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

[Fact Checking]

what struck me about it was at the very end where the oncoming car looks to be on the center line and the projected path appears to be going into the oncoming lane. ouch!

Path prediction is most important to estimate road curvature and safe, comfortable speeds, as it takes time to brake in a comfortable fashion.

Those overlapping paths visualized are not a problem, as the car will never drive there: the near distance lane markings that determine steering and lane centering are always accurate.

So imagine those extended paths with a different color and "dotted lines", they only matter to gradual, comfortable speed reduction, not to car positioning.

 

[scaesare]

huh? Nvidia Xavier has 16 GB LPDDR4

You clearly don't understand what on-die SRAM brings to this solution.

 

[CyberDutchie]

I guess I shouldn't be surprised that the earnings report is tomorrow and literally no one is talking about the stock in this thread?

That’s because it’s already priced in

 

[BioSehnsucht]

Regarding NVidia vs Tesla performance

Tesla HW3 : 144 TOPS (quoted as NN specific compute power, not including CPU/GPU) at 72W, 2 TOPS/W performance, costing 20% less than HW2.5 compute board
Tesla HW2.5 : < 12 TOPS at 57W, < 0.21 TOPS/W performance, costing ??
NVidia Drive Xavier : < 30 (20 GPU / Tensor) TOPS at 30W, < 1 TOPS/W performance, costing ?? (Tegra Xavier dev kits are $1300 on amazon but I'm sure actual cost is way below that)
NVidia Drive Pegasus : < 320 (260 GPU / Tensor) TOPS at 500W, < 0.64 TOPS/W performance, costing ?? (~RTX 2080 performance x2 plus Tegra Xavier x2 easily puts it over $2000, probably more)

I say the NVidia solutions are all less than claimed because there's no way a Tesla NN actually runs at claimed TOPS (vs HW3 which would) when considering the lack of architectural features needed to prevent stalls and work with batch size of 1, etc. I suspect they might not use batch size of one and instead take the latency hit for better overall throughput on HW2.5, but I could be wrong.

In fact, we can use Tesla's own data here to figure out the effective TOPS of HW2.5. They claim that HW3 is 21x faster than HW2.5 running the same NN, and this means they get ~6.86 TOPS from a claimed 12 TOPS performance, making real world performance only 57% of claimed. The newer NVidia chips might improve things somewhat but they probably won't run batch size of 1 as well, and they will never handle huge NNs as well (due to memory latencies and lack of large addressable SRAM memory on chip). As memory latency only gets worse in terms of clock rate (as it will be similar in terms of absolute time) the newer NVidia chips may not see particularly massive improvements when running small batch size large memory size NNs, even as their clock rates and number of compute units increases.

Plus NVidia is combining all sources of computing power to arrive at their TOPS ratings (CPU, GPU, and any Tensor cores), which is disingenuous from the perspective of running a single large NN. Ignoring the CPU cores, you get 20 TOPS on Xavier and 260 TOPS on Pegasus. The split of regular GPU and Tensor core power makes things even more complex...

So giving a generous 80% efficiency to NVidia's newer solutions when running Tesla's NN would mean Pegasus would be getting effectively 208 TOPS (still more than HW3) at 500W (still too much) for a horrendous 0.416 TOPS/W. If you scaled back that solution to 144 effective TOPS then it would still be 346W power consumption. Limiting it to 100W you'd be looking at only 42 TOPS. Hell limiting the claimed 320 TOPS to 100W would net you only 64 TOPS.

So no matter what, NVidia is hot and power hungry, and not a suitable candidate for Tesla's needs, even if they can reach the needed performance when unbounded by thermal and power constraints.

 

[lklundin]

and I thought short were smart lol
View attachment 399729

I can easily imagine some weak longs who become unsettled enough by these superficially meaningful statements, to sell or not buy more TSLA.

So I wouldn't brush it off as ridiculous.

 

[jerry33]

No, it doesn't. It assumes that we never figure out what the nature of the problem is.

Like a politician?

 

[scaesare]

I’d be shocked if there isn’t some flexibility in their chip(indeed, such flexibility seemed to be confirmed since they stated they have support for tanh and sigmoid activations). Karpathy wouldn’t want to be locked into one architecture forever.

The chip is just a giant function accelerator. The values are loaded at runtime and changeable.

Like your 4-function calculator. Despite fixed functions, you can solve a variety of problems based on the values you plug in to it.

The trick is knowing the funtions you need in order to commit to silicon. Fortunately this is straight-forward for the NN's they plan to be using.

 

[DaveT]

I was able to see the Tesla ridesharing app from a Tesla employee at the event. It's much better than the screenshots in the presentation. The app actually shows all the local Tesla cars on the map (like Uber/Lyft) app and then you can "summon" the car and see the car coming to your location. After you finish using the car, there's a "Return car" button.

 

[neroden]

OT

Please tell me why this negates Buddha's and some 12 step observations that the only person you can change is yourself.

Oh, it doesn't. I don't know why I try to explain things to people who don't want to learn; I think it's a compulsion.

