FSD v12.x (end to end AI)

[kabin]

From the left lane? I'd say that it's certainly unusual to swing so far out before a turn. Also, the signal was unusual in how high it was. Why did the signal look orange? I thought the orange-shifted reds had been addressed. Anyway, I can easily see this being outside the training data - far enough out that the network couldn't extrapolate to cover it.

There should be some aspect of training data commanding the brakes. The lead vehicle brake lights, turn signals, the broadside of the pickup, quickly closing in on a near stationary object in path, ...

Maybe it's another example of the boy's throwing caution to the wind in an attempt to make FSD feel more capable, more responsive, more human like than it really is?

 

[rlsd]

I'm not quite getting your drift, FSD "will have to deal with it" regardless! (If it wants to keep Tesla in business...)

Exactly what you said. FSD has to be fixed in the future version if Tesla wants to sell FSD.

 

[kabin]

In today's earnings call Elon said FSD is no longer training constrained and FSD improvements are 'profound.'

 

[JulienW]

In today's earnings call Elon said FSD is no longer training constrained and FSD improvements are 'profound.'

He also said all Tesla cars had the FSDS trial. I guess he has NEVER heard of 24.8.x software.

 

[JulienW]

Also it is crystal clear that Dojo is DEAD.

 

[MP3Mike]

Also it is crystal clear that Dojo is DEAD.

Based on what?

 

[JulienW]

Based on what?

Elon talked about all the Nvidia chips (A100) online (>30K) and no longer compute constraint. Then said something like they would have >80k of the A100 chips online in the next year.

 

[MP3Mike]

Elon talked about all the Nvidia chips (A100) online (>30K) and no longer compute constraint. Then said something like they would have >80k of the A100 chips online in the next year.

I think he said H100 equivalants. Which includes A100, H100, DOJO, etc... More on the slide deck:

So some of those ~37k equivalents are DOJO. They aren't sharing what the makeup of that is. (There could even be AMD chips in there.)

They even included a picture of a DOJO cluster in the slide deck:

 

[JulienW]

I think he said H100 equivalants. Which includes A100, H100, DOJO, etc... More on the slide deck:



So some of those ~37k equivalents are DOJO. They aren't sharing what the makeup of that is. (There could even be AMD chips in there.)

They even included a picture of a DOJO cluster in the slide deck.

Yeah I see the pic is there but why no mention of Dojo and why use a Nvidia chip number as the standard. (I was working and didn't hear "equivalent")? But still why not say Dojo and equivalent if Dojo is still relevant? Makes no sense to promote Nvidia and NOT your own product Dojo if it is still relevant.

 

[AlanSubie4Life]

I wonder how many miles on FSD (City Streets) they have done. This is info they have not published.

I guess it doesn’t matter. Will be interesting to see that curve dramatically reduce slope soon. Though maybe they will just extend free FSD.

Elon now pumping stock by telling people to drive FSD 12.3.

 

[MP3Mike]

Yeah I see the pic is there but why no mention of Dojo and why use a Nvidia chip number as the standard. (I was working and didn't hear "equivalent")? But still why not say Dojo and equivalent if Dojo is still relevant?

Because nobody, outside of Tesla, knows the computing capacity of a DOJO chip, but everyone, that matters, knows the capacity of a H100. It is just putting it into terms that are comparable to other companies installed training clusters.

Makes no sense to promote Nvidia and NOT your own product Dojo if it is still relevant.

They aren't "promoting" Nvidia. But, in any case why would they promote DOJO, it isn't like they are planning on selling it. All that matters is installed training capacity.

 

[Daniel in SD]

But, in any case why would they promote DOJO, it isn't like they are planning on selling it. All that matters is installed training capacity.

“Dojo also has the potential to become a sellable service that we would offer to other companies in the same way that Amazon Web Services offers web services, even though it started out as a bookstore. So I really think that yes, the Dojo potential is very significant.“

 

[MP3Mike]

“Dojo also has the potential to become a sellable service that we would offer to other companies in the same way that Amazon Web Services offers web services, even though it started out as a bookstore. So I really think that yes, the Dojo potential is very significant.“

But that is selling the service not the chips/servers. And they will probably sell/price it based on some industry standard unit, like H100 equivalents. (So people can properly compare availability/pricing between cloud services.)

 

[AlanSubie4Life]

“Dojo also has the potential to become a sellable service that we would offer to other companies in the same way that Amazon Web Services offers web services, even though it started out as a bookstore. So I really think that yes, the Dojo potential is very significant.“

It is great that Dojo is alive and well. Nvidia needs some stiff competition.

 

[lzolman]

I think the original idea of Dojo was a totally custom designed chipset, but now they've pivoted to using off-the-shelf hardware instead. Why not, it probably saves them money. Whatever "Dojo" becomes, it's just not the same Dojo, but that doesn't mean it won't do the same work or better...

 

[Daniel in SD]

It is great that Dojo is alive and well. NVidia needs some stiff competition.

This was last year. Happier times.

 

[kabin]

I think he said H100 equivalants. Which includes A100, H100, DOJO, etc... More on the slide deck:

View attachment 1041100

So some of those ~37k equivalents are DOJO. They aren't sharing what the makeup of that is. (There could even be AMD chips in there.)

They even included a picture of a DOJO cluster in the slide deck:
View attachment 1041101

Last I heard they were having trouble with Dojo reliability/connections.

 

[kabin]

“Dojo also has the potential to become a sellable service that we would offer to other companies in the same way that Amazon Web Services offers web services, even though it started out as a bookstore. So I really think that yes, the Dojo potential is very significant.“

The end of the conference call got even wackier with Elon and team promoting the millions of sold Tesla vehicles for use as a future supercomputer revenue. This from a guy who tried to save money by cutting all connections and rolling Twitter's data racks onto a rental truck.

I don't think anyone is crazy or desperate enough to put the company jewels in Elon's hands.

 

[kpanda17]

In today's earnings call Elon said FSD is no longer training constrained and FSD improvements are 'profound.'

Feel 12.4 will be rhe second profound step after 12.3
Seems obvious, but we might be talking about a logarithmic step
Wow

 

[Sharps97]

Dojo was a waste of money. Building chips in house is both difficult and expensive. And rarely necessary since NVidia, Intel, and AMD do custom IP designs for various clientele which would allow the use of decades of architectural experience that Tesla did not have at the time (and does not have now). And if Elon mandated any part of that design based on his intuition, not only is it useless for allowing 3rd parties to use (since it would lack a robust software infrastructure even approaching CUDA with questionable architectural trade-offs that don't make sense), it may not even be fit for purpose (at least on a dollar per watt / performance) relative to either H100, MI300, or even Intel's latest offering. And at least all three companies (NVidia, AMD, and Intel) have experience with designing this type of hardware.

The only 'benefit' to Dojo now is that unless you are a preferred customer, you aren't getting H100 any time soon and MI300 is just starting to ramp up (but has a much weaker software ecosystem). Best bet would have been to prototype Dojo, keep it small scale, and stop investing in it once you start to acquire the real deal, as it were.

The only thing left for Tesla to do now is to re-think the entire FSD approach into something that might work any time soon and off to the races =)