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

[Boomer19]

how can anyone that knows anything about tesla/tsla history not see this as a buying opportunity

 

[Fact Checking]

Don’t forget, as demonstrated yesterday, that cameras easily render a 3D pic as well.

Yeah, tried to cover that via:

• "If your LIDAR creates a 3D point cloud then that removes a big practical incentive from having visual networks provide really good 3D data."

The consequence is that if you are using camera and radar vision then you are forced to develop really good 3D mapping/ranging via neural networks - which 3D data is hidden in plain sight, as wonderfully summarized by @KarenRei:

I thought it was great how they illustrated how completely pointless LIDAR is, given how many different ways they already have for measuring distances, including the point cloud illustration.

1) Inherent object sizes - The neural net knows how big various objects "should" be, and thus distance is determined (usually with quite high accuracy) based on how big they appear. This is the same method humans use with one eye closed.

2) Parallax between cameras - where image data is matched up between different cameras, the relative positioning in the frame determines the distance. This is why we humans have two eyes - binocular vision.

3) Parallax of motion - the relative rate at which corresponding image data moves between consecutive frames determines its distance. This is akin to animals that bob their heads to assess distance.

4) Radar ranging - Radar returns a low-resolution distance map of the world ahead of the vehicle, regardless of weather (unlike LIDAR, which is even more weather sensitive than passive vision). Up close, ultrasonic sensors also come into play.

Tesla unifies data from four separate range-finding methods to assess distance - with a much higher max scan rate and resolution than LIDAR, at that.

But another argument that is worth stressing is not just that LIDAR is dead when compared to NN assisted camera vision, it's also money spent on the wrong thing: the money a carmaker spends on LIDAR is much better invested in sensors and computing capacity.

I.e. if we take the CoGs budget of say a $40,000 car, and we assume LIDAR costs come down from $50,000 to $5,000 due to new technologies and economies of scale, even in that case spending $5,000 on a LIDAR sensor reduces the safety of the car, because it takes away money from other safety features. It's a zero-sum game: a $5,000 LIDAR is crowding out $5,000 of safety features!

LIDAR would have to reach the cost of passive camera network, which is in the ~$100 range, to be a useful sensor to add to a mass manufactured car that improves safety.

This is a basic economic safety argument that Waymo and the other LIDAR proponents are missing.

 

[BestRadar]

More than 75% of articles from investors yesterday are not very positive. I see after hours trading is currently down $3 to $258 so I don't think investors were that impressed.

 

[Prunesquallor]

I was going to post the same thing. The UAE already has a fleet of Tesla taxi and they have a desire to be at the cutting edge. This would be a great place to test and generate data on the Tesla Network.

And dodging the street-racing Ferraris would be a good edge case.

 

[anthonyj]

Can’t wait for a cap raise. Would kill a lot of FUD. Seems pretty much guaranteed now since Elon guides for cash flow neutral here on out

 

[Dan Detweiler]

Seems to me that everything the stock market says they want to see comes down to battery cell production and hence the ability to put products on the roads or in the storage products. Everything that will prove beneficial for the long term health of the company is based on technology like chip development, software and FSD.

So...in short term we gotta get Panasonic up and running at full steam. That will get the analysts at least partially off our backs and then let the tech do its thing long term.

See...it's easy. Now just go do it. They should hire me as CCO (Chief Cheerleading Officer)

Dan

 

[Prunesquallor]

Irony is that if Robotaxi EVs come, Middle East petrostate economies' are going to get utterly rekt.

Some of them see it coming and are course-correcting.

 

[Dan Detweiler]

Oh and BTW no mention here of the snake charger? Elon said this was a trivial problem. Will we possibly see that soon? I'm not too bothered but the internet would go nuts if you could order one for your garage.

The thing is there is no real need for it until the Tesla Network is up and running and/or FSD is available to customers. So, why waste time on something they could get up and running in a very short period of time. Kind of a waste of resources at this point.

Dan

 

[Fact Checking]

I’m just fascinated to see what all the others say in response. Musk has dissed LIDAR for years but one hopes that after this, these same analysts will now ask the question back to the other players of the game.

UBS perma-bear "Colin Langan" is (unsurprisingly) doubling down on their LIDAR misunderstanding and FUD:

"UBS analyst, Colin Langan, reiterated a Sell rating and $200 price target on Tesla after attending the company's Autonomy Day. The highlights include a target date for he deployment of "feature complete self driving" capabilities, the launch of the Robotaxi in 2020 and featured the in-house designed TSLA chip has more power and uses less silicon than the prior gen NVDA chip."

