Tesla’s ‘Dojo’ Supercomputer For FSD Is Dead
The data center was built to analyze video footage from cars running FSD and learn from it for future decisions.
Photo by: InsideEVs
By: Iulian Dnistran
Aug 8, at 4:31am ET- Tesla has pulled the plug on the Dojo supercomputer that was designed to make its Full Self-Driving software better.
- The data center used multiple custom-built chips known as D1 to train artificial intelligence (AI) for driving assistance systems and the Optimus humanoid robot.
- The supercomputer was estimated to add $500 billion to Tesla’s value.
Tesla is killing its Dojo supercomputer, which helped train the algorithms that power the company’s advanced driving assistance systems like Autopilot and supervised Full Self-Driving, as well as the Optimus robot.
The decision, first reported by Bloomberg quoting people familiar with the matter and then confirmed by Tesla CEO Elon Musk, comes as the American company is slowly moving away from in-house development for AI solutions and instead striking deals with third parties like Nvidia, AMD and Samsung.
 Tesla Dojo supercomputer cabinet
 Photo by: Tesla
As part of the Dojo supercomputer shutdown, Peter Bannon, who was the team leader, will leave the company. Meanwhile, other team members will be reassigned to other data centers and compute projects at Tesla, according to Bloomberg’s sources.
Earlier this year, during the most recent quarterly earnings call, Musk hinted that the company could drop the Dojo supercomputer and instead work more closely with external partners. “Thinking about Dojo 3 and the AI6 inference chip, it seems like intuitively, we want to try to find convergence there, where it’s basically the same chip,” Musk said last month.
Despite killing the Dojo project, which lived in New York, Tesla is not abandoning its AI-powered ambitions. A bigger and better supercomputer dubbed Cortex entered the construction phase last year in Austin, Texas, powered by over 100,000 Nvidia H100 and H200 chips. Another data center is live in Memphis.
So, even without the Dojo supercomputer, Tesla’s AI-training cluster is still growing, as the company is working hard to bring unsupervised Full Self-Driving to the market after years of broken promises.
While the data centers’ main job is to analyze video data from Tesla EVs driving around the world, there’s also a hardware component to the company’s autonomous driving efforts. To up the ante here, Tesla has struck a $16.5 billion deal with Samsung to secure AI semiconductors that would be the base for the automaker’s AI6 (also known as Hardware 6) architecture. The deal, which is valid through 2033, includes building a factory in Texas to produce the chips locally.
The current generation of vehicle hardware chips, AI4 (or Hardware 4), is also made by Samsung, while the upcoming AI5 (Hardware 5) units will come from TSMC, initially from Taiwan and then from Arizona.
insideevs.com |
