Anyone have any sense on what code it would be? Might be time to dust off the old textbooks.
Flashlight app here I come.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable
- Thread starter AudubonB
- Start date
Flashlight app here I come.
OT weekend thinking about chips (the kind that go tick/tock) and maybe this should go into the Engineering thread (tag me to discuss in that thread how abouts...)
TL;DR - For Tesla, as a business, I think the need for having their own fab(s) is at or near the tipping point, as lack of scale is the show stopper and that issue is about to be eclipsed. Fabs are a huge capital investment and the timeline is 2+ years (even for Elon) once you have your chip roadmap and architectures nailed to build the machines that make the chips. Dojo would be a good Trojan Horse or a way to 'cut your teeth' on making near infinitely valuable compute that would go into the droid. If Tesla's ideal chips could come in a smallish set of types/classifications then it is a forgone conclusion they should pursue fab(s).
TL;DR - For Tesla, as a business, I think the need for having their own fab(s) is at or near the tipping point, as lack of scale is the show stopper and that issue is about to be eclipsed. Fabs are a huge capital investment and the timeline is 2+ years (even for Elon) once you have your chip roadmap and architectures nailed to build the machines that make the chips. Dojo would be a good Trojan Horse or a way to 'cut your teeth' on making near infinitely valuable compute that would go into the droid. If Tesla's ideal chips could come in a smallish set of types/classifications then it is a forgone conclusion they should pursue fab(s).
(Some high level tech PM weekend hand waving below...)
Today, Tesla can flash any chip in the vehicle. If they can do that, why not just send and receive constantly from a set of centralized compute cores to those vehicle systems and sub-systems to remove the localized compute? For one, they need to get off of the CANbus which is licensed and static. Move to their own deterministic communication method. A crazy amount of work, but not intractable.
I think an ideal draft design would be to have a set of two massive compute chips in lock-step, each with fully redundant power, communication and physically separated which contain all necessary cores for independent compute and networking layers, memory separation as well. This might not be the next step for Tesla depending on what the current designs include. Based on that, I'd assume you'd build a roadmap to get to the final design as jumping from current design to the most efficient design, would most likely cause build bottlenecks. I could see them developing more like 5 different chips based on classifications in the shorter term, as a stepping stone, and eventually moving to just one.
As some of us know, TSMC is one if not the most profitable company on the planet at making chips. They do this at scale and are pushing the envelope constantly so they can charge a pretty penny for their products over that products lifetime. They can also move customers to new products sooner than they'd want and charge another premium to help build the runtime stack, service stack, app platform, support stack, dev tools...etc. It is immensely profitable and will only become more and more so overtime.
It isn't, however, lost on me how powerful the current AMD Ryzen chip is though and that could be a good central compute for the time being. However, it might not be great at running Tesla's Linux as efficiently as a home grown architecture. And also, might not be robust enough to handle the compute needs for a fraction of the other ECUs compute needs.
Realize that Dojo and the Inference V4 chip will share similar components to generic generalized CPU compute. All custom AI chips require some amount of generalized compute. The ability to actually manufacture these chips would possibly be the next step, but Tesla might be utilizing some 3rd party patents for some parts of the Dojo or inference chip. So the next step would be to design a patent-free chip or establish a perpetuity free use license for critical IP. After this is known good and in parallel, learn to manufacture while iterating on a patent free or license-able fab that can meet your scale and classification needs.
Tesla has the chip design folks, but they'd need to either, hire fab design folks, buy IP for fab design or just buy a fab and dig in. Potential cost savings are staggering **IF** you knew you were going to scale to needing at least 1 billion chips/yr total and less than 10 fabs/lines. A "one chip to rule them all" design might be best and then just de-pop or fuse what you don't need for different needs. The ability for a single chip design, single footprint, single OS and stack would be a huge benefit across the board **IF** you had a high enough volume to justify the capital investment. Just a single interface, layer, deterministic design for networking would be enough to justify IMO as the gains in efficiency would be staggering. To not have to deal with any legacy interfaces and to know that delivery of messages throughout the car were going to arrive in a very tight tolerance makes the overall communication infrastructure quite simple relatively speaking.
Due to their specific, yet somewhat generic needs, the amount of chip designs they'd need for all their announced products is nearing 100 million/year (my estimate as each car, powerwall, charger, supercharger, eventually the droid..etc uses several chips per product)
This starts, for me, with the canonical question, how many of these chips are you going to need?
And ends with, let's say you have all the ideal chips you need, what then?
There's so much there to unpack...but hey, it's the weekend right?
Q: How many compute nodes are in a Tesla today? What does the BOM say?
A: Let's call this variable X
Q: Do all these nodes perform the same level of compute?
A: For simplicity sake, I might break this into classifications. Basic Linux OS runtime, rudimentary computations, specialized critical multi-layer, AI NN training, AI NN inference.
Q: Are there other needs outside of the car that the car classification of chips isn't covering?
A: I think they are covered
Q: Does every chip classification need a networking layer?
A: Totally, this is how we get to ideal communication in the ecosystem
Q: So, how many chips are we talking about?
A: Let's say there's a need for 100 million in 2025 to 1 Billion by 2030
Q: What is stopping Tesla from pursuing this?
A: Engineering first principles logic is NOT as simplistic as I've outlined so the answer is nuanced.
Today, Tesla can flash any chip in the vehicle. If they can do that, why not just send and receive constantly from a set of centralized compute cores to those vehicle systems and sub-systems to remove the localized compute? For one, they need to get off of the CANbus which is licensed and static. Move to their own deterministic communication method. A crazy amount of work, but not intractable.
