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

[Krugerrand]

Telsa has installed this automated/robotic metal brake to bend the stainless steel for Cybertruck: (Trumpf TruBend 5320)

Joe Tegtmeyer on Twitter: "Today at Giga Texas, a large delivery with a Trumpf Truebend 5320 machine was delivered. Here are a few images of the delivery and a link to the company website showing this exact machine in action! https://t.co/3Y9wSJkhkI https://t.co/dau5cUd2ZC" / Twitter | June 09, 2023​

Here is this machine in action: (Robots building Robots...)

TRUMPF Automation: BendMaster with TruBend 5320 - Demo

Cheers!

Yeah, look at it whiz through that part. 3 1/2 minutes with time lapse. Hope Tesla has more than one brake press.

 

[WarpedOne]

Is this recent H100 cluster a separate computer?

Yes.

Then what's up with Dojo

It is coming into life / production.

Dojo is optimized ground up chip, motherboard, cabinet and cluster. As such it presents a high risk project that is impossible to predict when it will be able to start working on real-world-problems and/or when it will be able to replace all other GPU clusters.

Until that happens Tesla utilizes old-school approach and takes delivery of orders that were placed years ago.

 

[Cosmacelf]

Sorry if this is a lame question... I thought Dojo used Tesla AI Chips with custom boards and amazing cooling designs.
Is this recent H100 cluster a separate computer? Then what's up with Dojo - or am I confusing one them?

Dojo uses custom Dojo specific custom AI chips (different from the custom AI inference chips used in all Tesla cars), with indeed amazing cooling designs. Did you know that some AI clusters are just now moving from air cooled to water cooled? When designing Dojo, Tesla correctly went water cooled from the get go.

Anyways, Teslas recent H100 cluster is indeed a separate computer and it does indeed do the same or similar function as Dojo. Tesla just needs that much compute that it needs both. Think building their own battery cells yet still buying cells from everyone…

 

[TN Mtn Man]

Actually, large pieces of stainless are bent routinely. Ever look at most elevator doors? Bent perfectly with no ripple, etc. probably Elon got the idea for the cybertruck in an elevator.

Exactly, and it is also formed/drawn routinely. Stainless steel sink for example. The difference here is using a roll-hardened, thick stainless (reportedly 3mm, ~1/8"). Different than stamping ~.030" annealed sheet.

 

[SOULPEDL]

Yep.


Yep.


It is online and scaling up.

Tesla needs as much compute as they can get, so they are building out both Dojo and Nvidia clusters. (And sounds like they may build out some AMD based clusters as well, if the cost on them works out.)

Could also be parallel plans, and Dojo is not up to the task quite yet. Could explain why FSD is so late (A100->H100 was suppose to be Dojo?). Tesla confidence and progress lately appears to be much improved, so perhaps Dojo is no longer Plan A as Plan B is working enough for FSD launch. (Could also be I'm just now catching on. ) Hopefully we hear an update on the next conf call.

 

[Singuy]

Yep.


Yep.


It is online and scaling up.

Tesla needs as much compute as they can get, so they are building out both Dojo and Nvidia clusters. (And sounds like they may build out some AMD based clusters as well, if the cost on them works out.)

I think Elon mentioned that they were spending about $2B/year on compute for AI training. (I don't know if that is all hardware costs, or if that includes energy use as well.)

I imagine when Elon say those things, he means the entire spent on AI, which includes power and engineering hours.

 

[Cosmacelf]

Telsa has installed this automated/robotic metal brake to bend the stainless steel for Cybertruck: (Trumpf TruBend 5320)

Joe Tegtmeyer on Twitter: "Today at Giga Texas, a large delivery with a Trumpf Truebend 5320 machine was delivered. Here are a few images of the delivery and a link to the company website showing this exact machine in action! https://t.co/3Y9wSJkhkI https://t.co/dau5cUd2ZC" / Twitter | June 09, 2023​

Here is this machine in action: (Robots building Robots...)

TRUMPF Automation: BendMaster with TruBend 5320 - Demo

I hope people realize the video is 13 years old! Not exactly cutting edge technology. Also the video showed bending of pretty flimsy, thin stock.

 

[SOULPEDL]

I hope people realize the video is 13 years old! Not exactly cutting edge technology. Also the video showed bending of pretty flimsy, thin stock.

Plus I'm thinking most Cyber Panels would only need a couple bends each, if at all.

 

[Singuy]

Could also be parallel plans, and Dojo is not up to the task quite yet. Could explain why FSD is so late (A100->H100 was suppose to be Dojo?). Tesla confidence and progress lately appears to be much improved, so perhaps Dojo is no longer Plan A as Plan B is working enough for FSD launch. (Could also be I'm just now catching on. ) Hopefully we hear an update on the next conf call.

Dojo is what Tesla said they are, which is using the D1 chips. It most likely took Tesla half a year to set up their H100 cluster. These are not amazon primed hardware that works a week later. A lot of time went into planning and design just to have the racks online. Then they have to optimize their software on the h100s as it's an architectural change, tho Nvidia will try to make it as easy as possible.

Just from what Tesla have been saying, Dojo is currently used for v11 training as it's used for video auto labeling. Diffusion is most likely not working well on Dojo yet. I imagine they have freed up all their Nvidia hardware for end to end while Dojo works on v11. Until v12 is out and is proved to be just as safe as v11, they will continue to roll out v11 patches to increase safety.

 

[Krugerrand]

Actually, large pieces of stainless are bent routinely. Ever look at most elevator doors? Bent perfectly with no ripple, etc. probably Elon got the idea for the cybertruck in an elevator.

Yeah, it happens allllll the time for large vehicle class A panels and just perfectly .

