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

[JRP3]

Ultrasound/radar/lidar might also be useful in the forward facing direction, if it helps more accurately measure the distance to objects.

They removed radar from the cars and don't use lidar. They won't use it on the bot which is based on the same AI software.

 

[dhanson865]

It's confusing when they should off 4x performance gain from Dojo and then shows off more performance gains elsewhere.

Anyways, when it comes to apples to apples comparison, it's all about performance per watt. 6 Dojo chips uses 100kw of power. Each Dojo chip replaces 6 DGX boxes which are 7.5kw per box. This puts the calculation to be 270kw equivalent of an A100 server vs 6 dojo chips. The efficiency gain is 2.7x. Perhaps this is on par with H100 which again, has a lot of marketing trickery and I doubt Tesla can get the gains Nvidia is claiming. If anything, this is actually pretty impressive considering it's on TSMC's 7nm vs H100's 4nm. So even if it's competitive vs the H100 going into Q1 2023 that's pretty ace in my book.

I am still wondering about the problem Tesla will face on an Nvidia system if the NN is too large to fit. I know catastrophic performance failures happen when you just run out of resources which can slow the process down by 100x or 1000x. I wish Tesla can elaborate a little more here.

Performance per watt isn't DOJO's goal metric. Time to process an autolabel batch is.

That's the apples to apples Tesla cares about. Any metric you pick that isn't actual FSD related code completion is immaterial to Tesla.

 

[Knightshade]

They removed radar from the cars and don't use lidar. They won't use it on the bot which is based on the same AI software.

OTOH, SpaceX does use Lidar, and Elon has been pretty clear he's a right tool for right job guy.

I don't expect they need it NOW- but I can absolutely see precision work that requires a lot more exact distance/speed/depth perception than the current FSD/AP system can offer with their 720p cameras.

If you're stopping for traffic ahead you're gonna stop feet away- an inch or a fraction of one, isn't going to matter. There's a slew of tasks humans do where even smaller differences matter.

Not only that- LIDAR that works for a car moving at high speed and tracking tons of objects is bulky and expensive. Lidar that'd work for this task is a cheap, tiny, part even my cell phone has today.

 

[TN Mtn Man]

I've seen suggestions that the head doesn't move.

Elon regularly calls out the human head as being a very slow gimbal.

IMO 2 forward facing cameras for perspective, one on each side and one at the rear would do the job.

Maybe they only need 4, but the cost difference between 4 and 5 is probably not significant.

Ultrasound/radar/lidar might also be useful in the forward facing direction, if it helps more accurately measure the distance to objects.

Honestly, I wonder how expensive it would be to install 8 (I think that's what the cars have). Seems like more code could be reused since the programming already exists to accurately integrate 3d spaces/objects with those cameras.

Next thought, seeing how TB is designed to grasp/operate the same controls that humans are, and are based on the same AI and general operating computer as FSD, how long before a TB can get in a non autonomous vehicle and drive it? Perhaps not a car right now, but maybe tractors or fork trucks to start.

 

[JusRelax]

An observation I made regarding Dojo, last night, was that they've solved all/most of the core architectural problems for building their own super computer....new and crazy power density /thermal solutions that will scale. They can fairly easily iterate the Dojo "tile" and have all the support systems complete. Sure they might iterate a bus or so, but all this legwork seems to have positioned them well to iterate far ahead of everything else. Tesla 101.

Another note - on their timeline slide they stated that they are already up to manufacturing one tile/day. It seems that Tesla's FSD iterative processing time will continue to shorten in the future until they've reached whatever internal goal they have for it. IIRC (which I could definitely be wrong about), in previous talks they stated a goal of 3 days or something?

 

[TN Mtn Man]

OTOH, SpaceX does use Lidar, and Elon has been pretty clear he's a right tool for right job guy.

I don't expect they need it NOW- but I can absolutely see precision work that requires a lot more exact distance/speed/depth perception than the current FSD/AP system can offer with their 720p cameras.

If you're stopping for traffic ahead you're gonna stop feet away- an inch or a fraction of one, isn't going to matter. There's a slew of tasks humans do where even smaller differences matter.