I think an ideal draft design would be to have a set of two massive compute chips in lock-step, each with fully redundant power, communication and physically separated which contain all necessary cores for independent compute and networking layers, memory separation as well. This might not be the next step for Tesla depending on what the current designs include. Based on that, I'd assume you'd build a roadmap to get to the final design as jumping from current design to the most efficient design, would most likely cause build bottlenecks. I could see them developing more like 5 different chips based on classifications in the shorter term, as a stepping stone, and eventually moving to just one.
As some of us know, TSMC is one if not the most profitable company on the planet at making chips. They do this at scale and are pushing the envelope constantly so they can charge a pretty penny for their products over that products lifetime. They can also move customers to new products sooner than they'd want and charge another premium to help build the runtime stack, service stack, app platform, support stack, dev tools...etc. It is immensely profitable and will only become more and more so overtime.
It isn't, however, lost on me how powerful the current AMD Ryzen chip is though and that could be a good central compute for the time being. However, it might not be great at running Tesla's Linux as efficiently as a home grown architecture. And also, might not be robust enough to handle the compute needs for a fraction of the other ECUs compute needs.
Realize that Dojo and the Inference V4 chip will share similar components to generic generalized CPU compute. All custom AI chips require some amount of generalized compute. The ability to actually manufacture these chips would possibly be the next step, but Tesla might be utilizing some 3rd party patents for some parts of the Dojo or inference chip. So the next step would be to design a patent-free chip or establish a perpetuity free use license for critical IP. After this is known good and in parallel, learn to manufacture while iterating on a patent free or license-able fab that can meet your scale and classification needs.
Tesla has the chip design folks, but they'd need to either, hire fab design folks, buy IP for fab design or just buy a fab and dig in. Potential cost savings are staggering **IF** you knew you were going to scale to needing at least 1 billion chips/yr total and less than 10 fabs/lines. A "one chip to rule them all" design might be best and then just de-pop or fuse what you don't need for different needs. The ability for a single chip design, single footprint, single OS and stack would be a huge benefit across the board **IF** you had a high enough volume to justify the capital investment. Just a single interface, layer, deterministic design for networking would be enough to justify IMO as the gains in efficiency would be staggering. To not have to deal with any legacy interfaces and to know that delivery of messages throughout the car were going to arrive in a very tight tolerance makes the overall communication infrastructure quite simple relatively speaking.
Due to their specific, yet somewhat generic needs, the amount of chip designs they'd need for all their announced products is nearing 100 million/year (my estimate as each car, powerwall, charger, supercharger, eventually the droid..etc uses several chips per product)
This starts, for me, with the canonical question, how many of these chips are you going to need?
And ends with, let's say you have all the ideal chips you need, what then?
There's so much there to unpack...but hey, it's the weekend right?
Q: How many compute nodes are in a Tesla today? What does the BOM say?
A: Let's call this variable X
Q: Do all these nodes perform the same level of compute?
A: For simplicity sake, I might break this into classifications. Basic Linux OS runtime, rudimentary computations, specialized critical multi-layer, AI NN training, AI NN inference.
Q: Are there other needs outside of the car that the car classification of chips isn't covering?
A: I think they are covered
Q: Does every chip classification need a networking layer?
A: Totally, this is how we get to ideal communication in the ecosystem
Q: So, how many chips are we talking about?
A: Let's say there's a need for 100 million in 2025 to 1 Billion by 2030
Q: What is stopping Tesla from pursuing this?
A: Engineering first principles logic is NOT as simplistic as I've outlined so the answer is nuanced.
Fun fact, I know a few of the guys over on that team at Tesla. They're former colleagues of mine. Former Microsoft and Apple guys. They're backgrounds are in apps/game development. No I don't ask them what they're up to over there since they know I'm an active investor in TSLA.
Having said that, it's been fairly common knowledge in the tech community that this is coming and has mostly been on the backshelf due to Tesla wanting to get more a "standard" hardware config across all of their vehicles that also gives them flexibility going forward.
The crazy thing is that when Apple unveiled the App Store, it was relatively easy to replicate the hardware/software needed to run it. In smart phones, it's relatively easy to have a min spec that's easy for every phone maker to stick to in order to run say Android.
Now think about legacy auto makers trying to do the same. It's going to be a mess.
Fire at Rivian plant. Hopefully the damage was minimal and doesn't impact production going forward.
driveteslacanada.ca
There isn't a lot of detail in the article, lets hope that the damage was limited to the one vehicle and not assembly or production equipment.
Rivian assembly line catches fire, plant evacuated, no injuries
Firefighters from Normal and Bloomington responded to a fire at the Rivian assembly plant in Normal, Illinois. The alarm went off around 9:20 pm local time. According to Normal Fire Chief Mick Humer, the assembly […]
driveteslacanada.ca
There isn't a lot of detail in the article, lets hope that the damage was limited to the one vehicle and not assembly or production equipment.
Do you happen to know how many people are on the FSD team? Given the rapid rate of progress it would be interesting to know how many are working on it.
Nope, the people I know are on the Apps/Entertainment software team. I don't know anyone on the FSD team
Nice article today in the WSJ describing how some used Teslas are selling for above the price of a new one and more so than other cars (front page, below the fold, Sat/Sun Feb 5/6).
The high prices and the disparity indicate that demand has been accelerating (and are not simply due to scarcity because supply growth has been great despite shortages).
This data is evidence bolstering both my theory that fashions change with crises and my assertion that Tesla will benefit from the fashion changes underway due to the pandemic crisis.
"Used Teslas Defy Logic of Car Buying" is the printed newspaper title. Here is a link to it online.
www.wsj.com
The high prices and the disparity indicate that demand has been accelerating (and are not simply due to scarcity because supply growth has been great despite shortages).
This data is evidence bolstering both my theory that fashions change with crises and my assertion that Tesla will benefit from the fashion changes underway due to the pandemic crisis.
"Used Teslas Defy Logic of Car Buying" is the printed newspaper title. Here is a link to it online.
