Dojo! Details of how Tesla's new supercomputer will make autonomous driving a reality! plus FSD beta
Vložit
- čas přidán 25. 07. 2024
- If you wish to support us in creating fun, informative content, please consider giving at our Patreon site here: / drknowitallknows . Thank you!
Wherein Dr. Know-it-all digs deep to find out as much as possible about Tesla’s new super computer, Dojo. What is it? How fast will it be? What will it be made of? When will it come out? And more importantly, what can it plus the Full Self Driving (FSD) 2 beta do for Telsa owners, and those looking for autonomous robotaxis in the near future? Watch and find out.
If you are looking to purchase a new Tesla and would like 1,000 free Supercharger miles, just click this link and use our referral link: www.tesla.com/referral/john11286. Thank you!
Music by Zenlee. Check out his amazing music on instagram -@zenlee_music
or CZcams - / @zenlee_music
If you enjoy more outdoorsy, adventure type videos--like climbing the Matterhorn or trekking to Mount Everest Base camp--and a few product reviews too, please check out our sister channel, Whole Nuts and Donuts, right here: / @wholenutsanddonuts5741
Please ask questions in the comments below, or at: DrKnowItAllKnows@gmail.com. Thank you!
Links to articles I looked it for research into this topic. Note: some information came from the comments sections, so you can look there too if you're interested.
seekingalpha.com/article/4310...
seekingalpha.com/article/4316...
www.tesmanian.com/blogs/tesma...
www.teslarati.com/tesla-dojo-...
electrek.co/2020/08/17/elon-m...
www.inverse.com/innovation/te...
cleantechnica.com/2020/08/16/...
www.indiatimes.com/auto/alter...
www.thestreet.com/tesla/news/...
motorlinks.net/2020/08/14/tes...
www.carswithcords.net/2020/05...
www.torquenews.com/11826/elon...
www.carandbike.com/news/tesla...
electrek.co/2020/10/27/tesla-...
analyticsindiamag.com/tpu-vs-...
electrek.co/2020/09/21/tesla-...
wholemars.net/2019/11/11/dojo...
www.autotimesnews.com/elon-mu...
10:53 “So now we get to the good stuff...”
Too modest: the NPU and sparse matrices leading up to that were great stuff. Please dive into that. Lots of Tesla channels cover highlights, but you’re distinguishing your channel by treating us as very intelligent. Thank you. Keep it up, please, and we’ll keep watching and upvoting!
Agreed Miguel - he’s the real deal
Thanks! An NPU/TPU/GPU episode is coming soon, and the one coming out tomorrow will be interesting to you too, I think! :)
I only just discovered this. This is an excellent summary. It's now September 2021, and Dojo is a reality.
Keep going with the channel. I watch a lot of youtube videos and have caught 2 of your recent ones and they are among the best I've seen. There is a niche for accessible *and* in-depth thought that you're filling. Excellent.
Alright dude, keep knowing it all we'll wait for some more, kudos for doing this!
Thanks 👍!
Enjoyed and learned a lot about the amazing progress to come for Dojo.
Thanks appreciated.
Now for your FSD 2.0 video 🤔
love it, keep up the superb work!
Wow I love your channel and your videos!!!! Please keep them coming!!!!
Solid content dude! Thanks and keep it up!
Great way of explaining such a dense and interesting topic. I’m so excited to see the capabilities of Dojo come to fruition
Great easy to understand video which was really interesting and well appreciated.
Really good information's! Thank you!
Been looking for this kind of in depth analysis. Thanks!
Exponential capabilities for DOJO. FSD iterations are going to get seriously radical. This is epic.
Been waiting for this video ! :)
I will not be surprised if this channel has 50K followers by the end of 2021. Your contents of Tesla are so informative IMO. Keep it up! You are awesome
I think it might be important to explain how the computer can label for itself. The computer can look through time and see it labeled something as a car after it came around the bush, and look back at the clues the camera could see through the bush, and label those keyhole clues as a car, so the computer will know a small wheel as it travels past a bush, is a car, as an example. So a glimpse of a stop sign through some leaves, can be labeled as a stop sign, even though the original labelling took place later in time as the Tesla passed closer to the sign where it was not as obstructed.
Agree. A video explaining in plain English what Unsupervised Learning is would be awesome even for people who don't like Tesla.
For example, GPT-3 was made using Unsupervised Learning.
So it's pretty clear that specifically, Unsupervised Learning is becoming a crucial technique increasingly.
