You're fabbing it wrong: Chip shortages due to lack of investment in the right factories, says IDC

[soAsian]

IDC: Chip shortages due to lack of investment in right fabs

Not enough money going into 40nm+ process nodes

www.theregister.com

why not move up the technology to 28nm or 14nm?

why 40nm? can't car vendors move up with technology advancement?

 

[Daniel Nenni]

IDC: Chip shortages due to lack of investment in right fabs

Not enough money going into 40nm+ process nodes

www.theregister.com

why not move up the technology to 28nm or 14nm?
why 40nm? can't car vendors move up with technology advancement?

Fab space is tight but we are not on allocation yet. And yes there is huge amounts of FinFET manufacturing capacity and much more being built every year. It really is the car companies fault but the end result will be a much stronger and optimized supply chain. The trend I'm seeing, and this is confirmed by the EDA folks, is that system companies are doing their own chips and this now includes the car companies. These chips will in fact be FinFET based thus saving power and getting better performance. Unfortunately it will take a couple of years before we see this in production but it will be well worth the wait. Until then we will continue the chip shortage narrative and build more FinFET fabs until there is a glut, absolutely.

And by the way, this scenario plays to TSMC's strength of building the best chips cheaper with a massive ecosystem to speed designs along so if I were the other foundries I would be concerned.

 

[jms_embedded]

why 40nm? can't car vendors move up with technology advancement?

Until the automotive companies or Tier 1 auto suppliers get into the fabless business, it's the main auto IC manufacturers to whom you need to ask this question.

There are lots of reasons why small node sizes may be less attractive for automotive/industrial ICs than for processors/SOCs.

- analog doesn't shrink, doesn't (I think; i'd love to see corroboration/refuting evidence) lead itself as easily to EDA software as digital
- low cost targets (high development costs: suppose your volumes are high enough that you can amortize them to 10c hit per chip. 10c hit on a $50 or $10 processor isn't a huge deal but 10c hit on a $1 microcontroller is)
- environmental specs: high temp range
- tech specs: leakage current, support for embedded flash

 

[jms_embedded]

The trend I'm seeing, and this is confirmed by the EDA folks, is that system companies are doing their own chips and this now includes the car companies. These chips will in fact be FinFET based thus saving power and getting better performance.

I can believe this for higher-end digital chips to support ADAS / zone controllers / infotainment, but not for lower-end chips or mixed-signal controllers or analog chips.

What's the expected design/one-time-mask-set cost for, say, a 16nm chip these days? $10M? (I'm just making a very uneducated guess here; semiengineering showed an IBS graph of $106M at 16nm but that's for an "advanced design" and includes software costs; the auto companies aren't going to spend more on software complexity just because they make a chip on 16nm instead of 40nm) Toyota allegedly sells around 10 million vehicles per year. Do the math, see what kind of sales price the price-per-transistor is more important than the one-time costs.

Someone (not me) could do a nice PhD thesis by taking a few totaled cars from different vendors in the latest model year, and doing some teardowns to measure die sizes and feature sizes, and trying to apply some cost modeling to look for some cost-saving opportunities in the future.

 

[hist78]

I can believe this for higher-end digital chips to support ADAS / zone controllers / infotainment, but not for lower-end chips or mixed-signal controllers or analog chips.

What's the expected design/one-time-mask-set cost for, say, a 16nm chip these days? $10M? (I'm just making a very uneducated guess here; semiengineering showed an IBS graph of $106M at 16nm but that's for an "advanced design" and includes software costs; the auto companies aren't going to spend more on software complexity just because they make a chip on 16nm instead of 40nm) Toyota allegedly sells around 10 million vehicles per year. Do the math, see what kind of sales price the price-per-transistor is more important than the one-time costs.

Someone (not me) could do a nice PhD thesis by taking a few totaled cars from different vendors in the latest model year, and doing some teardowns to measure die sizes and feature sizes, and trying to apply some cost modeling to look for some cost-saving opportunities in the future.

