Google and Tesla Will Compete With 10-100 Exaflop AI Clouds

In early 2022, Google revealed it is operating a full cluster of eight of these AI supercomputing pods in its Oklahoma data center. There is combined 9 exaflops of peak aggregate performance. Google believes this makes this “the world’s largest publicly available ML hub in terms of cumulative computing power, while operating at 90% carbon-free energy.”

Google plans to buidl a dozen AI data centers.

Tesla Announced AI Dojo Training at the Scale of Google AI Cloud

Each Tesla Dojo AI training tray consists of six training tiles. Tesla said each 135kg tray offers 54 petaflops (BF16/CFP8) and requires 100kW+ of power.

Each cabinet holds two trays and accompanying interface equipment. In Q1 2023, 10 cabinets will be connected into one ‘Exapod’ that will be the 1.1 exaflops (BF16/CFP8) Dojo system.

Tesla plans to install a total of seven ExaPods in Palo Alto, potentially offering 8.8 exaflops (BF16/CFP8). This will be competitive with the Google Oklahoma data center.

Musk said Tesla may offer Dojo to companies as a AI Cloud service.

Tesla is building next generation Dojo chips and tiles that will provide another 4X in performance.

8 thoughts on “Google and Tesla Will Compete With 10-100 Exaflop AI Clouds”

  1. Curious how the advanced cloud companies are getting weirder and weirder and ever more aloof and distant, factorizing out the control of computing from their customer’s side into their opaque and remote sancta sanctorum.

    With approaches and technology fewer and fewer can understand and change, only a priesthood of researchers from exclusive schools, despite their desire to make it approachable and have more possible candidates for their postings.

    If this continues, people will have cars, robot maids and workers that think and act thanks to being tethered to the Tesla or Google cloud, creating a potentially catastrophic civilization-wide bottleneck, depending on permanent communication with the AI mother lodes.

    Yes, we are already kinda there, but it will be up to the point where cars will crash, drones will fall from the sky, robots will cease working and create accidents and people will die soon if the tether is broken.

    • That is definitely not impossible, but I think you are fundamentally misinterpreting how the “cars and robot maids” will probably work: running their own neural networks offline, and only depending on the link to the great cloud for software updates, much like our own smartphones already depend on either the Apple or the Play Store, whichever the case may be.

    • There’s some of that going on, sure, but Telsa is pretty clearly devoted to the notion that the end products should be independently functional; Navigating based on vision systems, for example, rather than trying to get transponders put into traffic lights.

      The supercomputers here aren’t directing the cars where to go, they’re taking data sent from the cars when users take back control, to teach the neural networks to handle more edge conditions. Then the learning is downloaded to the cars’ own neural nets, to function independently.

      The bots are expected to operate in the same way.

      So the central server goes down, your Telsa bot will still be able to water the lawn, but if you plant a new shrub won’t be able to download it’s specific care requirements. And won’t ‘wake up’ one day knowing what to do if the neighbor’s kid draws a hopscotch pattern on your driveway.

      • Yeah, I get your point.

        No SDC manufacturer can allow their cars crashing because they have bad 5G connection. The same as no domestic bot company can allow the bot to stop working due to the same reason.

        Or would they? the IoT is full of grievous examples of things that should work without being tethered to a remote server, that on the contrary, cease working the very moment their mother company goes under.

        SDCs seem so far to have the full computing capability to do their own job without outsourcing anything urgent to he cloud, except updates and sending the vehicle experiences to be amassed.

        I presume Tesla bots will be the same: having enough computing on themselves to not depend so strongly on the cloud.

        • They’re literally, at this point anyway, using the same self-driving computer for the cars and the bot.

          I think the bots are a little safer for them to have lose functionality if a network connection is lost, because they’re not as likely to be in the middle of something where shutting down is life threatening. You’d just need them to maintain some kind of local safe shutdown sequence locally. Shut off the stove and sit in the corner in the event of a lost network connection while cooking, for instance. They’re certainly not going to try to run them all remotely from a server, way too much processing and communications to do that.

          Things like home automation that bricks if a local server is shut down is deliberate, to turn a purchase into a rental, but that seems not to be a path Telsa means to go down.

          • To be clear, while I think dependence on a remote server is a little safer for Teslabots than for self driving cars, I really do not think Musk means to build in that kind of dependence. It would really limit market penetration, for one thing. And makes no sense at all for internal use.

            I expect that they’ll maintain commonality between the car and robot self-driving computers, too, treating the cars as just special purpose robots. It’s consistent with their philosophy of reusing engineering.

Comments are closed.