Tesla launches production of Dojo supercomputer

Tesla, a renowned manufacturer of electric vehicles, has stepped up production of its innovative Dojo supercomputer, which is specifically designed to train autonomous driving systems. The company has allocated a $1 billion budget for the project, with the ambitious goal of achieving 100 exaflops of performance by next October.

Dojo is equipped with chips designed by Tesla’s own engineers. This project, first announced in 2019, continues to evolve, as Elon Musk confirmed in 2021.

According to Musk, the basis for the platform is “system-on-wafer” (System-On-Wafer), where each chip is a full-fledged 300 mm silicon wafer. Tesla calls these plates “Training Tile”. Each plate contains 25 D1 accelerators, while they consume 15 kW of electricity. A rack containing six of these platters is capable of delivering 100 petaflops, so 10 racks generate 1 exaflops.

Tesla affirmatively states its intention to achieve a total system performance of 100 exaflops by October next year. To illustrate, the most powerful supercomputer to date, Frontier, has a maximum computing power of 1,679 exaflops. Thus, Tesla set itself the goal of outperforming the Frontier by almost 60 times in just 15 months.