Automobile manufacturers can be penny wise or lacking logical thinking ability from time to time. I can't forget GM's ignition switch scandal that caused at least 124 people's deaths.

He uncovered GM crisis, took on automaker when NHTSA didn't

The lawyer of a car crash lawsuit that lead to one of the biggest safety recalls in history talks about how best to keep cars safe for consumers.

www.freep.com

 

[lilo777]

IDC: Chip shortages due to lack of investment in right fabs

Not enough money going into 40nm+ process nodes

www.theregister.com

why not move up the technology to 28nm or 14nm?

why 40nm? can't car vendors move up with technology advancement?

It's hard to get the full picture. Was there an absolute or a relative supply shortage? Daniel had a post where he pointed out that there was a step-function surge in demand for semiconductor products during Covid. Who could have predicted it? Now, even if there was a spare capacity for, say, 14nm it could not have been utilized without a delay. Someone had to redesign those chips which would require additional investments which might be difficult to justify given a very uncertain market picture. Everything is much more obvious in hindsight

 

[jms_embedded]

It's hard to get the full picture. Was there an absolute or a relative supply shortage? Daniel had a post where he pointed out that there was a step-function surge in demand for semiconductor products during Covid. Who could have predicted it? Now, even if there was a spare capacity for, say, 14nm it could not have been utilized without a delay. Someone had to redesign those chips which would require additional investments which might be difficult to justify given a very uncertain market picture. Everything is much more obvious in hindsight

Let's ask the question a bit differently. If someone (automotive IC manufacturers / Tier 1 automotive suppliers / automotive companies) is going to design new ICs with die that is manufactured at the foundries, which process node should they use, to maximize their profits, while constraining an appropriate level of functionality/reliability, and achieving certain supply chain goals (e.g. they want to be able to obtain new parts from this design for the next N years), given the situation today? (no hindsight from Jan 31 2022)

And what questions do we need to answer in order to make this decision? (e.g. type of IC, some kind of complexity metric, expected annual volumes, environmental requirements, number of years N, etc.)

 

[Daniel Nenni]

Let's ask the question a bit differently. If someone (automotive IC manufacturers / Tier 1 automotive suppliers / automotive companies) is going to design new ICs with die that is manufactured at the foundries, which process node should they use, to maximize their profits, while constraining an appropriate level of functionality/reliability, and achieving certain supply chain goals (e.g. they want to be able to obtain new parts from this design for the next N years), given the situation today? (no hindsight from Jan 31 2022)

And what questions do we need to answer in order to make this decision? (e.g. type of IC, some kind of complexity metric, expected annual volumes, environmental requirements, number of years N, etc.)

TSMC 28nm. Cheap and easy to design to and plenty of capacity, the biggest CMOS node for TSMC.

 

[Paul2]

why 40nm? can't car vendors move up with technology advancement?

Too expensive. Anything 1m/year< will never leave 130nm-180nm because of lot sizes.

For most of mortals, a tapeout on 200mm is the only thing they can hope for in their lifetime.

Design of physical IP on immersion nodes is too expensive too. It's not what 10 years old pirated Cadence can do, nor what Indian, or Bangladeshi VHDL outsourcing shops specialise on.

 

[mgoldsmith1979]

TSMC 28nm. Cheap and easy to design to and plenty of capacity, the biggest CMOS node for TSMC.

The challenge has been eNVM on these nodes; 40nm eFlash qual for Auto was in 2018 (https://www.tsmc.com/english/dedicatedFoundry/technology/specialty/eflash), and the cycle time for auto components is ~3 years for the platform certification and then a ~10 year support lifetime commit from the manufacturer. So anyone that was designing a "new car" platform in 2019 would not have considered anything less than 40nm as appropriate if including eNVM. Renesas worked on porting their MONOS Flash to 28nm, and that was supposed to be ramping in 2020 (https://www.renesas.com/us/en/about...boration-next-generation-green-and-autonomous) - no idea if that panned out. N28/22 for Wireless / IoT applications with eNVM was focused on STT-MRAM or RRAM rather than eFlash, and the former has challenges with magnetics (so not going into any ECU), and the latter performance / endurance may not meet the spec of the part (incl temp range). If one were designing a new part in 2021/22 expected to launch in '24-'25 timeframe, then I would agree that 28nm is a good option, but their capacity is not limitless either - that's why they are expanding into Japan, to support the Sony ISP business.

