Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable

[heltok]

Probably Toyota or VW.

Solid state batteries are like the holy grail of batteries: cheap, high range, high charge rate, and high energy output. However, they are a tough problem to also solve. Typically liquid electrolytes acts as a conductor for energy transfer and they are easier to mass produce, scale and cool in a traditional form factor. They can be easily scaled. Since this is solid state, it relies on an anodeless design. The material that acts as a dendrite barrier is also a mystery as it must also allow for energy transfer while preventing a short circuit. It is also drastically different than conventional batteries since each cell is like a playing card. So far stacking these haven't really been tested. I really hope they do indeed have the formula and this isn't another Theranos.

Time will tell.

9:30 and forward

Imo people need to stop hyping Solid State. It will not be the killer technology that saves Toyota or VW from Tesla. It will prevent the pain GM is going through that Tesla has solved, it might lower costs slightly by the end of the century and it might improve range a little. But 4680 accomplishes a lot of these goals also. And in the end it will be costs and scale that matter and there 4680 LFP will be king for the next few years. Maybe later solid state will improve things a little bit more, but by then Tesla might sell more cars than Toyota and have their own tech that is even better.

 

[Krugerrand]

Nasty. So prediction from last week is coming true. Macro in total correction territory. Spy has about another 6.5% downside. Tesla may fair better but let's just hope it can keep beating macro like today.

Gap close at 749, strong resistance at 755, need to break to get to 800. On the downside 714 is support and 692 is the support that will reverse tsla's bull trend to a bear trend. Expect more red in the macro environment ahead as vix closed above 23 two days a row while all indexes are closing below 50 Ema 3 days in a row. Just gonna be a sh%# show the next couple of weeks.

Didn’t you say, though, that there would be a rotation of funds into high tech???

 

[RobDickinson]

Actually, with respect to chips...
There's two divisible parts of building an IC (saying this as a person who had actually done this.. with a cast of many).

1. Designing the device. This is typically in Verilog or VHDL or some similar, more complex tool set, where one (actually, a bunch of people) are architecting, documenting, and laying out the device. It's a bit like writing software.
   • When I write software, being a EE, I tend to be very aware of the fact that each line of code, after compilation, becomes assembly code with JMP, LOADREG, indirect (machine) addressing, and all that jazz. Want a SWITCH statement? The compiler generates different code for one of them as compared to playing multiple IF/THEN/ELSE blocks.
   • When I'm writing VHDL - every line of VHDL translates into honest-to-golly NAND, NOR, AND, OR, D-Flip Flops, and so on. In addition, specialized VHDL/Verilog code can pull up different kinds of RAM blocks and all that; and, yes, one thinks hard about where all this hardware is going on the die that one is going to build. At the end of all this, the compilation process results in gates and such with signals, propagation delays, I/O buffers, and lots of floor planning.
2. The Foundry. The Foundry supplies libraries and tools to the people doing #1; what they get back from the designers is stuff that's extremely close to actual layout, but not quite. The Foundry people then take the design data and generate actual mask sets that are used for photolithography; and these, then, are actually made into actual silicon (or GaAs, etc.)

The serious, billions of dollars of construction, care and feeding, and making it all work: That's in #2. #1 is coding, tools sets, simulation (the actual device files bounce back and forth between the chip designers in #1 and the foundry types in #2, refining the design, doing simulations, Simulations, SIMULATIONS until everybody on both sides of the fence are ready to die. Mind you, a good deal of the expenses in #2 are, when going to a particular sub-nanometer technology, is building test chips with all sorts of gates, memories, I/O buffers, and what-all which then get characterized. The characterization gets stuffed into the libraries that get handed to #1, designed to work with the design tools (Synopsys, Mentor Graphics, tons of others), and then the #1 types do the design.

Tesla designed the neural network chips for the computer in the car; but they didn't run the foundry. The foundry, an outside company, built the actual devices, and both crowds validated working silicon.

Costs for the #1 crowd is expensive in people, software seats, and specialized simulation and design support hardware; but it's not billions of dollars.

And, if you're wondering, the fact that Intel still does both #'s 1 and 2 is a continuing source of amazement to the industry at large.