Whether to tone it down or not is a tradeoff between the social value of "shut up because all you're doing is making other people angry at you, they don't want to hear it" and the value of having been the one who pointed out that the Emperor had no clothes, just in case someone was actually listening.

I'm always #2. Something in my brain makeup means I get zero mental reward for the first (and walk away angry at the stupidity and worthlessness of humanity) and some mental reward for the second (and I walk away feeling that at least I've done my duty, and humanity can go hang). Sense of responsibility probably.

 

[StarFoxisDown!]

Would love to know what the amount of shares short are heading into Q1 earnings.....gotta be near an all-time high.

Also while I don't necessarily agree with Neroden's pessimism.....there should have been more focus at the event on how certain FSD features, when released even under human supervision, will set Tesla's very far apart from the competition and increase value and demand for Tesla cars. It felt like there was almost too much focus on Robotaxi and they missed an opportunity to show how the advance FSD individual features will drive demand and increase profits/margins. It wouldn't have hurt to make it clear that by early next year or possibly end of this year, Tesla owners who purchases FSD(and it's Hardware 3.0) will be able to navigate from their garage to work, to restaurant, etc.. while the car does all of the driving and the caveat is that they will still need to pay attention(have hands on wheel).

 

[humbaba]

Once an outlier is found, it's not hard to imagine similar but different cases, so the system can be gamed this way (this is one way where simulation can help because simulation helps if it can be imagined). Also, the fewer the instances of a particular category, the less likely they are to be encounters. If we take the lava flow example, the system doesn't know that it's a lava flow, but it does know the size (visual) that it's substantial (radar), and that it is stationary (visual), so it can fall back to the stop if something big is ahead and not moving. If there was thermal imaging, it would also know about heat. There is some question as to what would happen around a blind curve, but a human driver wouldn't fare any better and maybe worse due to reaction time.

not a criticism of your post, just a response to the "reaction time" comment. While undoubtedly true in principle, I have to wonder why my Tesla has such sluggish reactions to traffic crossing in front of me (e.g., oncoming traffic turns left across my traffic lane).

The Tesla is so slow in hitting the brakes that I have plenty of time to see the situation and press on the accelerator to avoid the unnecessary braking. Unnecessary? Yes, because usually the vehicle has left the roadway before braking starts -- and in the counter examples the vehicle is so far ahead that there's no point in braking. If it was a situation where I would have braked, in every case I would have slowed down well before the Tesla does.

Because the camera-to-brake loop is much faster than I could hope to be my only conclusion is that the currently deployed neural net cannot handle crossing traffic. One could always argue that the development version is better, but I'll fall back on the Missouri motto: "Show Me".

Path prediction is most important to estimate road curvature and safe, comfortable speeds, as it takes time to brake in a comfortable fashion.

Those overlapping paths visualized are not a problem, as the car will never drive there: the near distance lane markings that determine steering and lane centering are always accurate.

So imagine those extended paths with a different color and "dotted lines", they only matter to gradual, comfortable speed reduction, not to car positioning.

Did you look at the video? Its at the very end and clearly shows the expected path as going through the other car. My bet (based on experience with autopilot in similar situations) is that the video terminated with a disengagement. But in the best case you are saying the actual path taken would have been sharper than projected resulting in a less than smooth ride for the occupants.

Hmm... I may seem rather negative here so I'll repeat, I love autopilot. But it does have warts

 

[azaz]

@DaveT would be interested in your longer thoughts on the event and if you got a FSD demo

 

[Nocturnal]

Would love to know what the amount of shares short are heading into Q1 earnings.....gotta be near an all-time high.

Also while I don't necessarily agree with Neroden's pessimism.....there should have been more focus at the event on how certain FSD features, when released even under human supervision, will set Tesla's very far apart from the competition and increase value and demand for Tesla cars. It felt like there was almost too much focus on Robotaxi and they missed an opportunity to show how the advance FSD individual features will drive demand and increase profits/margins. It wouldn't have hurt to make it clear that by early next year or possibly end of this year, Tesla owners who purchases FSD(and it's Hardware 3.0) will be able to navigate from their garage to work, to restaurant, etc.. while the car does all of the driving and the caveat is that they will still need to pay attention(have hands on wheel).

Info about how Tesla owners can use these systems while they drive would have been very nice to know.

 

[neroden]

Regarding the “game, set, match, Tesla...” comment I think Tesla is so far ahead of competitors that they are confident enough to show their cards on stage by delving deep into how their hardware is designed and software is implemented.

FYI, as someone with some math / CS background, I didn't think they went very deep. The only thing on the software side which started to go deep was the one slide of the neural network architecture (the one which looks like a flowchart), which was too small to read. This was a broad architectural outline, nothing more. At this level of generality, it's also really the only possible architecture. There were a few details scattered here and there, but nothing trade-secret-like.

(OK, "some math/CS background" may be an understatement. I only have a bachelor's degree in math, but both my parents are math professors and I've studied all kinds of random math and CS stuff extracurricularly, so I actually have absorbed an awful lot of stuff from the "atmosphere" around me.)