Unrelated, but interesting was a paper from TSMC at VLSI in 2019 (I think) that showed their Wafer-on-Wafer attach putting a N40 die with eFlash atop an N16 logic die, specifically to bring that eNVM capability to a more advanced node, b/c the only current eNVMs for FinFET are still STT/SOT or RRAM based.

 

[hist78]

Too expensive. Anything 1m/year< will never leave 130nm-180nm because of lot sizes.

For most of mortals, a tapeout on 200mm is the only thing they can hope for in their lifetime.

Design of physical IP on immersion nodes is too expensive too. It's not what 10 years old pirated Cadence can do, nor what Indian, or Bangladeshi VHDL outsourcing shops specialise on.

My question is that if a shortage of cheap $3 semiconductor parts can cause a assembly line to shut down for a month or cause a car company to lose billions dollar in sales, will that change the "too expensive" calculation?

 

[benb]

Too expensive. Anything 1m/year< will never leave 130nm-180nm because of lot sizes.

For most of mortals, a tapeout on 200mm is the only thing they can hope for in their lifetime.

Design of physical IP on immersion nodes is too expensive too. It's not what 10 years old pirated Cadence can do, nor what Indian, or Bangladeshi VHDL outsourcing shops specialise on.

This sounds correct to me.
Here in Austin, NXP and Infineon fabs are hiring, in old 200mm buildings.

So to summarize Automotive constraints:
200mm
65nm node (this is the last dry 193nm litho node)
Dry 193nm photolithography
Aluminum wiring?

What the problems are:
You can't expand 200mm, no tools are available, and your fab is probably full
200mm capa has been stagnant for decades
Automotive parts require 200mm due to cost and batch size, and can't move to 300mm

Question
I get why auto ICs are exposed to 200mm stagnant capa, but it's not new, it's decades-old. At some point, and 2021 was the point, the 200mm capa was going to run out. What was the plan then?

 

[jms_embedded]

My question is that if a shortage of cheap $3 semiconductor parts can cause a assembly line to shut down for a month or cause a car company to lose billions dollar in sales, will that change the "too expensive" calculation?

Well stated! It's quite the conundrum, an illogical one, perhaps like this case:

- Dad tries to fix the leaky faucet because he doesn't want to call a plumber for $300
- The leak gets worse
- Dad strips the thread on the pipe, rigs up something temporary, heads to the hardware store hoping he will be able to find a solution, or that the problem will fix itself

 

[jms_embedded]

Question
I get why auto ICs are exposed to 200mm stagnant capa, but it's not new, it's decades-old. At some point, and 2021 was the point, the 200mm capa was going to run out. What was the plan then?

I'd state it slightly differently: somewhere in the 2022 - 2025 range perhaps was when the capacity was going to run out, but because of the large disruption in 2021, we ran into capacity problems early. (signal processing / control systems analogy: impulse disturbance causes a controller to hit saturation early when it was close to saturation anyway)

 

[jms_embedded]

TSMC 28nm. Cheap and easy to design to and plenty of capacity, the biggest CMOS node for TSMC.

I believe this for high-volume ICs that aren't low-cost. What kind of mask set costs (rough order of magnitude) are we talking about for 28nm?

IDK what the semiconductor companies need to justify producing a new IC in the $1-$2 range (high volume cost to customer) but I suspect the one-time costs for design and production need to be amortized over expected volume to be less than 20 cents*. (For required gross margin / operating margin, look at recent earnings reports.)

(*yes I work for a certain semiconductor manufacturer, no I do not work in IC design or production; this is a vague guess given that I don't want to find out a real number and deal with confidentiality constraints.)