Thing is.. Once one has a working silicon design, the cost per chunk of silicon is very, very small. So doing #1, despite the expense, only makes sense if one is going to buy zillions of a device. If you're only going to do 1000 of a chip, you buy somebody else's FPGA with the right feature set and pay $50-$1000 per chip. If you're going to buy 10 million, then it's cheaper by far to pay the NRE (Non-Recoverable Expense) to the Foundry people for the first hundred devices and pay your design team to work their butts off for a year or more.

So, consider Bosch. They do engine controllers. And the devices they design (assuming that they don't just buy a general purpose controller from somebody else) they sell to everybody who'll buy the box with the chip in it. They can sell a million each to three or four different car companies and make money; if the car company tries to roll their own, they won't beat Bosch on price since they'll have the fixed costs to get the design done, but won't have the volume to make it better. Hence, while I could be wrong on this, I betcha none of the big car companies do any of their own silicon designs (#1), because they can get the same, changing slowly over time, designs built by a specialized design house (a la Bosch) who can sell to all the big design houses.

Which is fine, reduces the cost of manufacture - but, when something Really Different shows up, like the neural chips Tesla is building, this whole incestuous relationship between the car companies and their suppliers falls apart:

1. The car companies don't even have #1 design teams. They have system integrators that mess around with catalogs of what the likes of Bosch & its competitors are building, Even if they wanted to do the whole-car-make-it-go-on-a-test-track, they would have to work with a selected supplier to get the silicon built, knowing all the while that any technology so built would also be going straight to the car company's direct competitors. Ouch.
2. The companies like Bosch who actually have #1 design teams would have to spend serious money with no guarantee they're getting it back, and they don't have, really, the knowledge that the car companies do. What do they do - partner with one car company? They could get stuck into a contract where any new stuff is only shared with that car company, and no other. That wrecks their whole business model of selling to everybody.
3. Eventually, the two points immediately above will get sorted - but this is one place where Tesla's vertical integration must be driving their competitors absolutely nuts.
4. Let's get this straight: Nvidia, a graphics card company for crying out loud, is a direct competitor to Tesla in the neural, self-driving computer space. Why? Because they make really, really, fast CPUs and parallel processors for (wait for it) better game playing on PCs. This is not stuff optimized for driving a real car around: It's stuff that's just close enough so that people without a #1 design team and the design rules to back it up can get something, anything out the door so they don't end up in the trash heap of failed businesses. I got no doubt that an Nvidea-based driving computer can do some self-driving; but is it going to be as good, or as cheap, as purpose-built silicon designed from scratch for the purpose?

This is what it looks like when your competitor is literally a few years ahead of you and you don't have the human and technological resources to compete. Because, when you did have them, you looked around, figured that nobody else needed to spend any money on New Stuff anyway, and fired them all, instead of competing because ten years down the road the barriers to entry were going to be low enough that a real competitor would show up and wipe you away.
You might have a chance if, when that competitor got started, you got started, too - but, from all accounts, the Big Iron Auto Companies spent most of their time saying, "Tesla will never succeed!" and staring at their navels.

And that's just the hardware. Don't get me started on the software.

Another part of this, highlighted by Intel this week is that companies like Bosch are using old designs on old production nodes. They do this to save money because they dont want to redo their stuff for newer nodes/production lines. This would cost them a fair bit.

But only a fraction of what it would cost intel to add new capacity for their old designs. something Intel etc just wont do.

So you are in a situation where there is limited production capacity, increasing demand, and the companies that would benefit most from the shift dont own the technology or have the capability to do anything but wait and hope for parts.


Remember Ford is talking about a mach-e needing 10 times the number of chips than a fossil car (3000 vs 300), but they have no way of integrating functionality or breaking out of the supplier fed structure they have now.

 

[Singuy]

Didn’t you say, though, that there would be a rotation of funds into high tech???

I said high growth. We see many high growth stocks doing WAY better than expected but a correction is strong. Tsla is beating macro. AMD, ARK, and Fiverr are some of mine beating macro. Lets just see how well they all hold on as the market rolls over.