Have you missed the pictures of the Cybertruck panels? Perhaps this is a case of having different definitions of perfect.

Let me give it to you straight. You can only bend SS so much before it cracks. That bend is also now a weak spot even if it doesn’t crack.

There is the size, thickness, composition and hardness of the piece of SS to consider. Have you any idea at all how big that side body panel is!?

Do you know how much bend or fold Tesla needs particularly at the edges of the panels for attachment purposes? You understand that edge bending is even more precarious, right? I’m pretty sure Tesla panel edging isn’t nearly as wide as in that demo video. And length of that edge bending plays a role; too long or too short and delicate = really hard.

Then spring back needs to considered during the process followed by lack of any give when attaching to other pieces. If the bends of two pieces don’t match precisely you can’t simply do some hand rework to make them fit like you could with thin aluminum or steel body panels.

Then there’s the fact you have to bend with the grain of the SS. I guarantee that before Tesla no foundry ever made rolls or sheets of SS wide enough with grain direction matching the size of Cybertruck side panels in mass production quantities.

Last I looked, elevator doors were flat, weren’t attached to many other elevator doors of various sizes and shapes to create a vehicle that must move at speeds from 1-100+ mph over various terrain torquing those elevator doors in multiple directions at one time. And let me check for those elevator door crash tests online…

I can go on, but I think you get - Tesla isn’t making an elevator door.

 

[Usain]

I think Elon mentioned that they were spending about $2B/year on compute for AI training. (I don't know if that is all hardware costs, or if that includes energy use as well.)

I would hope that includes not only energy use, but also expenses to pay the engineers who build and maintain the systems. Plus it would include the cost of the data center itself.

$2B per year on hardware alone doesn't sound possible.

 

[Artful Dodger]

I know this Forum frowns upon poetry, but I can not resist this gem:

"If you can see the sky, you can contact your family and friends with your usual smartphone, even at sea. If you can see the sky, you can get information with your usual smartphone, even in an emergency. If you can see the sky, you can connect anywhere."​

This connectivity will of course apply to Tesla cars* too, since they already come equipped with 4-G modems similar to the ones in smartphones.

Domo Arigato, Elon!

* World-wide coverage, except where prohibited by local laws.

 

[SOULPEDL]

Dojo is what Tesla said they are, which is using the D1 chips. It most likely took Tesla half a year to set up their H100 cluster. These are not amazon primed hardware that works a week later. A lot of time went into planning and design just to have the racks online. Then they have to optimize their software on the h100s as it's an architectural change, tho Nvidia will try to make it as easy as possible.

Just from what Tesla have been saying, Dojo is currently used for v11 training as it's used for video auto labeling. Diffusion is most likely not working well on Dojo yet. I imagine they have freed up all their Nvidia hardware for end to end while Dojo works on v11. Until v12 is out and is proved to be just as safe as v11, they will continue to roll out v11 patches to increase safety.

Did Elon mention which system produced his demo drive? Are you suggesting it was H100 born? If so, it spit out a NN end to end that quickly? Wasn't it just commissioned days ago? Likely OT, but it sure feels like we could get some crazy good drives coming soon! Looking forward. Thanks for info to all.

 

[SOULPEDL]

A double MMD? Interesting battle underway...

 

[Artful Dodger]

$2B per year on hardware alone doesn't sound possible.

Do some math. $30K for an H100, which Elon says is the equivalent of 3 A100s. Tesla published how many A100-equivalents they will have online next year. This is straightforward.

 

[2daMoon]

A double MMD? Interesting battle underway...

It's like a Hobbit's "second breakfast"

 

[ggies07]

Nice interview with one professor about what goes in Jeff Danh lab and the partnership with Tesla

EXCLUSIVE: Inside the Tesla-funded Jeff Dahn Research Group

Electric Autonomy finds out what makes the Jeff Dahn Research Group so appealing to Tesla in their goal to build to world's best batteries

electricautonomy.ca

Doing the math.....woah!

In one of the Dahn Group’s labs there are several long shelves holding small cardboard boxes. Each box contains 500 battery cells with a combination of letters and numbers scrawled on the outside in black sharpie.

“It’s basically the history of lithium-ion cells back to 2016 when the Tesla project started,” explains Metzger. “I don’t think any other laboratory would have something like that. This is really unique.”

Metzger is open about the goings on in the research group — surprisingly so — but he doesn’t allow photographs of the dead storage boxes. This is because the codes are the chemical compositions of every battery the Dahn Group has worked on for Tesla.

Another nice tidbit:

Meanwhile, the NMC battery in the desk line-up is already famous as Dahn and Tesla’s “million-mile battery.” But, Metzger says, it’s due for a major name change.

“This cell it has now 19,500 cycles [and counting]. Each cycle is 300 kilometres. So, if it were at 20,000 cycles it would be 6 million kilometres.”

That is to say, 3,728,227.15 mile

 

[Singuy]

Did Elon mention which system produced his demo drive? Are you suggesting it was H100 born? If so, it spit out a NN end to end that quickly? Wasn't it just commissioned days ago? Likely OT, but it sure feels like we could get some crazy good drives coming soon! Looking forward. Thanks for info to all.

V12 came from A100s.

Just remember all novel training techniques works on Nvidia hardware first as it's what the researches use to validate their thesis.

 

[SOULPEDL]

V12 came from A100s.

Just remember all novel training techniques works on Nvidia hardware first as it's what the researches use to validate their thesis.

Bullish TSLA AI.

 

[Max Plaid]

It's like a Hobbit's "second breakfast"

There's no such thing, that's was made up for the film adaptation of LoTR, along with other awful ideas like "dwarf tossing"

Visually impressive films, but they destroyed the source material... and don't get me started on The Hobbit films

Elen síla lúmenn’ omentielvo...