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

[kengchang]

Something just hit me. Since he is targeting the TN in the US, then sales need to continue focusing growth in the US. So relying on new EU countries or China sales although good for cash flow (higher $ versions), he needs more "units" in US. Two competing agendas - Profit vs more networking units. He said SR+ will be their focus. So should we expect a slow year on the financials, or does he make it up with great margins and FSD sales to keep profits going +? If that is the plan, are we there now and could show profit already in Q1?

Tesla is pushing SR+ due to battery cells constraints...

 

[Fact Checking]

Latest tweet from Elon about HW3 performance:

Elon Musk on Twitter

"Exactly. Also, you can’t actually use computation from a separate GPU effectively, as you get choked on the bus, so most of the computation is irrelevant. High power, high cooling, but low true, usable TOPS. Worst of all worlds."​

This is important: the "300 TOPS" (trillion/tera operations per second) figure from Nvidia's Pegasus is not effective performance to calculate neural networks, it's a synthetic benchmark.

The underlying problem is that real life neural networks used by Tesla are very large, so if they are used in GPUs then the weight data is much larger than the cache, and has to be fetched from DRAM all the time. This slows down execution dramatically and much of the GPU is idling around waiting for DRAM data.

Tesla's HW 3 chip not only has on-chip SRAM of 32 MB, but also has extremely wide buses able to fetch 1 TB/sec data.

This means that Tesla's chip should be significantly faster than even Nvidia's very latest, when running the large FSD networks.

 

[kengchang]

Still scratching my head how they freaken made this chip as a side project and suddenly become as competitive as Nvidia while blowing AMD's attempt out of the water. It's really crazy and I don't think these investors really understand what they were looking at.

ASIC is very simple/different animal compare to CPU/GPU

 

[powertoold]

So here are the timestamps for the people who were questioning me about Tesla's recent FSD development timeline. Per Karpathy, Tesla JUST implemented stuff like path prediction for cloverleafs and cut-ins 3-5 months ago:

Cut-in detection 3 months ago (2:12:44):

Tesla clover leafs only 5 months ago (2:15:57):

 

[Alphacrux]

Still scratching my head how they freaken made this chip as a side project and suddenly become as competitive as Nvidia while blowing AMD's attempt out of the water. It's really crazy and I don't think these investors really understand what they were looking at.

Elon hired the best: Karpathy and Bannon, true experts in their fields. And starting from “first principles”: we don’t need a GPU with Tensor extensions, just a CNN implemented in hardware. Nvidia has a GPU solution because they are in the business of selling GPUs. And oh, btw, their GPU based self-driving board consumes 500 watts.

 

[SOULPEDL]

An interesting point here is that Intel has been struggling with they <10nm chips, as explained here:

The video also explains chiplets which the twitter chip expert bear(who someone here linked) was unaware of.

The main takeaway is that as the transistors get smaller yields have been falling and many chips needs to be discarded. This is improving, but for Tesla going 14nm instead of 7nm in 2018 likely improved their costs and time to scale significantly. 14nm was good enough for now, 7nm or better might be ready for cost/scale in 2021.

Sad that we all get that point, yet FUD will spin it as weak computer. They can only do this for so long.
You realize if he nails L5 capable this year, we'll all be using it just with the nag until cities approve. By the way, how do you hold the steering wheel while its turning I wonder...

 

[kengchang]

Elon hired the best: Karpathy and Bannon, true experts in their fields. And starting from “first principles”: we don’t need a GPU with Tensor extensions, just a CNN implemented in hardware. Nvidia has a GPU solution because they are in the business of selling GPUs. And oh, btw, their GPU based self-driving board consumes 500 watts.

Let's not forget Jim Keller

 

[Singuy]

ASIC is very simple/different animal compare to CPU/GPU

So if it's so simple, why is Nvidia dicking around with Volta tensor cores if they can just get a few guys and build ASIC?

I get that certain ASIC are simple, especially those used to mine bitcoins. But it's completely foolish for Nvidia, supposedly the leader in autonomous driving, lost their biggest costumer and maybe the autonomy leadership just because they didn't dedicate a few guys to do a side project like Tesla here?

 

[loco]

My interpretation of the investor meeting is that Tesla is moving away from the "profitable all quarters going forward" cash cow narrative back towards cash burning growth mode. The real reason for this is that they failed to deliver financially and the official reason (or excuse) is FSD and the need to capture as much of the new market as possible. I think that is good. They should be a growth company with crazy valuation, not a 10x profit stock. That would be very bad for us investors. Capital raise confirmed with the sentence about doing what everyone expects them to do.

 

[SOULPEDL]

Tesla is pushing SR+ due to battery cells constraints...

And goal is "max number of units" for TN in that context specifically.

 

[SOULPEDL]

So if it's so simple, why is Nvidia dicking around with Volta tensor cores if they can just get a few guys and build ASIC?

I get that certain ASIC are simple, especially those used to mine bitcoins. But it's completely foolish for Nvidia, supposedly the leader in autonomous driving, lost their biggest costumer and maybe the autonomy leadership just because they didn't dedicate a few guys to do a side project like Tesla here?