Not only that- LIDAR that works for a car moving at high speed and tracking tons of objects is bulky and expensive. Lidar that'd work for this task is a cheap, tiny, part even my cell phone has today.

Just curious-what phone has Lidar and what kind of apps is it used for? Something I wasn't aware of. I do have a robot vacuum that has it.

 

[Knightshade]

Just curious-what phone has Lidar and what kind of apps is it used for? Something I wasn't aware of. I do have a robot vacuum that has it.

Iphone (and ipad) has had it for 3 generations now... I would've THOUGHT some android stuff would by now but a quick google says otherwise which is surprising.

Personally I've used it for measuring (~1cm accuracy) and there's even apps to create entire 3D model of a room then use SW to see how it'd look moving stuff around, etc... and the camera app uses it by default for better/more accurate focus in low light.

Apparently you can also use it to create 3D models of objects that you can import to blender and even things like... the Unity engine Tesla uses for their simulator.

And a lot of AR apps use it too... there's an entire dev kit from Apple around AR development using it.

Prob. a ton of other stuff I'm not even aware of.

Range is only 5 meters- pretty worthless on a car... but it'd make precise actions by a bot with hands/arms much shorter than 5 meters and moving at (relatively) low speed pretty accurate.

 

[Singuy]

Performance per watt isn't DOJO's goal metric. Time to process an autolabel batch is.

That's the apples to apples Tesla cares about. Any metric you pick that isn't actual FSD related code completion is immaterial to Tesla.

Uh, yeah the time it takes to process is the performance. The amount of energy to do such a thing is the watt. This is the only metric that matters for compute in the data center. And it's Tesla that is telling you the performance per watt by using pictures of dojo chips and GPU clusters. Notice Tesla spent some time explaining their power usage and cooling solution because this is all datacenter 101. Insane performance while using insane power still need to be distilled down to performance per watt. That's how you compare hardware to hardware. It's like $/watt when it comes to buying solar. There maybe 100 different type of panels but at the end of the day, the end user only cares about $/watt.

Are you saying the H100 will be worst than the A100 or doesn't work at all? This is the only argument you can make by throwing all metrics out the window

 

[BioSehnsucht]

I've seen suggestions that the head doesn't move.

Elon regularly calls out the human head as being a very slow gimbal.

IMO 2 forward facing cameras for perspective, one on each side and one at the rear would do the job.

Maybe they only need 4, but the cost difference between 4 and 5 is probably not significant.

Ultrasound/radar/lidar might also be useful in the forward facing direction, if it helps more accurately measure the distance to objects.

They showed a video clip with six camera views while the bot was maneuvering, and they were still labeled with FSD labels like "B pillar". Since they're porting their existing 8-camera setup from vehicles over to the Bot platform, I wonder if they'll keep all 8 cameras (extra cameras don't cost that much, in the scheme of things) or reduce them.

Assuming 360 degree coverage, 3 to 6 cameras is probably a good expectation, since in the forward direction Bot probably doesn't need to be able to see and process action hundreds of meters out at high velocities (3 of the FSD cameras on cars are all looking forward at different zoom/focal lengths to better handle up close and far away at the same time - you can probably just use one or at most two, perhaps calibrated for very near field and then whatever the normal closest distance camera is from vehicles, rather than using the normal FOVs of FSD cameras only)

I wouldn't worry about stereoscopic vision since apparently their depth from vision process is working just fine with processing motion plus time, at worst case they could just have the Bot move it's head (which might mean it's torso) a bit side to side or forward and backward to get the necessary data, but as long as it remembers what is there, it'll probably be fine. Should only be needed if an occluded space is revealed or when first "waking up". Combined with switching one of the forward cameras for one optimized for closer focal lengths, stereoscopic is probably unnecessary.

Perhaps as a customer-specific option they could later include a forward looking LIDAR for precision work, but that should be unnecessary for most customers. A similar option might be to swap out the structural components for lighter weight options (i.e. carbon fiber) if they need more (almost literal) run time due to needing the Bot to move about a significant amount (since it was pointed out up thread that the Bot would use 500W to walk vs 50W for a human, lots of walking would necessitate more frequent charging).