How to Boost the Price of Your Tesla? Drive It Off the Lot
Used models can sell for more than new ones. But parting with a used Tesla comes with side-effects: marital tension, a sense of loss.
www.wsj.com
We know there are 2000 people doing labeling. No one has said how many others are working on it.
Only if there is a dire need. They already have too much on their plate, such that sometimes, for instance, solar gets insufficient management attention. So if it's between HVAC and chips, I would prefer they work on HVAC.
Tesla has need of different types of chips, not all of which offer as juicy of margins. Neither Infineon nor STMicroelectronics are quite as lucrative as TSMC, as two examples. (That said, both SpaceX and Tesla will have a voracious appetite for STMicro's chips.)
Lastly, note that Tesla has avoided being on the absolute cutting edge with its designs, instead keeping to a node behind.
Last edited:
adiggs
Well-Known Member
Yeah, I used to think (after some here made some good arguments) that Tesla is in no need to have their own FAB.
Not so sure anymore. Especially after the last Earnings Call.
It seems they might NEED one. Sweet!
Edit: Elon liked this earlier:
A few observations about Tesla having a NEED for owning their own Fab. Apple (as one visible example) doesn't own their own Fab, nor do they need to. Then again I'm pretty sure that phones don't have as many different chips in them as a car does. From what I've just been reading, phones DO have a wider variety of chips in them from a manufacturing perspective. The chips that listen to audio have different properties than high end processors that makes for a different manufacturing process.
But we'll ignore that and pretend that a chip is a chip. That's some poor reasoning - Apple doesn't need it and they're bigger than Tesla. Heck Ford and GM have much higher unit volumes (cars) and they don't need their own fab. But that is also some pretty poor reasoning - X doesn't need it, therefore Y doesn't need it either.
We need to separate the two really big buckets in which chips fall into. Simplistically there are the high end processors that we all know and love. The Intel or AMD pc processors, the Apple M chips, the nVidia / AMD graphics processors, these sorts of chips. The shared characteristic of all of these chips is that to make them functional (energy and heat consumption, as well as size) is that they need to be manufactured on a cutting edge technology node. This isn't a nice to have - some of the advanced designs are so big that you can't fit them onto an older node wafer any longer. And even if you could the heat generated can't be removed from the chip fast enough. And even if you could remove the heat without burning the house down, the power supply to power the thing isn't available in residential or most commercial sites.
There are 3 or 4 manufacturers of these chips in the world - TSMC, Samsung, Intel, and (I think) Global Foundries. The latter I'm not sure of - I just haven't followed them as closely. What I DO know is that manufacturers of these chips have been disappearing from the world for more than 2 decades as manufacturers figure out that they don't have captive volume to keep the fabs busy, they don't have the money to fund the ongoing R&D for the -next- cutting edge node, etc.. The economics of being a manufacturer of these sorts of parts demands huge volume of these very expensive parts in order to pay for that cutting edge R&D, and the cutting edge tools and factory (and labor, and ...), in which to manufacture these. And because those economics are so dependent on stupendous volume, the bottom end volume producers are driven out of business. AMD selling off their foundries to Global Foundries being the most recent one of these events that I'm familiar with.
Intel gets the volume through their own products - the only remaining captive or integrated manufacturer.
TSMC gets the volume exclusively through external semiconductor design houses (like Apple, Qualcomm, Tesla; in all 3 cases I don't know which specific manufacturer the design houses work with, only that they don't manufacture these themselves).
Samsung gets the volume through both - internal products as well as foundry for others.
You can get a reasonably quick and easy financial statement level view into the capital and R&D for this stuff using TSMC or Intel financials. Intel's are 'polluted' by also having product R&D mixed in. Maybe 50/50 product and manufacturing split there? I really am guessing on that point.
Here are a couple of sources with more info on the economics for both sides of the relationship (designers and manufacturers).
Cost per wafer from TSMC, different nodes.
Die per wafer
Yield - the number of sellable chips after die are cut from the wafer, packaged, and tested. Not all die on a wafer are error free, and not all error free die get through packaging successfully. My understanding is that design houses buy wafers, not successful die. So the cost of a wafer with no good die, and a wafer with 100% good die is the same.
I can personally attest that wafers with 0 defects are sufficiently rare that the plant manager gets a page when it happens, a company email follows shortly (celebrate!), and in at least one circumstance the golden bb didn't get cut up and packaged - it got framed and went up on the wall
.
The other bucket are the itsy bitsy things that don't do much. An anti-lock brake controller for instance. These live under the same economics with an important and huge difference. They don't get made on the latest and greatest technology node. Heck I think there are circumstances where you actually want a larger feature size. So the factories tend to have been bought and paid for using previous products, the processes are ridiculously mature (really high die yield), and the products being made are really, really small (way more die per wafer).
To make the economics work here you need ridiculously high volume, though the cost to own and operate the manufacturing is much lower than that cutting edge stuff using EUV lithography (those people are doing magic - sufficiently advanced technology looks like magic - it does to me anyway).
With that background the question becomes what products/chips/parts does Tesla need to acquire for which owning and operating their own fab makes sense? Really the question is which bucket do these parts need to come from? For the FSD parts, maybe the screen graphics / processor, those come from bucket 1. With 1M units last year and say 5 of these per car - heck lets make it 10 to be generous (2 on the FSD board, processor for the screen, a graphics co processor for the screen, say 2 flash parts like those in a 2TB thumb drive, and 4 more for I have no idea) - that is annual unit volume of 10M parts in 3 or 4 designs.
Even if Tesla were doing 10M cars per year some day then they will need 100M parts per year. Either way this is small to medium volume for these very high end parts. Nobody working at these volumes can economically operate a fab, nor do they. They buy wafers with their designs from TSMC or Samsung (or Global Foundries).