@@MrSchweppes and Oliver, I will be happy to! It's a fairly simple concept, the the devil, as they say, is in the details. I'll work on one about unsupervised learning and how it functions in more detail soon. Thanks for the suggestion! (and GPT-3 is nuts!! lol)
Do you think they could use accelerometer data to train for potholes? That should provide instant feedback to self-supervise the labeling process.
Totally agree would love to know more about how recently this has been possible, where else it’s being used successfully, how easy is it for others to be able to do this too, is this only possible because of the power of dojo, how powerful a computer do others need to have to be able to do something similar, what checks and balances are in place to ensure this being done correctly, how much faster can the labeling be done because of this now and in the future? Great content 👍🏻
nice I have been wondering about dojo now I know more and I agree is will be very important for FSD
this is amazing stuff dude! love the channel and deep content and ur positive vibes.. good luck with ur channel.. glad to see how u moved from 1k to 12k subscribers :D
Valuable insight. Thank you.
I loved the video! Especialy the fact that you did not dumb down the content (at least not much), but explained the programming stuff nicely so us layman people can understand and follow. First video from this channel I saw - > subscribed 💪
Top notch content and research.
The most significant advancement is the unsupervised learning!!!
Thank you for distilling all the info. Please keep it up!
I love this! Please do that video on TPUs & NPUs! Can't wait to watch your channel blow up even more!
Super interesting video, thank you :)
This is my first time watching any of your videos and I absolutely love your level of detail and enthusiasm in your presentation!
Totally want to see differences between TPU and NPU! I think it would be really interesting. I’d also like to know more about unsupervised learning, how it translates to exponential learning. (Like how fast are we talking here?) What do you think about Hardware 4 that they’re working on and it’s implication in things? What do you think the advantages are over the current HW3 chips in some of our cars currently? Dude you rock #1 subscriber secured.
Thanks so much! I'm definitely doing a TPU/GPU/NPU video soon. I'll have to look into HW4, but that one would be cool too.
@@DrKnowitallKnows - At "Autonomy Day", last year, in April, I recall them stating, that while they were just then releasing the AP Version 3.0 Chip, and the Version 4.0 Chip was already progressing well in Design! But hadn't gone to FAB yet, but, that could be happening in about 18 Months - was my take-away of it!
So, baring Covid-19 Delays, we are now at that "18 Month" window, wherein, it could already be going to FAB, for First Samples, which would Lead "Volume Production" by 6-12 Months, as I grasped the info!
In Other Words, from now, to about Cybertruck Production, HW 4.0 could be coming onboard cars / trucks, as they ship! Possibly just Cybertruck, at first, as it starts to ramp up, and Chips Ramp Up! Then Other Tesla Products (cars)!
exponential.. edge case.. machine that builds the machine.. funding secured 420
Mr. Good content you put out there 👍
Really well done video and I agree the name is perfect. Would love to see other videos comparing NPU/TPU/etc.
Will do!
Love your videos doctor! Thoroughly enjoyed this video.
Good job I have subbed and look forward to seeing your channel grow. Nice content.
Excellent understanding of the concepts. Subscribed!
Excellent content. Thank you. BTW Iam a descendent of JS Bach so I was very happy you made mention of him. Keep up the good work.
Cool video! Thank you!
Thanks man.. learned a lot
Great video!! More on TPU and NPU would be really interesting!
“I concur!”
Agree, I'll be happy to see it!
yes, your explanation was very good already but a little more never hurt anybody!
I got
@@mcRydes llo
Very in depth, accurate and organized informations. Great video
That was one of the most amazing information I have heard. And the explanations, so good! thanks for transmitting all this knowledge and research.
Good good thank you for finally explaining this to people in a way people can understand, this misunderstanding has gone on far too long!
Another great video!!!
Thanks for clarify this.
I wondered whether individual cars were capable of learning/improving upon whatever neural network was downloaded into them.
New Subscriber - well done video THANK YOU
This is good content, not only rumors... Kudos... 👍
Great video mate! Keep it up
Again a very informative video. Thanks a lot. Please more npu detail stuff
At 2:14 AM, Dojo became self aware and immediately proceded to beat Donkey Kong high score and Super Mario speedrun time.
It would become aware at 4:20 pm lol
@@petermyles8224 🤣🤣🤣 to you both! Thanks for the laugh
Hollywood is already working on the script for the Terminator 69 movie,
in which Dojo and Skynet fight each other for final domination of Earth.
