Dojo

Tesla's custom-built supercomputing architecture designed for exascale AI training via the D1 chip.

Dojo represents a vertical integration of silicon and software designed to break the bottlenecks of traditional GPU clusters. At its core, the D1 chip delivers 362 TFLOPS of BF16 power, interconnected into a massive functional unit called a Training Tile. By utilizing a high-bandwidth fabric (9 TB/s per tile) and custom compiler stacks, Dojo accelerates the processing of millions of video clips from the Tesla fleet. This infrastructure reduces the latency of large-scale neural network training, enabling rapid iteration for Full Self-Driving (FSD) and the Optimus humanoid robot.

https://www.tesla.com/AI