 

[TN Mtn Man]

Iphone has had it for 3 generations now... I would've THOUGHT some android stuff would by now but a quick google says otherwise which is surprising.

Personally I've used it for measuring (~1cm accuracy) and there's even apps to create entire 3D model of a room then use SW to see how it'd look moving stuff around, etc... Apparently you can also use it to create 3D models of objects that you can import to blender and even things like... the Unity engine Tesla uses for their simulator.

Some AR apps use it too... and the camera app uses it by default for better/more accurate focus in low light.

Prob. a ton of other stuff I'm not even aware of.

Range is only 5 meters- pretty worthless on a car... but it'd make precise actions by a bot with hands shorter than 5 meters and moving at (relatively) low speed pretty accurate.

Wow, that's really interesting. Given the SW (I assume Solidworks) interface, it sounds like it would be a useful tool to get dimensions/positions off real-world objects to design off existing mounting/fastener points. I've considered designing some motorcycle accessories-but getting real-world positions and angular relationships without a CMM is a major challenge (and a CMM is expensive)-perhaps using LIDAR scanning would address that. Would need considerably better than 1cm accuracy though.

 

[BioSehnsucht]

Uh, yeah the time it takes to process is the performance. The amount of energy to do such a thing is the watt. This is the only metric that matters for compute in the data center. And it's Tesla that is telling you the performance per watt by using pictures of dojo chips and GPU clusters. Are you saying the H100 will be worst than the A100 or doesn't work at all?

Well, normally Perf/W is the only important metric, simply because even the slowest performing options are still usually not too terribly slow, and with modern cloud and virtualized focused systems, the performance of any given request/job is not going to be measurably impacted by the hardware running it, the hardware just defines how many can be run on a given chunk of hardware and at what cost, so the service provider will optimize for overall (vs single request) Perf/W to optimize their costs.

But when some NN jobs might take a month to compute on existing clusters, then there's probably a willingness to give up Perf/W for faster execution, as long as the resulting power needs aren't too insane.

If they reduced a NN job from 30 days to 1 day and it took the same amount of total power (just delivered at 30X the rate), that would be an even trade and a no-brainer, as long as their local substation doesn't trip in the process.

If it took 10 or 100 times the total power, that might or might not be acceptable, even if it resulted in significant speedup. Somewhere in the middle, if your jobs are taking too long to execute, you're going to be happy to pay more money (both in hardware costs and infrastructure costs, which includes energy and cooling all the heat that energy usage generates) for a "worse" Perf/W just to get faster results, but obviously an infinite cost is not acceptable for any level of performance, so there are limits.

If Tesla is managing to be both faster AND maintain equivalent or better Perf/W, then of course, Dojo is a win on all fronts, with the exception of possibly needing to get their own private substation built...

 

[Electroman]

I'm 47% sure I had a long LinkedIn argument with Trevor way back in the day. I gotta figure out how to search my activity log and frame that.

What a doofus.

Please do

 

[heltok]

Ok I don't think I've ever replied to my own post, but I have to clarify what I said. I dug a little deeper into Boston Dynamics and feel like I've oversimplified their work.

BD's target customer historically was the military. The buzzing dog model was intended to help transport cargo in terrain where wheeled vehicles would struggle, like a dense forest. They've since made newer models that can carry more while being significantly quieter (no more buzzing). They also have mini robots intended to be close-contact recon/spy devices. And while they never push this idea in their YT videos (likely due to sparking Terminator fears), their humanoid robots may have soldier-like applications, where agility and parkour-like abilities would be extremely helpful.

So, BD robots are designed to be useful in narrow military applications, where the primary product advantage is to spare humans from a high-risk environment. They are also pushing the physical capabilities of robots, like faster locomotion, better balance, etc. Some of their robots have some capability to understand their surroundings, like Spot, the yellow dog robot. This robot is used primarily for reconnaissance and can be useful for military and commercial applications. This robot has potential to be useful in a factory setting, perhaps enticing Hyundai to buy BD back in 2020.

So the main takeaway is that BD is also building useful robots, and they've advanced the field of robotics significantly, but their robots' usefulness tends to be narrower in scope than what Optimus aspires to. Tesla is intentionally focusing on the humanoid form, particularly on the usefulness of the hand, since we have one of the most dexterous hands in the animal kingdom. Coupled with a decent brain/AI, there's a lot of potential for usefulness.