@@danchatka8613 I'd watch that
I think it'll upload itself through over the air update to every Tesla, then beat us on every chess game just to troll us. :)
Great content, keep it up.
Good work crunching the information down to fit in a decent length video.
"What do you guys do in your Dojo?"
"We EXArcise!"
Excellent run down on Dojo. I "used" to be a lot smarter on new tech, but now I rely mostly on other folks to distill it for me so I can enjoy retirement. Thanks for the great work.
Here my 2¢ : rename the channel to
Dr. know ( drop the - it all )
Your content deserves a bigger audience.
just found your channel, great work! you are going to do well!
Stick with it, you will have 100k plus subscriber in a year.
Tesla is a good way to build your channel
Thanks!
I would LOVE a video on the differences between TPUs and NPUs!
Great video, but you are overlooking one of the crucially important "detail"! The "video processing" or 4D training as Elon Musk refers to is not only faster, requires less/no human labeling, but it's compared to "2.5D" training used by Tesla in the past, conceptually better approach! A typical way of approaching full self driving of cars is to use sensors (cameras, radars, lidars, ...) to figure out, where the car is at the moment and what's around it. With lidars it's easy, with only cameras it's harder, but still relatively easy compared to what follows.
The most difficult and for level 4/5 autonomy required is at least decent level of understanding of the situation around the car and FORESIGHT of what probably happens. The system simply has to be able to infer from what happened in the moment ago and currently happening what's the best/safest action for the car to do. Simply what experienced human drivers do without actually thinking about it.
Neural network in the human brain learns that, because we are capable of perceiving the "time continuum" by remembering the past, observing the presence and constantly trying to predict the future. This is not possible with "2.5D" training Tesla did and (to my knowledge) all its competitors still do (if using neural networks at all).
For example if you have just a picture of a car close to the middle of the road it's hard to predict with certainty if it's going from it's lane to your or vice versa or just went accidentally to close to the middle of the road. You need to program this to your system "by hand" and the car needs to be more cautious than a human driver would be.
If you however give the neural network a sequence of frames in time (a video) what happened before and after, it will learn this pattern and it become suddenly capable of deciding based on the whole action of the other car perceived in time, not just by a single frame / frozen moment in time.
That is HUGE!
Now Tesla can teach the neural network itself how to react in particular situation and not hand-code all the myriads of possible situations and reactions to them by army of developers that inevitably miss large portion of what's possible in the real world. The inherent generalization capabilities of the neural networks come very handy here - the car will be able to react not only to situations any of Tesla cars already encountered, but also to similar situations that never happened to a Tesla vehicle before...;-)
Thank you. Yes, I should've been much clearer about this for sure! The research I did for the video, and the amount of cuts I had to make to get it to a reasonable time meant I overlooked some aspects and this is the biggest. I think Tesla's referring to this as "trajectory projection," which basically means FSD is able to track an object, then (based on what it is--a box, a car, a dog, a person) create a reasonable projection of what is likely to happen. That is, as you say, HUGE. What do you think? An entire episode on this? I think it might help a lot of folks understand just how big a deal "4D" is...
@@DrKnowitallKnows Yes please and thank you. A dedicated episode will illuminate the intricacies behind 4D and what a massive change it is going forward.
@@DrKnowitallKnows an entire episode? Yes please! Love the content and your style. Subscribed
Great observation...you deserve a CZcams channel of your own!😂
You said "only with cameras" - no, Tesla also uses one forward-looking radar (which can "see" through fog, rain, etc. while lidar is based on lasers that can be blocked more easily). In the video, he also misses this point.
Thanks once again. Cleared this up for me, some anyway...)
I'm enjoying your educated and quite knowledgeable videos. So much on YT is so dumbed down that it's not very informative. You make mistakes but they're usually corrected, either by comments or by yourself, so please keep doing that. It's interesting to me that things I learned about sparse matrix calculations over 30 yeas ago are still being applied today.
After watching this video, I could only claim to have found one (very, very nit-picky) error: "the length of each byte is 32 bits". No, bytes are always 8 bits. A 'byte' is a unit of measure. Words can vary in bit size and I think that's what you meant.
As someone who has been following Dojo news (usually extracted from Musk's tweets or interviews) Dojo is indeed online and is now providing the "4D" updates. You may be right that it's not performing as it's expected to as time goes on, but it's more a matter of taking longer to calculate a new model than it is missing "parts". I believe - again from what Musk has 'said' - they were waiting for Dojo before launching beta FSD. You can see the 'proof' of this in the rapid turn-around times we're now seeing.
