Tesla’s relationship with artificial intelligence (AI) has been rocky at best. But even though CEO Elon Musk voiced strong concerns about the technology earlier this year, the company hasn’t stopped developing internal uses for it. One of which is a massive supercomputer, dubbed “Dojo,” which the carmaker plans to use to train its self-driving algorithms.
The supercomputer went into production last month and will serve as one of Telsa’s “four main technology pillars” to support its self-driving efforts. Harnessing a massive amount of computing power, the system will drastically accelerate the carmaker's push toward autonomous vehicles.
Nearly five years ago, news broke that Tesla was working on an in-house chip for self-driving cars. In 2019, the company debuted the chip and began shipping it in the Telsa Model S, Model 3, and Model X.
Now, the six-billion transistor chip is featured in the company’s new Dojo supercomputer. At its launch, Telsa claimed the D1 silicon delivered 21 times faster performance than the Nvidia chips it previously used. Since the chip was designed for self-driving workflows and analysis, it outperforms generic silicon at these tasks. Featuring dual neural network arrays, the chip can handle 72 trillion operations per second.
Dojo is poised to be a powerhouse. Musk claims it will handle an exaflop (one quintillion) of floating-point operations per second. Tesla plans to accomplish this with a scalable approach by combining six tiles of its in-house D1 chip to form a system tray. Then two trays are stacked together in a computer cabinet capable of managing 100 petaflops. Putting 10 of these cabinets together will let Dojo break the exaflop barrier, forming an ExaPOD.
According to a company tweet, the supercomputer is expected to be “one of the world’s five most advanced supercomputers by early 2024.”
With its upgraded capabilities, Dojo will be able to process more real-world video data from Telsa’s feeds. It will also process data from Telsa’s simulation models. This will accelerate the development and improvement of its self-driving algorithms for the Tesla Autopilot and full self-driving (FSD) systems. Given the vast amount of processing power the supercomputer will provide, Dojo could also provide bandwidth for Tesla’s non-vehicle projects, including its humanoid robot.
In a statement, the company said, “The better the neural net training capacity, the greater the opportunity for our Autopilot team to iterate on new solutions.”
Tesla is clearly much more than an electric car company. Under the guidance of Musk, it has branched out into several areas of technology, with a focus on machine learning and self-driving capabilities leading the way. Supporting those efforts has meant developing in-house technology and decreasing reliance on solutions from third-party providers.
In its second-quarter earnings report, Tesla outlined four key areas it plans to focus on while attempting to “solve vehicle autonomy at scale.” This includes, “extremely large real-world dataset, neural net training, vehicle hardware, and vehicle software.”
“We are developing each of these pillars in-house. 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 report says.
While it’s very exciting to see Dojo enter production, Tesla has some lofty goals for the machine. It’s already pursuing an aggressive timeline with its prediction of reaching exaflop processing power so soon. Whether or not the carmaker can eclipse nearly all other supercomputers in operation today in a matter of months remains to be seen.
Managing its power consumption will be one essential component. The supercomputer reportedly overloaded the local power grid during a limited test run. So, shoring up this insufficiency before the final version goes online is crucial.
Even a slower timeline, though, will surely yield positive results for Tesla’s self-driving ambitions. With far more computing power under its control, the company is poised to make big leaps with its autonomous features in the coming years.