 

[lafrisbee]

Who is responsible for accidents and the insurance at the time is an interesting question for L4/L5 vehicles.

I have provided a quite likely scenario.
FSD will ONLY be available as a subscription in the future, and the monthly rate will be seen as paying an insurance premium and getting a driver for free.

 

[Paul_SF]

Actually, with respect to chips...
There's two divisible parts of building an IC (saying this as a person who had actually done this.. with a cast of many).

1. Designing the device. This is typically in Verilog or VHDL or some similar, more complex tool set, where one (actually, a bunch of people) are architecting, documenting, and laying out the device. It's a bit like writing software.
   • When I write software, being a EE, I tend to be very aware of the fact that each line of code, after compilation, becomes assembly code with JMP, LOADREG, indirect (machine) addressing, and all that jazz. Want a SWITCH statement? The compiler generates different code for one of them as compared to playing multiple IF/THEN/ELSE blocks.
   • When I'm writing VHDL - every line of VHDL translates into honest-to-golly NAND, NOR, AND, OR, D-Flip Flops, and so on. In addition, specialized VHDL/Verilog code can pull up different kinds of RAM blocks and all that; and, yes, one thinks hard about where all this hardware is going on the die that one is going to build. At the end of all this, the compilation process results in gates and such with signals, propagation delays, I/O buffers, and lots of floor planning.
2. The Foundry. The Foundry supplies libraries and tools to the people doing #1; what they get back from the designers is stuff that's extremely close to actual layout, but not quite. The Foundry people then take the design data and generate actual mask sets that are used for photolithography; and these, then, are actually made into actual silicon (or GaAs, etc.)

The serious, billions of dollars of construction, care and feeding, and making it all work: That's in #2. #1 is coding, tools sets, simulation (the actual device files bounce back and forth between the chip designers in #1 and the foundry types in #2, refining the design, doing simulations, Simulations, SIMULATIONS until everybody on both sides of the fence are ready to die. Mind you, a good deal of the expenses in #2 are, when going to a particular sub-nanometer technology, is building test chips with all sorts of gates, memories, I/O buffers, and what-all which then get characterized. The characterization gets stuffed into the libraries that get handed to #1, designed to work with the design tools (Synopsys, Mentor Graphics, tons of others), and then the #1 types do the design.

Tesla designed the neural network chips for the computer in the car; but they didn't run the foundry. The foundry, an outside company, built the actual devices, and both crowds validated working silicon.

Costs for the #1 crowd is expensive in people, software seats, and specialized simulation and design support hardware; but it's not billions of dollars.

And, if you're wondering, the fact that Intel still does both #'s 1 and 2 is a continuing source of amazement to the industry at large.

Thing is.. Once one has a working silicon design, the cost per chunk of silicon is very, very small. So doing #1, despite the expense, only makes sense if one is going to buy zillions of a device. If you're only going to do 1000 of a chip, you buy somebody else's FPGA with the right feature set and pay $50-$1000 per chip. If you're going to buy 10 million, then it's cheaper by far to pay the NRE (Non-Recoverable Expense) to the Foundry people for the first hundred devices and pay your design team to work their butts off for a year or more.

So, consider Bosch. They do engine controllers. And the devices they design (assuming that they don't just buy a general purpose controller from somebody else) they sell to everybody who'll buy the box with the chip in it. They can sell a million each to three or four different car companies and make money; if the car company tries to roll their own, they won't beat Bosch on price since they'll have the fixed costs to get the design done, but won't have the volume to make it better. Hence, while I could be wrong on this, I betcha none of the big car companies do any of their own silicon designs (#1), because they can get the same, changing slowly over time, designs built by a specialized design house (a la Bosch) who can sell to all the big design houses.