'The analyst stated "The primary sensor is often vision, but most experts believe LiDAR will be needed since it provides more unique data, even if it's small, that makes the entire system safer. Regulators will likely want more sensor data to ensure the highest level of safety. TSLA may also need more nonsimulated testing. TSLA reported 0 AV miles driven in California in 2018 vs. GOOG which reported 1.2m".'​

There's two false claims in these two short paragraphs already:

• He is misconstruing the fact that Tesla did not opt to report disengagements in California into a false claim that Tesla only performs "simulated testing". In reality Tesla got disengagement events from about a billion miles of Autopilot driving, a three orders of magnitude larger data stream than the Google case UBS cites ...
• They also don't realize that even a very cheap ~$5,000 LIDAR sensor will crowd out much more effective sensors, i.e. in Tesla's model LIDAR will actively reduce car safety and kill people.

He is hiding behind the "most experts believe" weasel words and false appeal to authority, instead of using easily verified facts and logic. I absolutely do not want to read his full analyst report: garbage in, garbage out.

In other, completely unrelated news, UBS is apparently positioning to IPO Waymo:

Alphabet's self-driving car business could book $114 billion in revenue in 2030, says UBS

"On the heels of Waymo’s commercial launch on Wednesday, investment bank UBS estimates that Alphabet’s self-driving car unit will reel in $114 billion in revenue in 2030."​

Which I suspect explains why they absolutely have to "misunderstand" Tesla's plans.

 

[Singer3000]

UBS perma-bear "Colin Langan" is (unsurprisingly) doubling down on their LIDAR misunderstanding and FUD:

"UBS analyst, Colin Langan, reiterated a Sell rating and $200 price target on Tesla after attending the company's Autonomy Day. The highlights include a target date for he deployment of "feature complete self driving" capabilities, the launch of the Robotaxi in 2020 and featured the in-house designed TSLA chip has more power and uses less silicon than the prior gen NVDA chip."

'The analyst stated "The primary sensor is often vision, but most experts believe LiDAR will be needed since it provides more unique data, even if it's small, that makes the entire system safer. Regulators will likely want more sensor data to ensure the highest level of safety. TSLA may also need more nonsimulated testing. TSLA reported 0 AV miles driven in California in 2018 vs. GOOG which reported 1.2m".'​

There's two false claims in these two short paragraphs already:

• He is misconstruing the fact that Tesla did not opt to report disengagements in California into a false claim that Tesla only performs "simulated testing". In reality Tesla got disengagement events from about a billion miles of Autopilot driving, a three orders of magnitude larger data stream than the Google case UBS cites ...
• They also don't realize that a $5,000 LIDAR sensor will crowd out much more effective sensors, i.e. in Tesla's model LIDAR will actively reduce car safety and kill people.

He is hiding behind the "most experts believe" weasel words and false appeal to authority, instead of using easily verified facts and logic. I absolutely do not want to read his full analyst report: garbage in, garbage out.

He does have one point that strikes a chord with me and that’s on the regulators. Tesla Network is going to p*** off A LOT of entrenched interests. Regulatory approval is the piece of the puzzle that Elon may well be the most over optimistic about.

 

[Pras]

I'm not aware of any formal estimates that are public, but for total HW 2.5 compute module production costs I'd guess "a couple of hundred dollars", of which the Nvidia chips would have been around $100 I think, based on super bulk pricing and a good relationship to Nvidia.

But the big business risk for Tesla was not in current Nvidia costs, but in the FSD upgrade path given their decision to rely on ~100 TOPS performance neural networks: the next gen Nvidia automotive chips able to compute such networks cost well over $1,000, and Nvidia is not unhappy to squeeze customers where they can squeeze them, as we've seen it in the GPU space. So had Tesla not developed their own AI chip they'd be looking at a $1,000-$3,000 per unit cost I believe, which is about 2-6% of the Model 3 margins - an unacceptably high cost.

Also it's unclear whether Nvidia GPUs are able to effectively compute the big neural networks Andrej Karpathy has created, high efficiency GPU workloads and benchmarks generally try to reduce bus traffic - a memory-limited GPU performance can easily be only 10% of benchmarked performance. I.e. instead of the advertised 320 TOPS of the top of the line Nvidia monsters Tesla might be looking at an effective performance in the ~30 TOPS range only - at much, much higher unit cost and power consumption.

I don't think the GPU based neural network accelerators are competitive at the moment - the only potentially competitive chip is Intel's MobilEye chip, which has about half the TOPS performance of Tesla's chip.

So yes, despite the 14nm process size and TOPS disinformation campaign of the TSLA-Q-Anon cult, Tesla's new AI chip very likely offers unparalleled, unmatched performance for large FSD neural networks, at any price.

Thanks. Also important to know is that nvidia could have negotiated higher price on the revenue made from robotaxi, knowing that Tesla’s business model is dependent on them.

 

[jerry33]

So they believe that Tesla can make more money in 10 years than the top 15 most profitable public companies *combined*? Umm.. ?!?!

I guess it would be a fail if they only made as much as the top two companies combined

 

[printf42]

OT

I don't get it... NVIDIA's entire advantage right now is massively parallel floating point multiplication logic unit machines. That's what a GPU is. Why is Intel better?