The bucket I don't know as well is the low end bucket. Might it make sense for Tesla to buy a low end parts semiconductor manufacturer - probably one that is already in the business of making automotive parts - in order to secure the supply chain? I'd guess not, but the need for new process R&D won't exist. And as a bonus Tesla could design their own controllers rather than relying on whatever Bosch or other supplier has designed / is using.
My guess here is mostly based on the observation that none of the car makers in the world own their own fab (or collection of fabs) but rather buy these parts indirectly through the automotive supply chain. So maybe it would make sense for a Bosch or some such to own their own fab, but again, I don't believe anybody does. Between the variety of chips and the unit volumes available from the car industry as a whole, the car industry as a whole gets their controllers via a larger network of chip manufactures that work for more than just the car industry.
Therefore - Tesla with <1% of the worldwide car market probably doesn't (yet?) have enough volume to make even one of these low end fabs for building simple controllers work economically.
What will work, pretty easily, is offer to pay 2x or 5x what anybody else is willing or at least accustomed to paying for these parts, and let their supplier collect the incremental revenue and handle the R&D / capital / etc.. of semiconductor manufacturing. Even if that means $2k per car instead of $1k per car for these cheap controllers that will be a lot more manageable than trying to own and operate their own low end fab for $1B in order to save $1B. I made up the numbers and I suspect the low end controllers are worth more like <$100 per car.
There are 3 or 4 manufacturers of these chips in the world - TSMC, Samsung, Intel, and (I think) Global Foundries. The latter I'm not sure of - I just haven't followed them as closely. What I DO know is that manufacturers of these chips have been disappearing from the world for more than 2 decades as manufacturers figure out that they don't have captive volume to keep the fabs busy, they don't have the money to fund the ongoing R&D for the -next- cutting edge node, etc.. The economics of being a manufacturer of these sorts of parts demands huge volume of these very expensive parts in order to pay for that cutting edge R&D, and the cutting edge tools and factory (and labor, and ...), in which to manufacture these. And because those economics are so dependent on stupendous volume, the bottom end volume producers are driven out of business. AMD selling off their foundries to Global Foundries being the most recent one of these events that I'm familiar with.
Intel gets the volume through their own products - the only remaining captive or integrated manufacturer.
TSMC gets the volume exclusively through external semiconductor design houses (like Apple, Qualcomm, Tesla; in all 3 cases I don't know which specific manufacturer the design houses work with, only that they don't manufacture these themselves).
Samsung gets the volume through both - internal products as well as foundry for others.
You can get a reasonably quick and easy financial statement level view into the capital and R&D for this stuff using TSMC or Intel financials. Intel's are 'polluted' by also having product R&D mixed in. Maybe 50/50 product and manufacturing split there? I really am guessing on that point.
Here are a couple of sources with more info on the economics for both sides of the relationship (designers and manufacturers).
Cost per wafer from TSMC, different nodes.
Die per wafer
Yield - the number of sellable chips after die are cut from the wafer, packaged, and tested. Not all die on a wafer are error free, and not all error free die get through packaging successfully. My understanding is that design houses buy wafers, not successful die. So the cost of a wafer with no good die, and a wafer with 100% good die is the same.
I can personally attest that wafers with 0 defects are sufficiently rare that the plant manager gets a page when it happens, a company email follows shortly (celebrate!), and in at least one circumstance the golden bb didn't get cut up and packaged - it got framed and went up on the wall
.The other bucket are the itsy bitsy things that don't do much. An anti-lock brake controller for instance. These live under the same economics with an important and huge difference. They don't get made on the latest and greatest technology node. Heck I think there are circumstances where you actually want a larger feature size. So the factories tend to have been bought and paid for using previous products, the processes are ridiculously mature (really high die yield), and the products being made are really, really small (way more die per wafer).
To make the economics work here you need ridiculously high volume, though the cost to own and operate the manufacturing is much lower than that cutting edge stuff using EUV lithography (those people are doing magic - sufficiently advanced technology looks like magic - it does to me anyway).
With that background the question becomes what products/chips/parts does Tesla need to acquire for which owning and operating their own fab makes sense? Really the question is which bucket do these parts need to come from? For the FSD parts, maybe the screen graphics / processor, those come from bucket 1. With 1M units last year and say 5 of these per car - heck lets make it 10 to be generous (2 on the FSD board, processor for the screen, a graphics co processor for the screen, say 2 flash parts like those in a 2TB thumb drive, and 4 more for I have no idea) - that is annual unit volume of 10M parts in 3 or 4 designs.
Even if Tesla were doing 10M cars per year some day then they will need 100M parts per year. Either way this is small to medium volume for these very high end parts. Nobody working at these volumes can economically operate a fab, nor do they. They buy wafers with their designs from TSMC or Samsung (or Global Foundries).
The bucket I don't know as well is the low end bucket. Might it make sense for Tesla to buy a low end parts semiconductor manufacturer - probably one that is already in the business of making automotive parts - in order to secure the supply chain? I'd guess not, but the need for new process R&D won't exist. And as a bonus Tesla could design their own controllers rather than relying on whatever Bosch or other supplier has designed / is using.
My guess here is mostly based on the observation that none of the car makers in the world own their own fab (or collection of fabs) but rather buy these parts indirectly through the automotive supply chain. So maybe it would make sense for a Bosch or some such to own their own fab, but again, I don't believe anybody does. Between the variety of chips and the unit volumes available from the car industry as a whole, the car industry as a whole gets their controllers via a larger network of chip manufactures that work for more than just the car industry.
Therefore - Tesla with <1% of the worldwide car market probably doesn't (yet?) have enough volume to make even one of these low end fabs for building simple controllers work economically.