Which is fine, reduces the cost of manufacture - but, when something Really Different shows up, like the neural chips Tesla is building, this whole incestuous relationship between the car companies and their suppliers falls apart:

1. The car companies don't even have #1 design teams. They have system integrators that mess around with catalogs of what the likes of Bosch & its competitors are building, Even if they wanted to do the whole-car-make-it-go-on-a-test-track, they would have to work with a selected supplier to get the silicon built, knowing all the while that any technology so built would also be going straight to the car company's direct competitors. Ouch.
2. The companies like Bosch who actually have #1 design teams would have to spend serious money with no guarantee they're getting it back, and they don't have, really, the knowledge that the car companies do. What do they do - partner with one car company? They could get stuck into a contract where any new stuff is only shared with that car company, and no other. That wrecks their whole business model of selling to everybody.
3. Eventually, the two points immediately above will get sorted - but this is one place where Tesla's vertical integration must be driving their competitors absolutely nuts.
4. Let's get this straight: Nvidia, a graphics card company for crying out loud, is a direct competitor to Tesla in the neural, self-driving computer space. Why? Because they make really, really, fast CPUs and parallel processors for (wait for it) better game playing on PCs. This is not stuff optimized for driving a real car around: It's stuff that's just close enough so that people without a #1 design team and the design rules to back it up can get something, anything out the door so they don't end up in the trash heap of failed businesses. I got no doubt that an Nvidea-based driving computer can do some self-driving; but is it going to be as good, or as cheap, as purpose-built silicon designed from scratch for the purpose?

This is what it looks like when your competitor is literally a few years ahead of you and you don't have the human and technological resources to compete. Because, when you did have them, you looked around, figured that nobody else needed to spend any money on New Stuff anyway, and fired them all, instead of competing because ten years down the road the barriers to entry were going to be low enough that a real competitor would show up and wipe you away.
You might have a chance if, when that competitor got started, you got started, too - but, from all accounts, the Big Iron Auto Companies spent most of their time saying, "Tesla will never succeed!" and staring at their navels.

And that's just the hardware. Don't get me started on the software.

Thank you for this! And please, do start on the software!

 

[heltok]

Karpathy retweeted this one:

https://twitter.com/x/status/1439914539194142721

Seems like a breakthrough in training transformers, for language models and for image recognition(that Tesla are using), that in itself might be 4x faster to train to the same level of performance. There might be other work that brings part of this benefit that Tesla are already doing, but it seems likely that Tesla can use this to train their neural networks faster. Code is opensource and out on github. Here is the paper:

https://arxiv.org/pdf/2109.08668.pdf

 

[RobDickinson]

Here is the paper:

https://arxiv.org/pdf/2109.08668.pdf

|
Yeah I looked through that paper, it has many words and pictures, I knew some of the words too.

 

[Singuy]

Karpathy retweeted this one:

https://twitter.com/x/status/1439914539194142721

Seems like a breakthrough in training transformers, for language models and for image recognition(that Tesla are using), that in itself might be 4x faster to train to the same level of performance. There might be other work that brings part of this benefit that Tesla are already doing, but it seems likely that Tesla can use this to train their neural networks faster. Code is opensource and out on github. Here is the paper:

https://arxiv.org/pdf/2109.08668.pdf

My speculation is that Karpathy's team either knew about this prior or also contributed to the use of "Primer" from this paper. Karpathy and team talked about them working on optimizing/reduce resource usage with current hardware during AI day.

 

[heltok]

|
Yeah I looked through that paper, it has many words and pictures, I knew some of the words too.

Haha, true. I understood like 90% of it. Anyway I think it can be summarized in the graph to the right here. The old transformer for 72k hours got beaten by the new Primer transformer after ”only” 24k hours of training. And if you let the new one run for 72k hours performance improved.

There is a small cost to intialize the system but if you train the same system over and over you don’t need to redo this step. So for Tesla who are retraining the same network often this is less of a problem…

 

[Blue horseshoe]

Haha, true. I understood like 90% of it. Anyway I think it can be summarized in the graph to the right here. The old transformer for 72k hours got beaten by the new Primer transformer after ”only” 24k hours of training. And if you let the new one run for 72k hours performance improved.
View attachment 712613

There is a small cost to intialize the system but if you train the same system over and over you don’t need to redo this step. So for Tesla who are retraining the same network often this is less of a problem…

This must be what it's like for Ford execs trying to figure out how Tesla does so well

 

[NZ10K]

Haha, true. I understood like 90% of it. Anyway I think it can be summarized in the graph to the right here. The old transformer for 72k hours got beaten by the new Primer transformer after ”only” 24k hours of training. And if you let the new one run for 72k hours performance improved.
View attachment 712613

There is a small cost to intialize the system but if you train the same system over and over you don’t need to redo this step. So for Tesla who are retraining the same network often this is less of a problem…

Are you sure what you were the paper is talking about? 72k hours is close to 8 years.

