Tesla has announced that it has started production of its Dojo supercomputer, which will be used to train its fleet of autonomous vehicles.
In its Q2 2023 earnings report, Tesla outlined the four main technology pillars needed to solve vehicle autonomy at scale: an extremely large real-world dataset, neural net training, vehicle hardware, and vehicle software.
According to Tesla’s report, the company is developing each of these pillars in-house and has begun production of the Dojo training computer to enable faster and cheaper neural net training.
Tesla already operates a powerful Nvidia GPU-based supercomputer, but the Dojo custom-built computer will utilize Tesla-designed chips. The Dojo training computer was named by Tesla CEO Elon Musk and is expected to have exaflop-level capabilities.
According to a report by Network World, an exaflop computer system can perform 1 quintillion floating-point operations per second, which is an incredible amount of power.
At Tesla’s AI Day in 2021, the Dojo project was still in progress, with executives revealing the first chip and training tiles. Tesla plans to combine these tiles into a full Dojo cluster or “exapod” to achieve over 100 petaflops per cabinet and break the exaflop barrier.
During AI Day 2022, Tesla showcased further progress on the Dojo project, including a full system tray. The company now aims to have a complete cluster by early 2024.