Why Carmakers Can’t Transition to Newer Chips(jalopnik.com)
jalopnik.com
Why Carmakers Can’t Transition to Newer Chips
https://jalopnik.com/i-asked-experts-why-carmakers-cant-just-transition-to-n-1847739665
325 comments
> “We were able to substitute alternative chips, and then write the firmware in a matter of weeks,” Musk said. “It’s not just a matter of swapping out a chip; you also have to rewrite the software.”
https://www.theverge.com/2021/7/26/22595060/tesla-chip-short...
https://www.theverge.com/2021/7/26/22595060/tesla-chip-short...
Vehicle electronics have some real environmental challenges as well.
They have to function correctly
They have to function correctly
* after they've been parked outdoors in Death Valley all day.
* after they've been parked at -40 degrees for a long time.
* when doused with slush containing road salt.
* when their wiring harnesses deteriorate after a couple of decades of hard use.
* at least for a few seconds after a catastrophic crash.
I did a short stint in the automotive industry ~10 yrs ago (on the software/safety side) at one of GM’s brands.
The lack of forward-thinking was absolutely horrible, as was the constant insistence that anything new would never work (you remember the constant doomsaying about tesla?).
I got out as fast as I could, I seriously think that being in that environment might cause serious mental impairment…
So, not in the least bit surprised about this. Just nice to hear that the world is moving on whether the big auto brands wants it to or not.
Admittedly it was a long time ago, but I worked in the semi industry and this isn't how things were done. We manufactured batches of totally obsolete EOLed devices for various customers when they had sufficient volume. Devices that are in volume production (e.g. for cars) are carefully managed by people who do nothing but ensure that they're available in the right place at the right time in the right volume. In order to not have product available to meet demand either a factory has to go on fire, or the customer has to screw up their forecast.
I always thought chips were dominated by fixed capital costs of fabs. If they were dominated by variable material cost of wafers, as the article seems to imply, it wouldn't make sense that we see 90nm microcontrollers that sell for $1 and a high-end 16nm PC CPU that sell for $1000.
So the question is, what's the reason that 90nm microcontroller sells for $1?
I'm trying, and failing, to figure out an economic model that explains the market dynamics we actually observe.
If building a new 90nm fab costs $billions, almost as much as building a new 16nm fab, why does the 90nm microcontroller sell for 0.1% of the price of the 16nm Xeon?
If building a new 90nm fab costs 0.1% as much as building a new 16nm fab, why can't existing chip companies, some startup or GM themselves spend $10's of millions building a fab that can unblock $100's of millions of product, and alleviate the shortage?
So the question is, what's the reason that 90nm microcontroller sells for $1?
I'm trying, and failing, to figure out an economic model that explains the market dynamics we actually observe.
If building a new 90nm fab costs $billions, almost as much as building a new 16nm fab, why does the 90nm microcontroller sell for 0.1% of the price of the 16nm Xeon?
If building a new 90nm fab costs 0.1% as much as building a new 16nm fab, why can't existing chip companies, some startup or GM themselves spend $10's of millions building a fab that can unblock $100's of millions of product, and alleviate the shortage?
Any discussion about why you can't transition to newer chips that doesn't mention the software is incomplete.
Chips are not fungible in part because of software compatibility.
OK, so you got a newer chip, and have redesigned the board and everything to fit. Now you have to get all the old firmware running on it and validate it.
If the new chip isn't a 100% backwards compatible version of the old one, including all the peripherals, that could be a considerable effort, fraught with risk.
Chips are not fungible in part because of software compatibility.
OK, so you got a newer chip, and have redesigned the board and everything to fit. Now you have to get all the old firmware running on it and validate it.
If the new chip isn't a 100% backwards compatible version of the old one, including all the peripherals, that could be a considerable effort, fraught with risk.
When a Power PC is perfectly capable for a MARS rover missions, I dont see any reason why they must run Snapdragons.
Unless some evil company, starts ditching physical controls or starts shoving Ads into Speedometer.
And everyone follows.
Im thankful that automobile tech didn’t discover electron yet. Imagine your speedometer consuming 200mb of ram.
Im thankful that automobile tech didn’t discover electron yet. Imagine your speedometer consuming 200mb of ram.
I look forward to the cyber-punkish future (real soon now?) where there's an old beige computer case with a Pentium MMX sticker on it on someone's passenger seat with wires snaking into the engine bay. "Yeah the ECU died, so I had to hack together this replacement...".
The middle road is missed here.
There are two problems, long duration supplies of a chip on the same process, and the cost of making chips.
If the semiconductor companies can figure out an ASIC/FPGA type strategy that would allow any of the current chips in a car be produced "dynamically", then the semiconductor company could achieve their economies by just fabbing one design, and car companies could achieve longevity by getting commitments for production of that one design.
Xilinx recently made some steps in a new direction by producing what they called an "RF" SoC. Where RF stands for radio frequency. Basically it had all of the analog to digitial and digital to analog bits on the chip in addition to some FPGA fabric and some ARM AARCH64 cores. Its expensive and small quantity but I think it will turn out to be important in the long run as the vanguard of chips that are not 'type specific' at the time of manufacture.