My favorite example of how quickly and yet thoroughly bugs are being fixed (or should I say "issues are being resolved" since Dojo is operating on a whole new level?) is the round-about. Immediately after receiving the first FSD build, one of the beta testers drove through a normal round-about. Despite it being night, FSD did a competent job. But the next day another beta tester drove into downtown Auburn, California. There, painted onto the street at a wide intersection was a circle with writing inside. It is a huge sign that says something like "Welcome to Auburn". The first build decided that it was a round-about and the driver had to take over. But the very next version didn't and treated the intersection correctly. The speed of updates is beyond what we as consumers have ever seen, and it's just going to get faster.
Thanks for the great information. I'd love it if you went deeper in future videos. The NPU stuff is fascinating!
Great video!
Great video, love it! I'd love to know what other companies are using for their FSD?
Wow, good question! I believe most are using Nvidia, as they are far and away the leaders. Waymo is likely using Google TPUs, however. Does anyone happen to know? I could work up a video in this if there's enough interest.
That was Great!
Awesome editing!
Good video! I just wanted to add that Tesla's last update on this mentioned Dojo would use the SoC in all new Tesla vehicles. Also, those processors are made right in Austin, Texas (Samsung chipmaker branch).
Oh nice: no shipping distance at all once Giga Austin gets rolling!
Thank you for your hard work. As you said cars are not figuring stuff out on their own. Following instructions and collecting data for the mother ship is what they do. I look forward to watching the speed of improvements as Dojo comes online. My son works for Tesla and he keeps telling me there is so much more going on behind the scenes then we will ever know. Except maybe Dr. know it all. LOL
hahah. I wish I knew even more, honestly. That's fantastic your son works for the big red T. Tell him to keep changing the world for the better!
@@DrKnowitallKnows It was the mission to make a better world which caused my son to take a pay cut and travel across the country to work for Tesla. He is now very well off because of the stock incentives and stock bonuses he received for his hard work. He has been working on the road runner project. He tells me it will be world changing.
I really wish you could tell us more about what's happening! :)
@@davidalton7378 Congratulations to you, Sir. You must be a proud dad!
@@davidalton7378 - Thanks for the Inspiring Tidbits, as much as you can "Let Out!"
Yes you definitely need to do video on npu.
🙋🏻♂️ more of this content, please.
great video and mad props for maya. I still have a fully sealed maya v1.0 in box in my display case
Hahahah! Wow.... I wonder what it would look like to run that bad boy now--or if you even could?!
@@DrKnowitallKnows I have a copy on an OLD octane box so might have to fire it up!
Thanks! So now I understand why Tesla did not chose to have the 4D approach from the beginning on.
Thank you very much sir
Imo too, if Starlink bandwidth is connected and accesible, directly to each vehicle, this can reduce the task load/timframe on labeling, for whatever system they have, current or the upcoming dojo...
I agree, I think the Dojo upgrade (or some of it), happened with last weeks FSD beta release, as there has already been a beta upgrade 3 days after the initial release!
Thank you for breaking this down.
Looks like a warm day there :-)
A processor session would be awesome!
Great video.
Excellent discussion ( my Model 3 with FDC is on order) and I am trying to be prepared for the experience when it is released.
While you were describing the time reference I was thinking of what a football quarterback has to process as he throws the ball while running backward, dodging tacklers, evaluating the angle and speed of his receiver, determining the elevation needed and the force necessary to ........ it goes on and on! This is what Tesla must achieve to attain true level 5. I’m eagerly awaiting at least the entry into level 5. 🤔😃
Great video! Definitley do a NPU and TPU special, please!
Great video! Also, what's that font you're using for your overlays?
I am super interested in knowing what is an NPU and a TPU in more detail. If you could make videos about it I'd watch them 100%
Great channel. Wondering if Tesla would be able to use all cars with HW 2.5 and 3 in a distributed compute farm. Like folding @ home or wcg. Think of all the power sitting in garages all night. Owners could volunteer to help out. The compute adds up quick.
First of all thanks ✌️ CZcams proposed some videos of you inside my bubble 😁. During my studies I had courses about cpu implementations, all those pipeline stuff RISC/CISC whatever...I learned most important part of cpus are the instruction sets. All those SSE, AVX stuff. You talked about difference about gpu/cpu/tpu/npu - would love to see a video about instruction sets of those units. From top of my head I thing the FSD chip also has only some instructions to execute neural nets.