 

[dhanson865]

View discretion advised. Tesla's have avoided many many accidents. We have driven 4 Teslas(own 3 now) and have not had an accident in them. In fact our Tesla's avoided several close calls such as a Semi moving into our lane or a a car slowing ahead which our responded to. We have EAP and FSD beta in all cars.
I don't feel safe in ANY car other than our Teslas. The regulators better check out the FACTS and forget politics. This is life or death for people in and out of cars.
I am amazed how stupid some individuals have become due to either organic brain issues or money.

Aaron Smith at 44:10 in that video watched a left turn signal on for 10+ seconds before entering the blind spot of an 18 wheeler. I'm not feeling much sympathy for the Tesla driver there. He could have slowed slightly and let the truck over without incident if he paid attention to that blinker.

 

[powertoold]

Are you sure what you were the paper is talking about? 72k hours is close to 8 years.

72k cpu hours. If you had 72k cpus, it'd take one hour. Do the math

 

[dhanson865]

72k cpu hours. If you had 72k cpus, it'd take one hour. Do the math

Pricing | Cloud TPU | Google Cloud looks like they stop at 2048 so it'd take ~36 hours with a rented setup that is maxed out. Of course Tesla has their own and would have a different rate and different number of CPUs to get the same result and wouldn't be limited by Googles rental cluster sizes.

 

[MC3OZ]

If you want a good laugh, I recommend this video:-

Whoever produced these estimates together makes Gordon look like a genius. Surely this is an elaborate joke?

 

[heltok]

Pricing | Cloud TPU | Google Cloud looks like they stop at 2048 so it'd take ~36 hours with a rented setup that is maxed out. Of course Tesla has their own and would have a different rate and different number of CPUs to get the same result and wouldn't be limited by Googles rental cluster sizes.

Probably more like 256 cores TPUv3 for 2weeks. And run in parallel the 4 different test setups. Training GPT-3 costs ~$1M before this, after this work anyone who wants can do it for only $250k ^^

Imo we should be scared, one paper getting us 4x closer to the singularity, then Dojo getting us another magnitude closer. Soon Codex might be able to code itself:

 

[Ohmster]

Full Speed Ahead……

Tesla Gigafactory Shanghai enters “Plaid Mode” in its end of Q3 2021 push — TESLARATI

As the third quarter comes to a close, Tesla China has practically entered “Plaid Mode” with its operations in Gigafactory Shanghai. Based on recently taken videos of the expansive electric vehicle production plant, it appears that strong efforts are now underway to ensure that Giga Shanghai...

apple.news

 

[Ohmster]

Pretty slick. Anyone else capable of this?

Tesla Model 3 and Model Y will now detect and slow down for emergency vehicles — TESLARATI

Tesla’s software update 2021.24.12 seems to include something noteworthy for the Model 3 and Model Y. While the software update’s Release Notes were identical to previous updates in the way that they highlight features like Disney+ streaming and Car Wash Mode, 2021.24.12 also included something...

apple.news

 

[Artful Dodger]

TSLA up ~2x vs. macros, after some early hi-jinx in the Pre-market:

Nasdaq 100 Dec 21 (NQ=F)​

CME Delayed Price. Currency in USD
15,079.75 +55.75 (+0.37%)

TSLA Pre-Market Quotes Live​

This page refreshes every 30 seconds.
Data last updated Sep 22, 2021 05:31 AM ET.

Consolidated Last Sale	$744.16 +4.78 (+0.65%)
Pre-Market Volume	24,482
Pre-Market High	$745.84 (04:48:05 AM)
Pre-Market Low	$738.95 (04:00:05 AM)