Slashdot is powered by your submissions, so send in your scoop

 



Forgot your password?
typodupeerror
×
Supercomputing

Tesla Starts Production of Dojo Supercomputer To Train Driverless Cars (theverge.com) 45

An anonymous reader quotes a report from The Verge: Tesla says it has started production of its Dojo supercomputer to train its fleet of autonomous vehicles. In its second quarter earnings report for 2023, the company outlined "four main technology pillars" needed to "solve vehicle autonomy at scale: extremely large real-world dataset, neural net training, vehicle hardware and vehicle software." "We are developing each of these pillars in-house," the company said in its report. "This month, we are taking a step towards faster and cheaper neural net training with the start of production of our Dojo training computer."

The automaker already has a large Nvidia GPU-based supercomputer that is one of the most powerful in the world, but the new Dojo custom-built computer is using chips designed by Tesla. In 2019, Tesla CEO Elon Musk gave this "super powerful training computer" a name: Dojo. Previously, Musk has claimed that Dojo will be capable of an exaflop, or 1 quintillion (1018) floating-point operations per second. That is an incredible amount of power. "To match what a one exaFLOP computer system can do in just one second, you'd have to perform one calculation every second for 31,688,765,000 years," Network World wrote.

This discussion has been archived. No new comments can be posted.

Tesla Starts Production of Dojo Supercomputer To Train Driverless Cars

Comments Filter:
  • During the gold rush, you don't get rich digging for gold. You get rich selling shovels to miners.

    A lot of larger miner conglomerates, aware of this, set up production of their own shovels expecting rush to last a long time. Makes sense.

    • by AmiMoJo ( 196126 ) on Thursday July 20, 2023 @08:34AM (#63701524) Homepage Journal

      Musk announced it way back 2019, with an operational date of 2020. They claim it's been used for development since at least 2021, but is only now entering production use.

      Musk tends to vastly over-estimate what AI can do for him. Not just self driving cars, which he claimed were a year away back in 2016, but also things like his robot and brain interface.

      In any case, it would be interesting to know exactly what they are doing with Dojo. Is it just machine vision learning, as in they are trying to replace the lidar and radar sensors that everyone else is using with pure vision, and that means having a system that can recognize pretty much anything a human can in all lighting and weather conditions. Tesla already has issues with their cars rear-ending others since they removed the front radar, as they simply can't recognize certain types of vehicle.

      Everyone else greatly simplified this task with lidar and radar.

      Then you get to the decision making for the actual driving control inputs, which everyone else does with algorithms. Are they planning to have an AI driver that is trained on the rules of the road?

      • It is possible to drive cars well enough using only cameras, as evidenced by the radar- and LIDAR-less wetware systems already in place. Of course it is harder, but trying to do that without lidar and radar will result in cheaper cars in the end.
        • Re: (Score:1, Troll)

          by drinkypoo ( 153816 )

          It is possible to drive cars well enough using only cameras, as evidenced by the radar- and LIDAR-less wetware systems already in place.

          Humans do a pretty crap job for the most part, so that's actually counterevidence to your point.

          Of course it is harder, but trying to do that without lidar and radar will result in cheaper cars in the end.

          Will it? LIDAR is getting cheaper pretty quickly. RADAR is already quite affordable. The cameras used on Teslas still don't offer the same resolution that our eyes can manage, so not only do the cameras need to improve, but more processing power needs to be added to process the higher resolution data. A nonmoving human eye can see about three times the detail that the latest cameras used on Teslas can pick up. A

          • Hey, as long as the "smart" driverless cars are smart enough to get the fuck out of MY way when I'm driving, I'm cool.

            They will stick to the limit, I never do...so, yeah as long as they get out of the way, it's a good thing.

            Better than having some human granny in the passing lane holding everyone up.

            • by Luckyo ( 1726890 )

              Man, there's a reason why Cayenne drivers are considered pariahs among the Porsche drivers. Now we know why.

              • Man, there's a reason why Cayenne drivers are considered pariahs among the Porsche drivers. Now we know why.

                Ugh...not me, I'll never own a SUV.

                All of my cars to date have been 2-seater sports cars...with the sole exception being my '86 911 Turbo (black, whale tail, the whole package)....which technically had 4 seats.

