrelated reasons), check this particular journalist's history.
Too lazy to look it up, but what exactly is it about said journalist? Good, bad? Because Reuters is one of the best journalistic sources out there.
Edit: Two sources confirming stuff is not just how journalism works, but also intelligence services.
Edit: Two sources confirming stuff is not just how journalism works, but also intelligence services.
yes, that is how journalism works.
Last week's news, submitted a few times (https://news.ycombinator.com/item?id=37528494 https://news.ycombinator.com/item?id=37547828).... anything new since?
Mon, Sep 18th, Taiwan shares tumble to end below 16,700 points as TSMC hit hard
https://focustaiwan.tw/business/202309180011
Mon, Sep 18th, Taiwan shares tumble to end below 16,700 points as TSMC hit hard
https://focustaiwan.tw/business/202309180011
The costs of building and idling Phoenix, AZ sites must be nontrivial. Without booked orders, buildout would be a waste of money.
It makes sense - lower sales from the iPhone 15 than expected - recession in lots of places, increased competition from China.
The iPhone 15 hasn't been released yet. Having expectations of high sales at this point might be premature...
That new equipment is mostly for making iPhones.
Any info if SAIC 7nm is having any impact? Or is it not feasible for non Huawei?
Latest rumor is expected sales of around 20m units - should cut into the iPhone 15 sales in China.
if you read a lot of chinese news, they said 2 biggest reasons: 1. many high skill engineers out from tsmc, moved to huawei and simc. 2. tsmc fear iphone 15 sales are falling especially in china
Sounds like market manipulation
https://arstechnica.com/gadgets/2023/08/report-apple-is-savi...