 

[Causalien]

Would love to know what the amount of shares short are heading into Q1 earnings.....gotta be near an all-time high.

Also while I don't necessarily agree with Neroden's pessimism.....there should have been more focus at the event on how certain FSD features, when released even under human supervision, will set Tesla's very far apart from the competition and increase value and demand for Tesla cars. It felt like there was almost too much focus on Robotaxi and they missed an opportunity to show how the advance FSD individual features will drive demand and increase profits/margins. It wouldn't have hurt to make it clear that by early next year or possibly end of this year, Tesla owners who purchases FSD(and it's Hardware 3.0) will be able to navigate from their garage to work, to restaurant, etc.. while the car does all of the driving and the caveat is that they will still need to pay attention(have hands on wheel).

The question I ask is, what is pushing TSLA to show their hands early now as opposed to waiting till it is ready to reveal like apple. The robotaxi should've been the "one more thing" moment. It is not like a competitor have achieved it before them forcing them to reveal something.

 

[Vitold]

Overall I like the TN story. So what if it will take 4, 6 years instead of 2 to get there as long as cars sell they can afford delays. And, opposite is also true - exponential growth does not matter as much due to upcoming TN.

TN and car production complement each other quite well from investors perspective. Less so from owners point of view as I doubt that current cars will ever be allowed to drive themselves due to side camera repeater’s short range and location.

 

[lklundin]

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.

If you are competent, ambitious and young (so you can handle the workload), then Tesla is your career's fast lane. You will have 'Tesla' on your CV and you will be learning a _lot_ by pushing the envelope together with others of your caliber.

So I can easily see how Tesla can attract all the talent - in any field they venture into.

I am going to have to watch Bannon and Karpathy a second time, this was great stuff.

PṠ. One more thing: Working for Elon Musk like Bannon's and Karparthy's teams is also amazing, because like in science you get to work with a solution derived from first principles, the difference being that with Elon Musk, what you create is then very quickly introduced to the real world. Bannon hinted at this, this is something surely attracts real, ambitious people.

 

[Fact Checking]

In fact, we can use Tesla's own data here to figure out the effective TOPS of HW2.5. They claim that HW3 is 21x faster than HW2.5 running their code, and this means they get ~6.86 TOPS from a claimed 12 TOPS performance, making real world performance only 57% of claimed.

Fantastic observation!

The newer NVidia chips might improve things somewhat but they probably won't run batch size of 1 as well, and they will never handle huge NNs as well (due to memory latencies and lack of large addressable SRAM memory on chip). As memory latency only gets worse in terms of clock rate (as it will be similar in terms of absolute time) the newer NVidia chips may not see particularly massive improvements when running small batch size large memory size NNs, even as their clock rates and number of compute units increases.

Exactly, and it's worse: the larger neuron count of larger NNs makes the "batching cost" higher on more parallel non-deterministic chips like Nvidia Turing, because of the batch completion problem; if you have 10,000 neurons in a layer that execute in parallel, then this layer is only completed once the slowest computation thread completes. As parallelism increases so does the distance between fastest and slowest computation increase. This means that stalls get super-linearly escalated as parallelism increases.

The only solution to scale to massively parallel computing hardware is either:

• multithreading, which has limited use in vision computations as you don't want to delay/overlap processing, because frames come only every 10 milliseconds,
• or highly deterministic processing where a parallel computation that uses the internal SRAM and registers will execute at exactly predictable cycles which can be instruction scheduled at compilation time. (I.e. somewhat similar to Intel's VLIW Itanium, but done right.)

Tesla has chosen the latter, which allows them to scale their computations almost linearly and (eventually) allows them to utilize the hardware close to 100%.

But Nvidia execution is not deterministic to such a degree, and I'd not be surprised if the 320 TOPS Nvidia chip could process the big AK-NET at less than 100 TOPS effective performance - and with vastly more power used.

 

[Causalien]

FYI, as someone with some math / CS background, I didn't think they went very deep. The only thing on the software side which started to go deep was the one slide of the neural network architecture, which was too small to read. This was a broad architectural outline, nothing more. It's also really the only possible architecture. There were a few details scattered here and there, but nothing trade-secret-like.

(OK, "some math/CS background" may be an understatement. I only have a bachelor's degree in math, but both my parents are math professors and I've studied all kinds of random math and CS stuff extracurricularly, so I actually have absorbed an awful lot of stuff from the "atmosphere" around me.)

Have to disagree with you here. Architechting that sort of software is most of the work. From what I am seeing, it is software that works across at least 5 different type of compute units and Firmwares that runs of each of them. Normal software projects don't deal with firmwares and deals with only the CPU and maybe, sometimes the GPU.

 

[jerry33]

Hmm... I may seem rather negative here so I'll repeat, I love autopilot. But it does have warts

No doubt, but we don't have many of the new hardware out in the wild yet (and I believe the ones that are out are running the older software which isn't optimized), and the demos are using the development version of the software.