                But since you couldn't fit more than a bag of groceries back there I don't really count that as a true 4-seater car.

                I like fun cars, and yes, I drive them fast....it's much nicer if the slow

                • by Luckyo ( 1726890 )

                  In my language, there's a rule of three T's. "Tee tietä typerykselle".

                  In English, it roughly translates to "make for for the dumbass".

                  The other version translates to "make way, he's in a hurry to his funeral".

            • by haruchai ( 17472 )

              "They will stick to the limit, I never do...so, yeah as long as they get out of the way, it's a good thing."

              Enjoy it while you can. The timelines from every automaker, especially Tesla, are too optimistic, imo but it's something that WILL be solved...at some point.
              When that's happens, that cars are good enough, YOU....WE...won't be considered good enough & insurance companies will make sure that no human can afford to have a driver's license for public roads where autonomous vehicles operate.

              • When that's happens, that cars are good enough, YOU....WE...won't be considered good enough & insurance companies will make sure that no human can afford to have a driver's license for public roads where autonomous vehicles operate.

                I truly hope this happens well after I leave this earth.

                I have had fun driving...I've owned cars that are fun to drive.

                I like the autonomy....and when cars become self driving, I believe you will lose the autonomy.

                At some point, the govt will control access and BANG...you

            • I sincerely hope they will stick to the limit and lower the speed to keep it safe if you come too close.
        • Re: (Score:2, Troll)

          Only high resolution mmwave imaging radar will be able to deal with extremely bad weather.

          Of course level3+ self driving is AI hard and level 2 is a menace. So the vision tech is moot.

        • by youn ( 1516637 )

          the cost differential is not that much... which is why all other manufacturers have them in, even in cheaper cars.

          Just Trying to be safe when you have a more secure option available today is not an ideal goal

          Sure when they solve the whole problem without additional sensors, they are free not to use that sensor, meanwhile, they could make their cars safer. Especially when teslas are not known to be the cheapest around

          • The Musk says you don't need them, so his cult members^H^H^H^H^H fans will die on that hill and so will he.

        • by AmiMoJo ( 196126 )

          I think Musk expected to have self driving in a few years, but because it has been so long since they started we have reached the point where radar and lidar have got affordable. Radar is common now, and Volvo is fitting lidar to cars.

          Vision only might be cheaper, it depends if it can work with cheap cameras. And if it can work at all - it's not clear at all if AI can reach human levels of understanding of the world though vision in the next few decades.

        • It is possible to drive cars well enough using only cameras, as evidenced by the radar- and LIDAR-less wetware systems already in place.

          In that case has Tesla done the necessary (and obvious) research of ensuring that a human can drive well enough using only the visual inputs they supply?

          Because those wetware cameras also have a lot of extra features, like being able to move independently to improve depth perception on specific objects, I'm guessing the Tesla cameras don't do that.

          Besides, just because something is doable for a human brain doesn't mean we're close to replicating it with AI. I suspect the progress of LLMs took a lot of resea

        • Your statement is pure conjecture, with no real-world evidence to back it up. Even the term "well enough" is completely subjective. No current "self driving" system is fully autonomous, be it run by LIDAR, camera vision systems, RADAR or magical fairy dust. Zero. None. Nada. And there is no evidence that they ever will be.

          End game is probably cheap labor in developing nations driving your car with a VR headset and emergency braking systems run locally.

      • With the new added super computer power, finally they will solve the full self driving cars within months

        same deadline as every year for a few years now :p

      • You don't understand what Dojo does. Its a full compute platform for processing the massive input of video files from the fleet and crunching it down for use by the Neural Nets. Instead of using racks of NVIDIA GPUs they can use this platform for increased scale, decreased cost and more work per watt. It's got nothing to do with lidar and radar - that's in the car.
      • Everyone else greatly simplified this task with lidar and radar.

        You can get off-the-shelf chipsets from Intel that can do this with vision only cameras. Available for years now.

        This "self driving cars must have LIDAR" cult will never die.

        • by AmiMoJo ( 196126 )

          And yet no actually self driving cars use these magic chips, they all have lidar.