Ya, they're bummin now I bet. Short Nvidia!
Not an advice (no really).

 

[Singuy]

My interpretation of the investor meeting is that Tesla is moving away from the "profitable all quarters going forward" cash cow narrative back towards cash burning growth mode. The real reason for this is that they failed to deliver financially and the official reason (or excuse) is FSD and the need to capture as much of the new market as possible. I think that is good. They should be a growth company with crazy valuation, not a 10x profit stock. That would be very bad for us investors. Capital raise confirmed with the sentence about doing what everyone expects them to do.

I think he is saying when regulators approve and FSD has substantial evidence that it's safer than human drivers with zero interventions, then he'll switch to cash flow neutral to build the Tesla network fleet. But if this unicorn scenario actually comes to fruition then the stock is probably in the 4 figures and we wouldn't even care where Tesla is going when it comes to cash flow..lol

 

[Fact Checking]

An interesting point here is that Intel has been struggling with they <10nm chips, as explained here:

The video also explains chiplets which the twitter chip expert bear(who someone here linked) was unaware of.

The main takeaway is that as the transistors get smaller yields have been falling and many chips needs to be discarded. This is improving, but for Tesla going 14nm instead of 7nm in 2018 likely improved their costs and time to scale significantly. 14nm was good enough for now, 7nm or better might be ready for cost/scale in 2021.

Another point people missed: bootstrapping a new CPU is extremely hard, and the tools and expertise to do this successfully close to on the first attempt (which Tesla apparently managed) is extremely limited.

Intel's tic-tock model applies here, Intel implements new microarchitectures on a robust, proven process - let alone entirely new CPUs:

Tick–tock model - Wikipedia​

On that basis I predicted a 14 nm process back in August 2018:

TSLA Market Action: 2018 Investor Roundtable​

Anyone complaining about Tesla "only" using 14nm doesn't know what they are talking about.

 

[SOULPEDL]

Tesla must hv such a lead to just open the chip specs to the public. I think we saw our first FSD ad.
It may not spin well for todays SP, but a compelling case was made to sell more FSD.

 

[Singuy]

Ya, they're bummin now I bet. Short Nvidia!
Not an advice (no really).

I'm pretty bear on Nvidia. They have been such dicks to the entire industry to the point that they are cornered. Nvidia is trying to sell their autonomous chip to companies without a training fleet that makes a difference unlike Tesla. Data centers and AI training are going ASIC. And the only area Nvidia excels at which is gaming can possibly lose to Google Stadia. They already lost all the consoles except low margin Nintendo. It's becoming a world in which Nvidia is no longer necessary as a company. AMD being the underdog could possibly steal all the gaming marketshare via Stadia and consoles while all companies are creating their own ASIC.

 

[kengchang]

So if it's so simple, why is Nvidia dicking around with Volta tensor cores if they can just get a few guys and build ASIC?

I get that certain ASIC are simple, especially those used to mine bitcoins. But it's completely foolish for Nvidia, supposedly the leader in autonomous driving, lost their biggest costumer and maybe the autonomy leadership just because they didn't dedicate a few guys to do a side project like Tesla here?

Just having the chip is half of the equation (prolly more like 10%) it's the software that's going to solve the autonomy driving not hardware.

 

[printf42]

Elon liked a tweet with this video, where Former Google/Waymo/Uber star self-driving engineer Andrew Levandowski said: the reason he is not using LiDAR anymore is because “not that there is any restrictions on LiDAR (from Google), but I do have restriction, personally, of not doing things that I know that aren’t gonna work”.

It only have 500 views so I guess it’s not shared here yet.

 

[SOULPEDL]

Agree that this is not going to happen overnight everywhere at the same time, but for the location it goes live, I can imagine them taking over Uber+Lyft+Taxi overnight, and slowly takes over private car ownership as well, just because of the economy of not having a driver in the cost structure.

So when one metro area goes online, it could soak up big chunk of supply immediately, so that cars will flow in from other part of the country and affect supply and demand on a much larger scale.

Ya, just move your car to that city. I could see that happening.

 

[Fact Checking]

Per Karpathy, Tesla JUST implemented stuff like path prediction for cloverleafs and cut-ins 3-5 months ago:

3-5 months for the latest iteration of features is what he says in essence, which is an entirely uncontroversial point. This also proves it conclusively that the original claim you made is false:

Tesla's current FSD development only started 3 months ago, per comments from the FSD demo today.

I repeat: it's ridiculous to suggest that Tesla only started working on the current FSD code 3 months ago.

Not sure why you just cannot admit that you were wrong suggesting it.

 

[S3XY]

So here are the timestamps for the people who were questioning me about Tesla's recent FSD development timeline. Per Karpathy, Tesla JUST implemented stuff like path prediction for cloverleafs and cut-ins 3-5 months ago:

Cut-in detection 3 months ago (2:12:44):

Tesla clover leafs only 5 months ago (2:15:57):

And? We all know this. Why do you keep harping on it and what makes you think there is any significance?