What will work, pretty easily, is offer to pay 2x or 5x what anybody else is willing or at least accustomed to paying for these parts, and let their supplier collect the incremental revenue and handle the R&D / capital / etc.. of semiconductor manufacturing. Even if that means $2k per car instead of $1k per car for these cheap controllers that will be a lot more manageable than trying to own and operate their own low end fab for $1B in order to save $1B. I made up the numbers and I suspect the low end controllers are worth more like <$100 per car.
Reasonable to know - I used to work at Intel but retired a year ago. I didn't work in or directly with the fab at any time (I was in IT). I just learned what little I know about semiconductor manufacturing by being an Intel investor and following the company closely. I did get to see a variety of presentations by the CFO over the years about the economics of semiconductor manufacturing. It's a business with gargantuan leverage.
The thing about leverage is that it works for and against. If you've got the volume then the leverage is for you and you're ridiculously profitable. If you don't have the volume then the leverage works against you and you're out of business in a snap of the fingers.
Everything here are my own ideas and observations of the world and are not Intel's.
EDIT to add: I also just saw @Discoducky ideas around this. I see merit in the one chip to rule them all idea at least as a first order approximation, and I see where he's going with that. I'm also pretty sure that he has a lot more relevant semiconductor industry experience than mine.
Last edited:
Point of clarification, if you please:
With respect to the ”Tesla’s App” conjecture - what kind of apps might it be appropriate for a Tesla, as opposed to a platform provider like Apple or Alphabet, to provide? Would it be likely that such apps would…could…should be available other than on vehicles’ screens? For a typical user, would it be easy to “parse” between one’s set of “phone apps” and “Tesla apps”?
A little bit separately, do any have well-founded opinions as to whether the revenues & margins Apple obtains through its universe of apps be appropriate to consider for Tesla?
With respect to the ”Tesla’s App” conjecture - what kind of apps might it be appropriate for a Tesla, as opposed to a platform provider like Apple or Alphabet, to provide? Would it be likely that such apps would…could…should be available other than on vehicles’ screens? For a typical user, would it be easy to “parse” between one’s set of “phone apps” and “Tesla apps”?
A little bit separately, do any have well-founded opinions as to whether the revenues & margins Apple obtains through its universe of apps be appropriate to consider for Tesla?
A few observations about Tesla having a NEED for owning their own Fab. Apple (as one visible example) doesn't own their own Fab, nor do they need to. Then again I'm pretty sure that phones don't have as many different chips in them as a car does. From what I've just been reading, phones DO have a wider variety of chips in them from a manufacturing perspective. The chips that listen to audio have different properties than high end processors that makes for a different manufacturing process.
But we'll ignore that and pretend that a chip is a chip. That's some poor reasoning - Apple doesn't need it and they're bigger than Tesla. Heck Ford and GM have much higher unit volumes (cars) and they don't need their own fab. But that is also some pretty poor reasoning - X doesn't need it, therefore Y doesn't need it either.
From all of this, my own interpretation is that Tesla is around 2 or 3 orders of magnitude too low volume to make owning their own fab to have a chance of being economically viable / good choice. If I'm off by 1 order of magnitude in Tesla's favor then we're still 1 or 2 orders of magnitude short. When we're a year or 2 away from 10M cars per year and if Tesla can make the business case, then an acquisition of an already operating manufacturer / fab is something I would only hate as an investor (as opposed to a "you've lost your mind and forgotten the mission" thing, probably causing me to divest) - at least that could be operated as its own already operating division rather than trying to build and hire all of this from scratch.We need to separate the two really big buckets in which chips fall into. Simplistically there are the high end processors that we all know and love. The Intel or AMD pc processors, the Apple M chips, the nVidia / AMD graphics processors, these sorts of chips. The shared characteristic of all of these chips is that to make them functional (energy and heat consumption, as well as size) is that they need to be manufactured on a cutting edge technology node. This isn't a nice to have - some of the advanced designs are so big that you can't fit them onto an older node wafer any longer. And even if you could the heat generated can't be removed from the chip fast enough. And even if you could remove the heat without burning the house down, the power supply to power the thing isn't available in residential or most commercial sites.
There are 3 or 4 manufacturers of these chips in the world - TSMC, Samsung, Intel, and (I think) Global Foundries. The latter I'm not sure of - I just haven't followed them as closely. What I DO know is that manufacturers of these chips have been disappearing from the world for more than 2 decades as manufacturers figure out that they don't have captive volume to keep the fabs busy, they don't have the money to fund the ongoing R&D for the -next- cutting edge node, etc.. The economics of being a manufacturer of these sorts of parts demands huge volume of these very expensive parts in order to pay for that cutting edge R&D, and the cutting edge tools and factory (and labor, and ...), in which to manufacture these. And because those economics are so dependent on stupendous volume, the bottom end volume producers are driven out of business. AMD selling off their foundries to Global Foundries being the most recent one of these events that I'm familiar with.
Intel gets the volume through their own products - the only remaining captive or integrated manufacturer.
TSMC gets the volume exclusively through external semiconductor design houses (like Apple, Qualcomm, Tesla; in all 3 cases I don't know which specific manufacturer the design houses work with, only that they don't manufacture these themselves).
Samsung gets the volume through both - internal products as well as foundry for others.
You can get a reasonably quick and easy financial statement level view into the capital and R&D for this stuff using TSMC or Intel financials. Intel's are 'polluted' by also having product R&D mixed in. Maybe 50/50 product and manufacturing split there? I really am guessing on that point.
Here are a couple of sources with more info on the economics for both sides of the relationship (designers and manufacturers).
Cost per wafer from TSMC, different nodes.
Die per wafer
Yield - the number of sellable chips after die are cut from the wafer, packaged, and tested. Not all die on a wafer are error free, and not all error free die get through packaging successfully. My understanding is that design houses buy wafers, not successful die. So the cost of a wafer with no good die, and a wafer with 100% good die is the same.