Thank you
When did tesla say dojo is online? Was that your hunch? Anyways, this is an interesting topic. This channel is fast becoming my favourite. 👍
Thank you! And as for timing, at Investor Day (9/22) Elon said mid next year. But at the same time, in August, he said there would be public access to Dojo in 6-8 weeks. It's a bit tough to square those two, which is why I'm thinking they're bringing it online a piece at a time. But that is totally an educated guess on my part!
very good video
Thanks so much for all the great information! This is exactly what I’ve been looking for… I have so many questions… I think I understand the machine learning part of it and the difference between what’s happening in the car and what’s happening at the mothership. You did however mention that although the car can’t “learn“, it can “remember“.
So, if there is a particular self driving challenge I encounter on a daily basis, when I disengage auto pilot to handle that situation, I understand the car itself doesn’t register that except to send the data to the mothership. Does the mothership then incorporate the data about that specific issue at that specific location and share it with the whole fleet, or does it apply that generally to situations like that unrelated to location?
In other words are routes that are more frequently traveled by teslas going to be more error free then places no teslas have driven before?
And does the fact that my car will “remember“ mean that I would expect full self driving to improve relatively quickly on a route that I take frequently, where some other Tesla going over that same route might not do as well?
It seems like it’s a combination of general ability to handle situations it’s never seen before along with location-based data that helps it in situations that have already been encountered at specific locations.. Is that correct?
Thanks so much again for taking the time to enlighten us
Good questions! So, first, I don't work for Tesla, so I don't know absolutely. What it appears to be to me: let's say on your daily drive, there's a weird road intersection where you have to make a soft left onto a road. FSD fails; you take over. FSD shoots that problem back to the mothership, and it's put into the bin for extra training on the next go-around. If you (and 500 other people around you) do the same drive every day, and have to take over each time, it'll get flagged 501 times/day, which pushes it up the list of edge cases. If there are 1,000 other intersections similar to this, and 500 people per day have to take over, then 500,000 cases per day are pushed to Tesla, making this a BIG DEAL issue that gets high up in the queue. Next FSD update, you notice the problem is gone: that's not because of your intersection, but because of the 1,000 intersections that are very similar to yours. So this is general. The memorization _might_ be (again, guessing here) that FSD notes your path through that intersection each day, and begins to memorize how you drive it. Then it starts to get more confident in its prediction since what you do and what it does (in shadow mode) start to align. That would be the local "learning" that it would be doing.
Great video, it probably took a lot of research for you to do it. 👍
It was a lot of work--so thanks for noticing! But what fun to get ti dig into the future. Dojo is perhaps Tesla's greatest achievement yet (hardware and software)--assuming they can pull it off.
interested in a video explaining the differences between TPU and GPU and how that help for training
Hey there! Would you mind doing a video explaining NPU’s please? 🙏
Definitely interested in a separate video on NPUs
Yes please. More on NPUs
Is the DoJo supercomputer functional, or are the current Beta updates by the current learning computer? I missed that issue. Thanks
Love your videos... FLOPs (floating point operations per second) are a general purpose speed benchmark ... but the real key to speed in an AI application like Driving is fast integer based matrix math
Love it🤩🤩🤩
Yes please; NPU vs TPU etc
Definitely do a deep dive into NPU's please.
+1 on TCU and NTU
yes please make a comparison vid for npu, gpu, tpu, cpu
I would love to see a video on NPU's!
Nice video. FSD is a critical enabler for Tesla to achieve its mission faster. It helps accelerate the adoption of EVs near-term as a feature people want, mid-term to enable robotaxis and long-term to accelerate retiring the global ICE fleet with each EV eliminating 10 or more ICE vehicles if used full time as a robotaxi. I would keep digging deeper into FSD, the technology (sensors, chips, dojo & AI) and roadmap. Just 20M EVs produced a year by 2030 could replace 200M ICE vehicles a year with FSD + robotaxis or the entire ICE fleet by 2035 by Tesla alone. With help from other EV OEMs then it can happen even faster.
Please do an episode on NPU, thank you
Do you think that eventually teslas will be able to communicate with other teslas in the general vicinity for a heads up early warning of accidents and such?
your 100% correct dojo will tackle the extreme edge cases required for driverless cars
Chasing those 9s is the way it'll get done!