Imagine a company that has the same basic FPGA architecture, packaged in a variety of automotive spec packages, with one-time programmability. That would reduce the number of SKUs considerably (basically by package type).
There are two problems, long duration supplies of a chip on the same process, and the cost of making chips.
If the semiconductor companies can figure out an ASIC/FPGA type strategy that would allow any of the current chips in a car be produced "dynamically", then the semiconductor company could achieve their economies by just fabbing one design, and car companies could achieve longevity by getting commitments for production of that one design.
Xilinx recently made some steps in a new direction by producing what they called an "RF" SoC. Where RF stands for radio frequency. Basically it had all of the analog to digitial and digital to analog bits on the chip in addition to some FPGA fabric and some ARM AARCH64 cores. Its expensive and small quantity but I think it will turn out to be important in the long run as the vanguard of chips that are not 'type specific' at the time of manufacture.
Imagine a company that has the same basic FPGA architecture, packaged in a variety of automotive spec packages, with one-time programmability. That would reduce the number of SKUs considerably (basically by package type).
This feels like lack of planning. The semiconductor industry was a known entity for many years before this level of integration came to vehicles. They should have known this would happen.
Ultimately the industry will have to maintain their own production for consistent silicon.
Ultimately the industry will have to maintain their own production for consistent silicon.
Why can't they just make pin-compatible versions of the chips using new processes? That would take time, but surely easier than investing in new fabs for old processes?
At the root is the question: "Do we make it more complex" or "Do we do try it 3 times".
The carmakers have been making the wrong call for the past decades.
One of their suppliers has eclipsed their impact on society, marking one of the first times in something like ~70 years that things didn't bend over to meet their needs.
----
And before you ask. Yes i would rather have a pace-maker with triple redundancy and thrice redesigned.
The carmakers have been making the wrong call for the past decades.
One of their suppliers has eclipsed their impact on society, marking one of the first times in something like ~70 years that things didn't bend over to meet their needs.
----
And before you ask. Yes i would rather have a pace-maker with triple redundancy and thrice redesigned.
But for the cost, couldn't (for instance) a 16nm fab produce 90nm chips?
If not, how complex would it be to "port" an existing 90nm chip, to produce a 16nm revision? Impossible, or quite easy but not cost effective? From the POV of the car companies, could such revisions be used directly, or would they need to be qualified in the same fashion as new parts?
If not, how complex would it be to "port" an existing 90nm chip, to produce a 16nm revision? Impossible, or quite easy but not cost effective? From the POV of the car companies, could such revisions be used directly, or would they need to be qualified in the same fashion as new parts?
Wouldn't the solution be planning better?
Sure, I understand that making a new car in combination with making a new ECU or brake controller on a new process is going to be stupidly dangerous and troublesome. So just don't.
Rather than doing it as a part of a single project, there should be a department or separate company making the ECU/controller. This way when the semiconductor company moves forwards, ECU Group starts a project targeting the new process, and releases the results when it's ready. And meanwhile what goes into the cars is the ECU built on the previous, well tested process.
Also, are modern cars really in much need of custom silicon? Micro-controllers are stupidly powerful now. There has to be off the shelf hardware capable of handling what a car needs. There are long standing architectures like ARM that don't require starting from scratch every time somebody makes a better chip.
Sure, I understand that making a new car in combination with making a new ECU or brake controller on a new process is going to be stupidly dangerous and troublesome. So just don't.
Rather than doing it as a part of a single project, there should be a department or separate company making the ECU/controller. This way when the semiconductor company moves forwards, ECU Group starts a project targeting the new process, and releases the results when it's ready. And meanwhile what goes into the cars is the ECU built on the previous, well tested process.
Also, are modern cars really in much need of custom silicon? Micro-controllers are stupidly powerful now. There has to be off the shelf hardware capable of handling what a car needs. There are long standing architectures like ARM that don't require starting from scratch every time somebody makes a better chip.
Really well written investigative piece. Its so often forgotten that safety critical hardware like a car ecu dances to a different rhythm than consumer electronics that just has to not explode in your pocket..
Is a part of the problem that the chips they are depending on are so trivial and small that you would end up losing more when cutting it up on a smaller process? I guess there is a limit to how many chips you could put on a single wafer even if the chip was only a single gate?
Maybe the old nodes are efficient enough for the chips they need. Then again it would make sense to update designs frequently enough to stay at nodes that have plenty of capacity, perhaps every 5 or 10 years or so?
Maybe the old nodes are efficient enough for the chips they need. Then again it would make sense to update designs frequently enough to stay at nodes that have plenty of capacity, perhaps every 5 or 10 years or so?
Can't is a really strong word. One potential solution that might open some paths to innovation would be to split up the product a bit more. Ship vehicles with extremely basic features only and some of those features intentionally designed to be swapped out in the short term. Use third party suppliers to swap in instrument clusters, engine controllers, and such and allow those to use all the latest chips.
Why can't they transition to fewer chips instead?