I can personally attest that wafers with 0 defects are sufficiently rare that the plant manager gets a page when it happens, a company email follows shortly (celebrate!), and in at least one circumstance the golden bb didn't get cut up and packaged - it got framed and went up on the wall
The other bucket are the itsy bitsy things that don't do much. An anti-lock brake controller for instance. These live under the same economics with an important and huge difference. They don't get made on the latest and greatest technology node. Heck I think there are circumstances where you actually want a larger feature size. So the factories tend to have been bought and paid for using previous products, the processes are ridiculously mature (really high die yield), and the products being made are really, really small (way more die per wafer).
To make the economics work here you need ridiculously high volume, though the cost to own and operate the manufacturing is much lower than that cutting edge stuff using EUV lithography (those people are doing magic - sufficiently advanced technology looks like magic - it does to me anyway).
With that background the question becomes what products/chips/parts does Tesla need to acquire for which owning and operating their own fab makes sense? Really the question is which bucket do these parts need to come from? For the FSD parts, maybe the screen graphics / processor, those come from bucket 1. With 1M units last year and say 5 of these per car - heck lets make it 10 to be generous (2 on the FSD board, processor for the screen, a graphics co processor for the screen, say 2 flash parts like those in a 2TB thumb drive, and 4 more for I have no idea) - that is annual unit volume of 10M parts in 3 or 4 designs.
Even if Tesla were doing 10M cars per year some day then they will need 100M parts per year. Either way this is small to medium volume for these very high end parts. Nobody working at these volumes can economically operate a fab, nor do they. They buy wafers with their designs from TSMC or Samsung (or Global Foundries).
The bucket I don't know as well is the low end bucket. Might it make sense for Tesla to buy a low end parts semiconductor manufacturer - probably one that is already in the business of making automotive parts - in order to secure the supply chain? I'd guess not, but the need for new process R&D won't exist. And as a bonus Tesla could design their own controllers rather than relying on whatever Bosch or other supplier has designed / is using.
My guess here is mostly based on the observation that none of the car makers in the world own their own fab (or collection of fabs) but rather buy these parts indirectly through the automotive supply chain. So maybe it would make sense for a Bosch or some such to own their own fab, but again, I don't believe anybody does. Between the variety of chips and the unit volumes available from the car industry as a whole, the car industry as a whole gets their controllers via a larger network of chip manufactures that work for more than just the car industry.
Therefore - Tesla with <1% of the worldwide car market probably doesn't (yet?) have enough volume to make even one of these low end fabs for building simple controllers work economically.
What will work, pretty easily, is offer to pay 2x or 5x what anybody else is willing or at least accustomed to paying for these parts, and let their supplier collect the incremental revenue and handle the R&D / capital / etc.. of semiconductor manufacturing. Even if that means $2k per car instead of $1k per car for these cheap controllers that will be a lot more manageable than trying to own and operate their own low end fab for $1B in order to save $1B. I made up the numbers and I suspect the low end controllers are worth more like <$100 per car.
Reasonable to know - I used to work at Intel but retired a year ago. I didn't work in or directly with the fab at any time (I was in IT). I just learned what little I know about semiconductor manufacturing by being an Intel investor and following the company closely. I did get to see a variety of presentations by the CFO over the years about the economics of semiconductor manufacturing. It's a business with gargantuan leverage.
The thing about leverage is that it works for and against. If you've got the volume then the leverage is for you and you're ridiculously profitable. If you don't have the volume then the leverage works against you and you're out of business in a snap of the fingers.
Everything here are my own ideas and observations of the world and are not Intel's.
EDIT to add: I also just saw @Discoducky ideas around this. I see merit in the one chip to rule them all idea at least as a first order approximation, and I see where he's going with that. I'm also pretty sure that he has a lot more relevant semiconductor industry experience than mine.
Thank you.
At what number would make sense for Tesla to have a FAB?
Is there a limit to the amount of stuff that would use a chip, any chip?
The need obviously needs to be there, if they decide to have a FAB.
adiggs
Well-Known Member
Just for fun, something I did later in my career was a bit of research into EUV for a project I was on. Again not working in or with the fab.A few observations about Tesla having a NEED for owning their own Fab. Apple (as one visible example) doesn't own their own Fab, nor do they need to. Then again I'm pretty sure that phones don't have as many different chips in them as a car does. From what I've just been reading, phones DO have a wider variety of chips in them from a manufacturing perspective. The chips that listen to audio have different properties than high end processors that makes for a different manufacturing process.
But we'll ignore that and pretend that a chip is a chip. That's some poor reasoning - Apple doesn't need it and they're bigger than Tesla. Heck Ford and GM have much higher unit volumes (cars) and they don't need their own fab. But that is also some pretty poor reasoning - X doesn't need it, therefore Y doesn't need it either.
From all of this, my own interpretation is that Tesla is around 2 or 3 orders of magnitude too low volume to make owning their own fab to have a chance of being economically viable / good choice. If I'm off by 1 order of magnitude in Tesla's favor then we're still 1 or 2 orders of magnitude short. When we're a year or 2 away from 10M cars per year and if Tesla can make the business case, then an acquisition of an already operating manufacturer / fab is something I would only hate as an investor (as opposed to a "you've lost your mind and forgotten the mission" thing, probably causing me to divest) - at least that could be operated as its own already operating division rather than trying to build and hire all of this from scratch.We need to separate the two really big buckets in which chips fall into. Simplistically there are the high end processors that we all know and love. The Intel or AMD pc processors, the Apple M chips, the nVidia / AMD graphics processors, these sorts of chips. The shared characteristic of all of these chips is that to make them functional (energy and heat consumption, as well as size) is that they need to be manufactured on a cutting edge technology node. This isn't a nice to have - some of the advanced designs are so big that you can't fit them onto an older node wafer any longer. And even if you could the heat generated can't be removed from the chip fast enough. And even if you could remove the heat without burning the house down, the power supply to power the thing isn't available in residential or most commercial sites.