I think you can find parallells between Tesla FSD vs Waymo et al and Tesla Optimus vs BD Atlas.

Waymo, Cruise etc use Lidar, HD maps and other tricks to get something working pretty well quickly, but struggle with scaling it and making it a mass market product.
Tesla FSD uses cameras, no HD-maps etc which makes it harder, but a lot cheaper and they get enough data to be able to fully use NN to iterate quicker.

BD Atlas uses Lidar, pneumatics and other tricks to get something working pretty well quickly, but struggle with scaling it and making it a mass market product.
Tesla Optimus uses cameras, electrical motors etc which makes it harder, but a lot cheaper and they get enough data to be able to fully use NN to iterate quicker.

Sure it took Tesla a while to be able to deal with complex intersections, but then a few years later they have 160k cars out in the field sold for actual money to real customers. Meanwhile Waymo/Cruise are around 1k cars on the street, still not making any money. Same for Atlas and Optimus, fast forward a few more years and Tesla will have 160k Optimus out in the world sold for a profit and BD will have 1k Atlas out in the field, still not making any money.

 

[traxila]

Since Moderators are not allowed to go after speeling misteaks, gramaticaliferous badbads and other such horrors, I am going to have to find some way to ban you to the 9th Circle of Hell for reasons other than for having created above a certain non-word.

Just you wait, Eliza Doolittle. Just. You. Wait.

Everyone knows the plural of Octopus is Octopii and so is it for Optimii.

 

[SOULPEDL]

This is exactly what likely went over the heads of many watching the presentation, despite Elon and others repeated mention of the economic impact.

The potential of the bot combined with essentially free energy from renewables opens the door for a world where Universal Basic Income is not just possible, it is imperative. (and wouldn't be dependent upon taxes to pay for it)

Given some thought on how everyone having what they need to get by changes motivations. If a person wants to be creative or productive they can be, not because they have to work to survive, simply because they find it rewarding to do so.

It boggles my mind thinking about this. But, that might not be saying much.

In that utopia, land becomes more sacred. It is finite (on earth anyway).

 

[Singuy]

Well, normally Perf/W is the only important metric, simply because even the slowest performing options are still usually not too terribly slow, and with modern cloud and virtualized focused systems, the performance of any given request/job is not going to be measurably impacted by the hardware running it, the hardware just defines how many can be run on a given chunk of hardware and at what cost, so the service provider will optimize for overall (vs single request) Perf/W to optimize their costs.

But when some NN jobs might take a month to compute on existing clusters, then there's probably a willingness to give up Perf/W for faster execution, as long as the resulting power needs aren't too insane.

If they reduced a NN job from 30 days to 1 day and it took the same amount of total power (just delivered at 30X the rate), that would be an even trade and a no-brainer, as long as their local substation doesn't trip in the process.

If it took 10 or 100 times the total power, that might or might not be acceptable, even if it resulted in significant speedup. Somewhere in the middle, if your jobs are taking too long to execute, you're going to be happy to pay more money (both in hardware costs and infrastructure costs, which includes energy and cooling all the heat that energy usage generates) for a "worse" Perf/W just to get faster results, but obviously an infinite cost is not acceptable for any level of performance, so there are limits.

If Tesla is managing to be both faster AND maintain equivalent or better Perf/W, then of course, Dojo is a win on all fronts, with the exception of possibly needing to get their own private substation built...

I think the biggest advantage of Dojo is complete freedom by Tesla vs using Cuda code stack by Nvidia. Their 2023 Q1 performance expectations are absolutely a moving target and I am expecting 10x by year end considering they managed to squeeze 30% more performance out of the A100 hardware. Lots of low hanging fruit trying to improve v1 Tesla compiler vs well optimized software from Nvidia.

With that being said, there is never any time where you would give up performance/watt. Your system having worst p/w just means some other hardware can also do the task faster at the same power, or do the same task using less power.

 

[Electroman]

120 starlinks donated for hurricane victims. 13 mile radius, 1000 users per terminal

is it just me? after about 2.5 minutes the audio cuts off for the rest of the video.

