Tesla announces the Dojo training supercomputer

At its first AI Day, Tesla unveiled Dojo, a supercomputer it said it was building specifically to train the neural networks behind its self-driving software. Ganesh Venkataramanan, the company’s senior director of Autopilot hardware, presented the D1, a custom AI training chip built entirely in-house on a 7-nanometre process, with Tesla claiming roughly 362 teraflops of compute per chip. Groups of D1 chips were arranged into “training tiles,” which Tesla said it would combine into an “ExaPod” rack configuration designed to exceed an exaflop — one quintillion floating-point operations per second — of aggregate compute.

The stated purpose was distinctive: rather than training on curated benchmark datasets, Tesla framed Dojo around processing video recorded by its own fleet of production vehicles, then numbering more than a million cars, to improve its Full Self-Driving system. Elon Musk said the system was intended to come online the following year and, longer term, that Tesla might make its training capacity available to other developers.

Dojo made Tesla one of the first non-chipmaker, non-frontier-lab companies to commit publicly to designing its own AI training silicon rather than relying solely on Nvidia GPUs, anticipating a wider trend of large AI buyers building custom accelerators. The programme’s timeline slipped in the years that followed, and Tesla later scaled back the in-house chip effort in favour of purchased Nvidia and Dojo-successor hardware.