There are 3 or 4 manufacturers of these chips in the world - TSMC, Samsung, Intel, and (I think) Global Foundries. The latter I'm not sure of - I just haven't followed them as closely. What I DO know is that manufacturers of these chips have been disappearing from the world for more than 2 decades as manufacturers figure out that they don't have captive volume to keep the fabs busy, they don't have the money to fund the ongoing R&D for the -next- cutting edge node, etc.. The economics of being a manufacturer of these sorts of parts demands huge volume of these very expensive parts in order to pay for that cutting edge R&D, and the cutting edge tools and factory (and labor, and ...), in which to manufacture these. And because those economics are so dependent on stupendous volume, the bottom end volume producers are driven out of business. AMD selling off their foundries to Global Foundries being the most recent one of these events that I'm familiar with.
Intel gets the volume through their own products - the only remaining captive or integrated manufacturer.
TSMC gets the volume exclusively through external semiconductor design houses (like Apple, Qualcomm, Tesla; in all 3 cases I don't know which specific manufacturer the design houses work with, only that they don't manufacture these themselves).
Samsung gets the volume through both - internal products as well as foundry for others.
You can get a reasonably quick and easy financial statement level view into the capital and R&D for this stuff using TSMC or Intel financials. Intel's are 'polluted' by also having product R&D mixed in. Maybe 50/50 product and manufacturing split there? I really am guessing on that point.
Here are a couple of sources with more info on the economics for both sides of the relationship (designers and manufacturers).
Cost per wafer from TSMC, different nodes.
Die per wafer
Yield - the number of sellable chips after die are cut from the wafer, packaged, and tested. Not all die on a wafer are error free, and not all error free die get through packaging successfully. My understanding is that design houses buy wafers, not successful die. So the cost of a wafer with no good die, and a wafer with 100% good die is the same.
I can personally attest that wafers with 0 defects are sufficiently rare that the plant manager gets a page when it happens, a company email follows shortly (celebrate!), and in at least one circumstance the golden bb didn't get cut up and packaged - it got framed and went up on the wall
The other bucket are the itsy bitsy things that don't do much. An anti-lock brake controller for instance. These live under the same economics with an important and huge difference. They don't get made on the latest and greatest technology node. Heck I think there are circumstances where you actually want a larger feature size. So the factories tend to have been bought and paid for using previous products, the processes are ridiculously mature (really high die yield), and the products being made are really, really small (way more die per wafer).
To make the economics work here you need ridiculously high volume, though the cost to own and operate the manufacturing is much lower than that cutting edge stuff using EUV lithography (those people are doing magic - sufficiently advanced technology looks like magic - it does to me anyway).
With that background the question becomes what products/chips/parts does Tesla need to acquire for which owning and operating their own fab makes sense? Really the question is which bucket do these parts need to come from? For the FSD parts, maybe the screen graphics / processor, those come from bucket 1. With 1M units last year and say 5 of these per car - heck lets make it 10 to be generous (2 on the FSD board, processor for the screen, a graphics co processor for the screen, say 2 flash parts like those in a 2TB thumb drive, and 4 more for I have no idea) - that is annual unit volume of 10M parts in 3 or 4 designs.
Even if Tesla were doing 10M cars per year some day then they will need 100M parts per year. Either way this is small to medium volume for these very high end parts. Nobody working at these volumes can economically operate a fab, nor do they. They buy wafers with their designs from TSMC or Samsung (or Global Foundries).
The bucket I don't know as well is the low end bucket. Might it make sense for Tesla to buy a low end parts semiconductor manufacturer - probably one that is already in the business of making automotive parts - in order to secure the supply chain? I'd guess not, but the need for new process R&D won't exist. And as a bonus Tesla could design their own controllers rather than relying on whatever Bosch or other supplier has designed / is using.
My guess here is mostly based on the observation that none of the car makers in the world own their own fab (or collection of fabs) but rather buy these parts indirectly through the automotive supply chain. So maybe it would make sense for a Bosch or some such to own their own fab, but again, I don't believe anybody does. Between the variety of chips and the unit volumes available from the car industry as a whole, the car industry as a whole gets their controllers via a larger network of chip manufactures that work for more than just the car industry.
Therefore - Tesla with <1% of the worldwide car market probably doesn't (yet?) have enough volume to make even one of these low end fabs for building simple controllers work economically.
What will work, pretty easily, is offer to pay 2x or 5x what anybody else is willing or at least accustomed to paying for these parts, and let their supplier collect the incremental revenue and handle the R&D / capital / etc.. of semiconductor manufacturing. Even if that means $2k per car instead of $1k per car for these cheap controllers that will be a lot more manageable than trying to own and operate their own low end fab for $1B in order to save $1B. I made up the numbers and I suspect the low end controllers are worth more like <$100 per car.
Reasonable to know - I used to work at Intel but retired a year ago. I didn't work in or directly with the fab at any time (I was in IT). I just learned what little I know about semiconductor manufacturing by being an Intel investor and following the company closely. I did get to see a variety of presentations by the CFO over the years about the economics of semiconductor manufacturing. It's a business with gargantuan leverage.
The thing about leverage is that it works for and against. If you've got the volume then the leverage is for you and you're ridiculously profitable. If you don't have the volume then the leverage works against you and you're out of business in a snap of the fingers.
Everything here are my own ideas and observations of the world and are not Intel's.
EDIT to add: I also just saw @Discoducky ideas around this. I see merit in the one chip to rule them all idea at least as a first order approximation, and I see where he's going with that. I'm also pretty sure that he has a lot more relevant semiconductor industry experience than mine.