 

[Cosmacelf]

• Optimus actually already runs on FSD computer, with the same occupancy network, retrained with indoors environment. This validates the basic idea of the program, that Tesla FSD real world AI is transferable to general purpose robotics. This puts Tesla on the leading edge of building an actually working brain for AGI robotic slave labor, mind blowing possibilities.

Let not get ahead of ourselves here. I heard nothing about AGI last night. That's a much bigger step. What I saw last night was a platform that could replace some factory workers. ie. replace workers who do the same thing over and over.

You will note that Tesla didn't address how to tell the robot what to do. No interactions with the robot. All actions were pre-programmed.

Tesla FSD has a relatively simple task of following a pre-set route. The user doesn't have to tell FSD what to do, it is all baked into the training data and algorithmic code. A general purpose robot is a different animal. I'm not saying Tesla can't get there, but they have a big task ahead of them on that front and they haven't started it yet.

• Their latest hardware is optimized for cost and large scale manufacturing. Being useful(while being cost effective) is much more impressive than being able to do back flips. This put them a few steps ahead of any other robotic companies. Remember, prototypes are easy, scale production is hard, no other humanoid general purpose robotic company has the manufacturing know how to do that.

I agree and I think that is the biggest wow moment for me last night. Very few people will appreciate it, of course. They think Boston robotics can pump out battery powered dancing robots at scale and low price and that simply isn't true.

The second robot, the one that was made of all Tesla parts and was cost optimized is the one that people ridicule, but it was the most interesting one.

 

[Cosmacelf]

Been skimming some comments from the AI day. Reddit, twitter etc is toxic AF... So many people with no technical understanding who are so fast to be dismissive, so much negativity and general hate. Sometimes I wonder if these guys are bots or not, but I think mostly they are just sheep being herded by some bot owner with bad intents. It's sad to see. And so many communities that used to be people interested in X are now filled with people who don't even like X. Finance is filled with people who hate people in finance, politics is filled with people who hate politicians, and technology is filled with people who hate technology. Really sad times we live in. Wonder how long this forum will remain readable. The Autopilot/FSD/Optimus subforum here is often very toxic, any discussion goes very quickly into long negative dialogues that makes me wonder if the posters are bots or not...

Amen brother. I suspect you are right in that it is mostly sheep being led by AI bots designed to incite division and rancor. Moderation is the only way out and kudos to the TMC mods for keeping this thread, the most challenging thread on TMC, civil and useful.

Anyway, I guess what I have learned from skimming the web is that most people were very unimpressed by Optimus and are downright negative towards Tesla FSD. People in the field or fanboys are very impressed, maybe a bit too much. There is very little inbetween...

Doesn't matter. Tesla and SpaceX keep creating massively useful products that people keep buying.

 

[Singuy]

Let not get ahead of ourselves here. I heard nothing about AGI last night. That's a much bigger step. What I saw last night was a platform that could replace some factory workers. ie. replace workers who do the same thing over and over.

You will note that Tesla didn't address how to tell the robot what to do. No interactions with the robot. All actions were pre-programmed.

Tesla FSD has a relatively simple task of following a pre-set route. The user doesn't have to tell FSD what to do, it is all baked into the training data and algorithmic code. A general purpose robot is a different animal. I'm not saying Tesla can't get there, but they have a big task ahead of them on that front and they haven't started it yet.



I agree and I think that is the biggest wow moment for me last night. Very few people will appreciate it, of course. They think Boston robotics can pump out battery powered dancing robots at scale and low price and that simply isn't true.

The second robot, the one that was made of all Tesla parts and was cost optimized is the one that people ridicule, but it was the most interesting one.

Yeah I feel like the a lot of it got lost after we saw the first Bumble-C. People would have paid more attention if Optimus was under a cover and Musk pulls it and shows a much sleeker V1 that is optimus. Maybe show off some arm articulation and that's all. How optimus came out with 4 guys frantically trying to carry him out on a rolling pole and then Musk had to explain how it can't walk was just so clunky and lost a lot of the message. Remember how people were wowed at V2 of the raptor engine because it's like half the complexity? That's all they needed.