Tesla, the renowned electric vehicle manufacturer, has officially activated the production of its groundbreaking Dojo supercomputer dedicated to the training of autonomous driving systems. The company has allocated an impressive budget of $1 billion for this project, with an ambitious target of achieving a performance level of 100 exaflops by October of next year.
Dojo is equipped with chips that have been specifically designed by Tesla’s own engineers. This project, which was originally announced back in 2019, has continued to evolve, as confirmed by Elon Musk in 2021.
According to Musk, the foundation of the platform is the “system-on-over” system-on-over), where each chip functions as a fully-fledged 300-mm silicon plate. Tesla refers to these plates as “Training Tiles.” Within each plate, there are 25 D1 accelerators, which consume 15 kW of electricity. A rack containing six of these plates is capable of delivering a performance of 100 petaflops, meaning that 10 racks can generate 1 exaflops.
Tesla has made a resolute declaration to achieve a comprehensive performance level of 100 exaflops across its systems by October of next year. To provide some context, the most powerful supercomputer presently in existence, Frontier, has a maximum computing power of 1.679 exaflops. This means that Tesla has set itself the challenge of surpassing Frontier by nearly 60 times in a mere 15 months.