The thing I remember most clearly was the litho machine - the one that shines the light through the mask to pattern the current layer on the wafer. That one machine is .. uhm ... huge. The article I found at the time, if I remember right (and am not finding that source again right now), said that the machine was 2 stories tall, the size of a reasonably typical suburban house, and needed its very own 1MW power supply.
But what's going on INSIDE of that machine is magic. Or well, sufficiently advanced technology anyway.
To do EUV you need an EUV flashlight (my term) of exactly the right wavelength of 13.5nm. Amazingly enough you can't grab an LED that'll do that at the corner hardware store.
Lots of technical details about how to do it at that link. My translation goes like this:
- squirt a really, really small bit of liquid tin into a chamber
- hit it with a laser, not to vaporize it, but to push that bit of tin into exactly the right shape
- for the second laser that does vaporize it
- that gets you lots of different stuff, some of which is that 13.5nm euv that you want. the rest is junk and needs to be separated out
- so grab the stuff you want, isolate the stuff you don't want (gotta clean the lint trap and the inside of that chamber periodically, and that means downtime)
- keep that running fast enough that you get a flashlight and not a flash
And now you've got a flashlight.
Except that you can't use mirrors, or at least regular mirrors, to get this light all lined up and going where you want it to. This wavelength goes through mirrors. So there are lots more problems downstream from here. But hey - at least now you've got a flashlight.
And after all of this pain and agony to get it working in a lab, the R&D boffins needed another decade+ to scale the power output up enough to make it somewhat economically usable in a production fab.
The really crazy part of this is that EUV lithography is just now starting to get into production fabs. Yay - its finally here. That means that the companies that are doing this have R&D types that are starting work today on stuff that'll go into production fabs in 10 years, and have had people working on this EUV stuff for 10 years (actually much longer - this particular technology has been any year now for about 10 years. I wonder where I've heard something like this before
).For cutting edge semiconductor chip manufacturing you need a large team of people that are working on the stuff that is 1-10 years out (and every year in between) in order to have the next technology node ready to go, after the current cutting edge technology node. It isn't just money (though there's a need for a lot of it) - its the team of magicians (scientists). This work is pretty much invisible to chip design houses, probably as it should be. The design skillset and the manufacturing R&D skillset in this world are pretty much orthogonal.
Home prices up 30% where I live, salary in IT almost doubled in last 5 years, restaurant visits twice as expensive since Covid.
Last edited:
I wonder if they will ever offer AMD Zen retrofits to owners of cars with Intel Atom. If not, I'll never be able to play The Witcher 3 on my car.
pitabun
Member
Is AMD Zen the GPU they advertised as PS5 equivalent when Plaid came out? I honestly don't know.
adiggs
Well-Known Member
That's part of the problem. We're talking about chips as if a chip is a chip. If that were the case there wouldn't be an Nvidia or AMD graphics processor business line - everybody would be using Intel processors, as many as needed, to also handle the graphics. Way, way back in the 486 days was when -the- processor didn't have the right stuff to do graphics as well as everything else. That's when the 486 graphics co-processor showed up as something you could add on to juice up your machine. And they did!
The details of chip design for a graphics processor are different from a general compute processor. Which are also different from the chips optimized for neural network training, etc.. etc..
Even in the simple controller bucket a chip isn't a chip. I don't know that category as well, and maybe there is a more expensive generic controller that could be designed that could replace any and all other controllers. One thing I do know is the economics are going to be ridiculously hard to fight - some of these cost $0.10 each in bulk and some are going to be more like $10. Even if that generic controller can be built for $10 who is going to put that into something when a $0.10 part will do the job? For 1M units - maybe. But for billions and billions? Its not economically feasible when you're only getting 0.10 worth of value from the part.
Clearly efforts to minimize the number of different chip designs will increase the volume of the stuff you do make and use, and that is good economically. There are good reasons we don't have a chip to rule them all today - some of them probably bad - but the economics also drives us in this direction.
That being said one technology vector that is being explored is a way to put many different slivers of silicon (die) together into a single package. If this just worked then you might imagine some general purpose compute in 1 sliver, some graphics accelerate in a second sliver, some USB controller logic (from a cheaper technology node) in a third sliver, some dynamic memory in a 4th sliver, some flash memory in a 5th sliver, etc.. And all of that in a package with a set of pins to plug into a single plug on a motherboard that looks like a single "chip" that does many things.
So you can get some of the economic benefits of one-chip-to-rule-them-all while also getting some of the economic benefits of manufacturing different controller / processor logic in different technology nodes.
I think this general idea is referred to as chiplets (vs SOC or system-on-chip). This link is to Cadence where they call this System-in-Package (vs system-on-chip).
Helpful, but you still end up needing many chip designs. Though in this case some chiplets can be designed in house and manufactured elsewhere, while some chiplets are designed and manufactured in house. Do you use an external packaging company in this case, or do you also do your own assembly / test? The problems they just keep on coming!
adiggs
Well-Known Member
The more direct answer - when Tesla has more of the very high end processor demand than Intel has demand for their high end parts. That'd be the entry level. I haven't seen that volume number published, but you might be able to back into it between revenue and average selling price info - that is stuff I've seen regularly published in the past (its reasonable to approximate Intel revenue as Intel processor product revenue for this scale of analysis). Revenue/asp will get you units.
Oh - and Apple across all of its products might be (probably is) at that volume, or higher already. They don't have, and pretty sure they want nothing to do, with owning and operating their own fab. It isn't because they can't afford it.
Makes sense.
But until now nobody was thinking about making an autonomous robot either.
I think we are talking about different time frame. I’m kinda thinking about 2030 to 2040, Moon, Mars.
Similar threads
- Locked
- Locked
- Locked