          • That was my point. The LIDAR === "Self Driving Cars" cult will never die, at least in the foreseeable future. All the "actually self driving cars" that rely on it is evidence of that. The project teams that developed those yet-to-be-proven products bought into a hyped technology the same way so many business plans were based on "block chain" technology.

            Musk knows about LIDAR. He developed a system for SpaceX that uses it for docking. You can't argue that he is ideologically driven to reject LIDAR.

      • by steveha ( 103154 )

        Tesla already has issues with their cars rear-ending others since they removed the front radar, as they simply can't recognize certain types of vehicle.

        I'm gonna need you to provide references.

        There have been documented cases of Tesla cars, driving themselves, hitting things and killing someone. To my knowledge, most or all of these cases involved very old versions of Autopilot, possibly even "Hardware 1" (which only had a single front-facing camera). And all the ones I know about involved the car hitting

  • really need to test with real cars on an real test track.
    not in VR but needs to be done on an test track with real hardware and the cars local CPU's only.

    • They've been testing with real cars in the real world on real streets, and the results are that Teslas make embarrassing mistakes that no other automaker's vehicles make. This is in turn because Musk decided that Teslas didn't need RADAR (let alone LIDAR) and they would do the whole job with machine vision. This means that even a boring "normal" car with AEB is better at not crashing into things than an allegedly full self driving Tesla.

      This is the same mentality as Oceangate, except instead of just killing a few billionaires, this is affecting real people out in the real world.

      • by Junta ( 36770 )

        Don't forget how they overpromote safety by disingenious claims about accidents per mile driven.

        Where they compare all miles driven by a human in any car to Tesla autopilot miles and declare Tesla to be safer, despite autopilot only bothering to activate in the most favorable conditions, and if autopilot disengages prior to an accident, well that's not an autopilot incident...

      • They've been testing with real cars in the real world on real streets, and the results are that Teslas make embarrassing mistakes that no other automaker's vehicles make.

        Are you referring to the Cruise taxis and Waymo units that routinely go nonlinear and just lock up in traffic until an operator comes to rescue them? They do that even though they are on software rails, one step up from a carnival ride. Citizens have taken to sabotaging them in protest.

        Have you seen those cars? They look like they have cancer, with LIDARs on every corner. MOAR LIDARS doesn't save them.

        • by steveha ( 103154 )

          I agree with nearly everything you said. I would have tried to word things a bit more politely, but you aren't wrong.

          In fairness though there is one point to discuss:

          Citizens have taken to sabotaging them in protest.

          I saw a TikTok video explaining how to sabotage them (step zero, steal a traffic cone; step one, put the cone on the hood of the Waymo car; hey presto, the car thinks it can't drive at all). That video gave a reason for doing it, and that reason was that self-driving cars are bad because every

    • Test Track (Score:5, Informative)

      by JBMcB ( 73720 ) on Thursday July 20, 2023 @08:58AM (#63701582)

      Fun fact: most car companies have an autonomous car division in Pittsburg for two reasons:

      1. Carnagie-Mellon university has an outstanding robotics program. Good pool of grads to hire from.
      2. Pittsburgh is an absolute nightmare for driving. Tunnels with no GPS signal leading to bridges where every lane becomes a different exit. Multiple streets stacked on top of each other. Flat areas, hilly areas, downtown areas, suburbs, one way streets, surface highways, elevated highways, highways with exits and entrances on both sides, weird left turns. It's a nightmare, and a perfect place to test autonomous cars under worst case conditions.

  • FSD (Score:4, Funny)

    by muh_freeze_peach ( 9622152 ) on Thursday July 20, 2023 @08:56AM (#63701574)
    FSD by 2017!
  • That means Tesla is going to stop using customers for this job?

  • by PPH ( 736903 ) on Thursday July 20, 2023 @09:21AM (#63701666)

    So is the goal to earn a black belt in dealing with pedestrians?

  • by groobly ( 6155920 ) on Thursday July 20, 2023 @11:05AM (#63701932)

    We are going to find out what are the limits of deep learning. The question is, how much can we train a system to do without really knowing what it is doing? The fundamental problem of AI is that everything a human does can at any point involve anything at all that the human knows.

After a number of decimal places, nobody gives a damn.

Working...