Elon Musk has been vocal about Tesla’s Dojo, an AI supercomputer Autonomy poised to be integral to the company’s AI-driven future. Musk recently announced a renewed focus on Dojo as Tesla prepares to introduce its robotaxi in October. But what exactly is Dojo, and why is it pivotal to Tesla’s strategy?
What is Dojo?
Dojo is Tesla’s custom-built supercomputer designed to train the “Full Self-Driving” (FSD) neural networks. Enhancing Dojo is crucial for Tesla’s ambition to achieve full self-driving and launch a robotaxi. Currently, FSD is installed in about 2 million Tesla vehicles, enabling some automated driving tasks but still requiring human oversight. Despite delays in the robotaxi reveal from August to October, Tesla remains committed to achieving full autonomy, investing significantly in AI and Dojo.
The Origins of Dojo
Elon Musk envisions Tesla not just as an automaker but as an AI company mastering self-driving technology by emulating human perception. Unlike other companies that use a mix of sensors (lidar, radar, cameras) and high-definition maps, Tesla relies solely on cameras for visual data, processed by advanced neural networks to make driving decisions. This approach aligns with Andrej Karpathy’s statement at Tesla’s first AI Day in 2021 that the company aims to build “a synthetic animal from the ground up.”
How Dojo Works
Dojo is crafted to function as a training platform for AI, particularly FSD. A supercomputer comprises thousands of smaller computers, each with its own CPU and GPU. CPUs manage overall operations, while GPUs handle complex tasks like parallel processing essential for machine learning. Tesla currently uses Nvidia GPUs for AI training but aims to create a more efficient system with its custom hardware.
Why Tesla Needs Dojo
Tesla’s camera-only approach necessitates enormous computational power to process vast amounts of video data from its vehicles and run simulations for AI training. This vision aims to replicate human visual and cognitive functions in driving. As Tesla amasses more driving data, Dojo will be pivotal in storing, processing, and training neural networks to improve FSD.
Tesla’s Custom D1 Chips
To reduce reliance on expensive Nvidia chips, Tesla developed the D1 chip, unveiled at AI Day in 2021. Manufactured by TSMC using a 7-nanometer process, the D1 chip boasts 50 billion transistors and is optimized for AI workloads. Despite not being as powerful as Nvidia’s A100 chip, the D1’s design allows for efficient machine learning operations.
Tesla combines 25 D1 chips into a tile, each with a compute power of 9 petaflops and 36 terabytes per second of bandwidth. Multiple tiles form racks, and racks form cabinets, collectively making up the supercomputer. Tesla plans to scale Dojo by deploying multiple ExaPODs, comprising numerous cabinets, to enhance computational capacity.
Future Implications of Dojo
Dojo’s custom chips could enable Tesla to expand its AI training capabilities cost-effectively and reduce dependence on Nvidia’s increasingly expensive and scarce hardware. Musk’s goal is for Dojo to handle half of Tesla’s AI hardware needs within 18 months, with Nvidia and potentially AMD chips covering the rest.
In the long term, Dojo could open new revenue streams for Tesla through AI services and robotaxis. Morgan Stanley’s September 2023 report suggested Dojo could add $500 billion to Tesla’s market value. While Dojo currently focuses on Tesla’s AI needs, future versions might cater to general AI training, potentially offering cloud computing services similar to AWS and Azure.
Current Status and Future Prospects
Production on Dojo began in July 2023, and Tesla expects it to be one of the top five supercomputers by early 2024. The company plans to scale Dojo to reach 100 exaflops by October 2024, requiring over 276,000 D1 chips. Tesla has also committed $500 million to build a Dojo supercomputer at its Buffalo gigafactory and is setting up a dense, water-cooled supercomputer cluster at its Austin facility.
Tesla’s AI team currently uses Tesla HW4 AI computers alongside Nvidia GPUs for training, with plans to integrate Dojo fully by the end of the year. This hybrid approach ensures Tesla has sufficient computational power to advance its AI ambitions.
Conclusion
Tesla’s Dojo represents a significant investment in AI, aiming to push the boundaries of autonomous driving. While the journey is fraught with challenges and uncertainties, the potential rewards are substantial, positioning Tesla at the forefront of AI and autonomous vehicle technology.
Link
Readmore about AI competition