When I say Intel, I mean their Nervana or Mobile eye might have some potential in CNN inference front, definitely not their x86 product line.
GPUs are optimized for parallel computing, but in a more general sense, its not optimized for convolutions and is now limited by memory bus bandwidth. Sure they can do convolutions, but they won’t be nearly as fast and efficient as dedicated hardware can do.

 

[Drax7]

The technical presentation was phenomenal, the business model will develop over
Time and may go through many iterations. Nonetheless Tesla has further widened its
Competitive moat . At this point vertical integration is of huge benefit.

In the short term it seems the stock price is pegged
To Q1 earnings and guidance from there onwards.
Can they increase production or is battery supply a limiting
Factor.

Moreover, Is their mission statement changing somewhat
Or just expanding.

 

[Bladerskb]

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.

I should point out that First of all 1 Turing Chip runs around 130 TOPS and that's just a single chip and you don't even need to access other GPU for additional power. In addition, Nvidia now uses NVLINK and is now on version 2. Drive PX 2 which didn't use NVLINK had the bottleneck problem. That has all be effectively removed.

Lastly, They showed the half Drive Px 2 (10 tops) running Tesla's NN at 110 FPS then ran the same network on the FSD computer and got 2300 frames.

Give the Drive Px 2, a 21x boost and you get 210 TOPS needed to get to Teslas 2300 frames.
But now remember this Drive Px 2 is two generations old. Nvidia is now on the Turing Architecture Tensor Core GPU. Which means the operations will already be orders of magnitude faster on Turing chips just because of improvement on the architecture. For an improvement from Turing architecture could be only need 100 TOPS to achieve 2300 frames

Edit: I should also point out the Drive PX 2 Tesla uses and did its comparison with didn't have Tensor Cores which started with Volta. "Tensor Cores can accelerate large matrix operations, which are at the heart of AI, and perform mixed-precision matrix multiply and accumulate calculations in a single operation."

 

[Artful Dodger]

All cars build after ~November 2016 are HW2 capable and are trivially upgrade-able to HW3

Yeah not quite, its actually earlier than that. Tesla halted production of both S and X at the start of Oct 2016 when they severed their relationship with Mobileye. The line was down for approx. 2 weeks to refit for HW2. The computer was designed from that point to be easily upgradable to the FSD computer.

So all Tesla vehicles build in Oct 2016 or later can be retrofited for Robotaxi service with a simple computer swap (which can be done by Tesla Mobile Service).

So Sep 2016 or earlier is AP1 hdw and won't be upgradable to FSD ("cheaper to build a new car than to upgrade"). But the vast majority of Tesla's fleet is already HW2+ thus FSD upgradeable. And of course S/X has the FSD computer since Mar 2019, and Model 3 also has it since as of April.

Cheers!

 

[Dan Detweiler]

He does have one point that strikes a chord with me and that’s on the regulators. Tesla Network is going to p*** off A LOT of entrenched interests. Regulatory approval is the piece of the puzzle that Elon may well be the most over optimistic about.

The regulators are only going to care about one thing...safety. If they give the go ahead to a system and people start getting killed in high numbers using that system then they look bad and their credibility going forward is screwed. If Tesla has billions of miles of actual, real world, driving miles that show beyond a shadow of a doubt that their system is many times safer than human drivers and light years ahead of any other manufacturer in terms of usage and practical applications, they will have no problem pulling the trigger in Tesla's favor.

Dan

 

[Remus]

He does have one point that strikes a chord with me and that’s on the regulators. Tesla Network is going to p*** off A LOT of entrenched interests. Regulatory approval is the piece of the puzzle that Elon may well be the most over optimistic about.

It may not be that hard. You see there are so many different government on this earth, getting approval from one big city is not that hard. Once they demonstrated that it works, it will overwhelmingly benefit the weak and disabled: the blind people, people have brain or heart conditions, elderly. Prohibitions would be seen by these people to rob them of their hope for mobility and freedom. Lawsuits can be filled by civil liberty groups to force the government hands.

 

[AudubonB]

Morning heads-up:

How many times is NN going to get this in its feed? (Thinking....thinking...thinking....Ah ha! Family taking toddler to dinner. Safe To Pass With Care)

 

[Fact Checking]

He does have one point that strikes a chord with me and that’s on the regulators. Tesla Network is going to p*** off A LOT of entrenched interests. Regulatory approval is the piece of the puzzle that Elon may well be the most over optimistic about.

Fortunately there's a bit of a prisoner's dilemma there: economic regions employing 'robotaxis' will immediately benefit and will be seen as high-tech embracing incubators. So the first mover regulator will reap positive PR and economic advantages, which creates pressure to approve truly safe solutions.

In that way it's somewhat similar to the pressure to reduce corporate tax rates and offer incentives to attract companies.