Earlier this week, reports emerged stating that Tesla has stepped back from its Project Dojo initiative. While the reports were initially framed as a negative development for the electric vehicle maker\u2019s autonomous driving efforts, CEO Elon Musk later noted on X that Tesla was indeed halting its Dojo initiative.<\/p>\n
Elon Musk\u2019s Confirmation<\/strong><\/p>\n As per Musk, Tesla was shuttering Project Dojo because it does not make sense for the company to divide its resources and scale two different AI chip designs<\/a>. Dojo, after all, is designed to train the company\u2019s autonomous driving program, and thus, it would not be rolled out to Tesla\u2019s consumer products. <\/p>\n In a series of posts on X, Musk stated that it would make sense to just use Tesla\u2019s AI5\/AI6 to train its FSD and Autopilot systems. \u201cIn a supercomputer cluster, it would make sense to put many AI5\/AI6 chips on a board, whether for inference or training, simply to reduce network cabling complexity & cost by a few orders of magnitude,\u201d Musk said.<\/p>\n Tesla\u2019s AI5 and AI6 chips are expected to be rolled out to the company\u2019s consumer products, from Optimus to the Cybercab to the next-generation Roadster.<\/p>\n AI6 is Dojo\u2019s Successor<\/strong><\/p>\n What was particularly interesting about Musk\u2019s comment was his mention of using AI5\/AI6 chips for training. As per Musk, this strategy could be seen as \u201cDojo 3\u201d in a way, since the performance of Tesla\u2019s AI5 and AI6 chips is already notable. Musk\u2019s comment about using AI6 chips for training caught the eye of many, including Apple and Rivian alumnus Phil Beisel<\/a>, who noted that \u201cAI6 is now Dojo.\u201d<\/p>\n \u201cDojo is Tesla\u2019s AI training supercomputer, built around a custom chip known as the D1. The D1 and AI5\/AI6 share many core design elements, particularly the math operations used in neural networks (e.g., matrix multiplication) and highly parallel processing.<\/p>\n \u201cDojo had a unique feature: chips arranged in a 5\u00d75 grid using a system-on-wafer design, with etched interconnects enabling high-speed data transfer. In a sense, Dojo will live on as the generalized AI6. Going forward, all efforts will focus on AI6,\u201d the tech veteran wrote in a post<\/a> on X.<\/p>\n