My old car runs fine controlled with a single atmega128 for the injectors, and that is the only chip in there (outside of the radio that I never use (I already carry a phone for all sort of entertainment))
My old car runs fine controlled with a single atmega128 for the injectors, and that is the only chip in there (outside of the radio that I never use (I already carry a phone for all sort of entertainment))
Can we manufacture designs for larger/older process nodes on newer ones? I understand it's a little bit like the DPI on a printer. I imagine there are many other changes, and expense, but would it work?
I'd love to see Ken Sherrif do a teardown of some vintage ECM.
https://www.autotrader.com/car-news/when-did-cars-get-comput...
(tldr)
> So instead of one clear answer, we have three possibilities: 1957 Rambler, 1958 Chrysler or 1968 Volkswagen.
https://www.autotrader.com/car-news/when-did-cars-get-comput...
(tldr)
> So instead of one clear answer, we have three possibilities: 1957 Rambler, 1958 Chrysler or 1968 Volkswagen.
The fine article quotes the Intel CEO:
> "I’ll make them as many 16 nanometer chips as they want"
> Carmakers have bombarded him with requests to invest in brand-new production capacity for semiconductors
> featuring designs that, at best, were state of the art when the first Apple iPhone launched.
> "It just makes no economic or strategic sense"
That website is unreadable on mobile due to ads.
But cars don't even need chips. Cars worked just fine for 70 years without chips. I don't get it.
Sounds like the auto-makers play down on the talent they have and prefer to push the work outside, probably fearing they are not able to execute.
This reminds me of being a 3rd party selling service to some company and they constantly ask for us to implement features that they could easily add - but can't. They don't say that flat out but we all understand what's happening..
This reminds me of being a 3rd party selling service to some company and they constantly ask for us to implement features that they could easily add - but can't. They don't say that flat out but we all understand what's happening..
Give me a microcontroller with between 32K and 2MB of onboard flash, some eeprom, multiple CAN peripherals, ADC, SPI, good timers, and some specialized peripherals - synched PWMs with dead time for driving power electronics, and some other special things. Dual-core lockstep for safety critical things (like the fronk latch). Some combination of all that at a price starting under a dollar.
Oh, and a temperature range of -40 to +125C ambient.
Can Intel do that on their obsolete 16nm multipatterning process for anywhere near the price target? Didnt think so.
Oh, and a temperature range of -40 to +125C ambient.
Can Intel do that on their obsolete 16nm multipatterning process for anywhere near the price target? Didnt think so.
https://news.ycombinator.com/item?id=28709481
> The aspect of the problem not talked about is that a lot of automotive chips are really, really old.
> And they are old, because of many certification requirements set partially by the industry, and a few odd governments.
> In reality, most of them are both less reliable, and harder to work with in comparison to open market parts.
> The only two things I seen chips with micron scale nodes used in my life were: air conditioner boards, and car parts.
For example, who in the world today makes PMOS TTL chips? I bet the few foundries who can still do that will bill car parts makers an arm, and a leg for keeping making something this old
Want to migrate? Find somebody who can can translate hand-drawn TTL chip to moderately modern CMOS under 60 years old.
At times, the cases about I heard of car makers were having troubles with were also nothing less of direct consequence of conscious overengineering: some BWMs have one STM32 per window switch just to blink the LED, and do ADC despite the switch having just 5 positions. And now they can't ship because of this single LED blinker.
> The aspect of the problem not talked about is that a lot of automotive chips are really, really old.
> And they are old, because of many certification requirements set partially by the industry, and a few odd governments.
> In reality, most of them are both less reliable, and harder to work with in comparison to open market parts.
> The only two things I seen chips with micron scale nodes used in my life were: air conditioner boards, and car parts.
For example, who in the world today makes PMOS TTL chips? I bet the few foundries who can still do that will bill car parts makers an arm, and a leg for keeping making something this old
Want to migrate? Find somebody who can can translate hand-drawn TTL chip to moderately modern CMOS under 60 years old.
At times, the cases about I heard of car makers were having troubles with were also nothing less of direct consequence of conscious overengineering: some BWMs have one STM32 per window switch just to blink the LED, and do ADC despite the switch having just 5 positions. And now they can't ship because of this single LED blinker.
Some things to emphasize:
OEMs (GM, Ford, Toyota, VW, etc) do not design components, and they do not want to. They design specifications for components, and then get suppliers to bid. This is great for efficiency in established ecosystems, not great for agility.
To my knowledge, GM did not cancel any chip orders, because GM itself had no chip orders (this is oversimplified). The suppliers cancelled chip orders.
For a 1st tier supplier to move to a different process/chip would also be difficult, because they do the same thing the OEM does - supply a specification, and get 2nd/3rd tier suppliers to bid on it.
A wholesale migration to a modern architecture is risky and costly.
Smaller process/feature size on a wafer is believed to be less resilient, for example to heat and vibration.
The risk to large automakers is that something goes wrong and they have to do a recall. The risk to up and comers (Tesla) is failure to grow. Also, Tesla has a small product line and absolute loads of cash.
If you were going to design a new car electrical architecture from scratch today, you would have something like a 40-60V system with a centralized controller (or pair of controllers in a safety redundant configuration).
Even with a largely cleansheet design, Tesla uses 12V because of the sheer ubiquity of 12V components.