Tesla has formally ended its Dojo supercomputer challenge, closing out a four-year effort to develop one of many world’s strongest A.I. coaching methods and marking a significant shift in Tesla’s A.I. ambitions. CEO Elon Musk introduced the shutdown in a sequence of posts on X over the weekend. He additionally confirmed that your complete Dojo staff of about 20 staff has been disbanded.
Dojo is powered by Tesla’s D1 chip, designed in-house (and manufactured by TSMC) to deal with large volumes of driving knowledge for coaching the corporate’s Full Self-Driving (FSD) system. With Dojo shelved, Tesla is pivoting to a streamlined chip technique, specializing in next-generation chips like AI5 (manufactured by TSMC) and AI6 (manufactured by Samsung) for each coaching and deployment. AI5 chip is constructed to energy self-driving and robotics capabilities, whereas AI6 can moreover deal with large-scale A.I. coaching duties.
“As soon as it turned clear that every one paths converged to AI6, I needed to shut down Dojo and make some robust personnel decisions, as Dojo 2 was now an evolutionary lifeless finish,” Musk defined on X.
First unveiled at Tesla’s AI Day in 2021, Dojo was pitched as a breakthrough that would scale back reliance on third-party chip suppliers reminiscent of Nvidia, whereas delivering higher bandwidth, decrease latency and decrease prices.
The choice follows months of inner turbulence within the Dojo division, together with the lack of a number of key engineers. Musk famous that it now not made sense for Tesla to divide its sources between two totally different A.I. chip product strains—one optimized for inference and one other for coaching. Inference refers back to the strategy of operating an already-trained A.I. mannequin to make real-time choices, reminiscent of figuring out objects on the street in a shifting automotive. Coaching, in contrast, is the computationally intensive strategy of instructing an A.I. mannequin by feeding it large quantities of information till it could actually acknowledge patterns precisely.
“In a supercomputer cluster, it will make sense to place many AI5/AI6 chips on a board, whether or not for inference or coaching, merely to cut back community cabling complexity and value by a number of orders of magnitude,” Musk wrote on X.
The shutdown additionally comes as Tesla faces mounting challenges in its core electrical car enterprise. Within the newest quarter ended June 30, Tesla’s EV income dropped 16 % from the earlier 12 months, whereas complete income fell 12 % year-over-year. The corporate’s U.S. market share has additionally slid sharply, falling to lower than 50 % from 75 % in 2022.