 

[Paul2]

My question is that if a shortage of cheap $3 semiconductor parts can cause a assembly line to shut down for a month or cause a car company to lose billions dollar in sales, will that change the "too expensive" calculation?

Point 1. The global 200mm capacity is grossly inadequate to meet demand
Point 2. Even with sky high 200mm service prices, an entry level 300mm will cost more per wafer, cost more in NREs, and need longer lead times than anything on 200mm.

20 years since, and 300mm service is still nowhere near of supplanting 200mm on cost.

 

[count]

The legacy automakers have product architectures that are currently locked into these older nodes. You have a whole bunch of different parts from different suppliers with different chips in them. EV startups like Tesla don't have this issue. Tesla built it's own supply chain, and was able design chips that consolidate multiple functions. So a Tesla will have fewer, higher value, chips in it, and those are being fabricated on newer nodes.

 

[jms_embedded]

Tesla built it's own supply chain, and was able design chips that consolidate multiple functions. So a Tesla will have fewer, higher value, chips in it, and those are being fabricated on newer nodes.

Makes sense; has anyone done a teardown of a Tesla to go over IC content?

Tesla still has some supply chain issues though; their recent earnings call specifically cited voltage references and oscillators among the "boring" chips that caused them pain and suffering. But these seem like good candidates to integrate into more complex chips. (At least, voltage references are. I don't know enough about the nuances of clock generation to know whether it's cost-effective to integrate a high-quality oscillator that works over the full temperature range into an IC. Oscillators are tricky at the temperature extremes.)

hmmm... I found these, but not sure exactly how to pick out the nuggets of information:

- https://www.eetasia.com/teslas-hardware-retrofits-for-model-3/ (US website https://www.eetimes.com/a-tesla-model-3-tear-down-after-a-hardware-retrofit/ annoyingly divided into 9 pages)

- https://www.rolandberger.com/publications/publication_pdf/roland_berger_computer_on_wheels.pdf

- https://www.motortrend.com/features/tesla-model-3-teardown-deconstructed-3/ -- not much technical commentary, but I still see circuit boards with lots of little ICs in addition to the big ICs. If they change to a more SOC approach maybe they can get the BOM count down and reduce # of "mature node" parts, but I don't see it yet.

 

[benb]

So the world can meet lets say 80% of the auto IC demand, and it is a hard limit the way bitcoin has a hard limit. I'm not 100% convinced of this but it seems like a working theory, and I can easily picture how it happened, like bankrupcy, a little bit at a time then all at once.
There's this concept of a "black start" that Texans have been learning about since Feb 15 last year. It's when all the power plants trip, and there is no power to even start the thing that starts the power plant. It takes 2-3 months to start one little thing, to start a bigger thing, to start a still bigger thing. Meanwhile, 25 million Texans have to walk to safety, and maybe half freeze.
Looks kind of like that for 20% of the auto industry.

 

[jms_embedded]

So the world can meet lets say 80% of the auto IC demand, and it is a hard limit the way bitcoin has a hard limit. I'm not 100% convinced of this but it seems like a working theory, and I can easily picture how it happened, like bankrupcy, a little bit at a time then all at once.
There's this concept of a "black start" that Texans have been learning about since Feb 15 last year. It's when all the power plants trip, and there is no power to even start the thing that starts the power plant. It takes 2-3 months to start one little thing, to start a bigger thing, to start a still bigger thing. Meanwhile, 25 million Texans have to walk to safety, and maybe half freeze.
Looks kind of like that for 20% of the auto industry.

Where is Willy Shih reporting on this when you need him?

edit: you got me curious...

If The Texas Power Grid Had Gone Down, It Would Need A 'Black Start.' How Long Would That Take?

It’s happened before in other parts of the country, but in Texas it would play out much differently. That’s because Texas has its own power grid with minimal connections to neighboring grids.

www.kut.org

<whistles slowly, walks away>