Tesla says Dojo 2 will be 10x better than Dojo 1

Tesla is working on its Dojo supercomputer to participate in the majority of full self driving (FSD) and to achieve this goal, the company already started development on Dojo 2 with 10x performance boost.

Introduced in 2019, Dojo is designed to process massive video data library for FSD training. It uses in-house designed D1 chips, processed with TSMC’s 7nm technology. The first Dojo data center is expected to house 3,000 D1 chips.

Three years later, Tesla installed the first Dojo cabinet to run tasks like Stable Diffusion. Dojo 1 is currently reported active in Giga New York and Palo Alto, handling around 5% to 10% of the training data for FSD.

Meanwhile, the company has employed more than 50,000 H100 GPUs at its facility to train the latest FSD version. Despite this massive data center, which will soon upgrade to 100,000 GPUs, the company is keen on mainstreaming its custom chipset. It is important to gain more control over vision related data and performance adaptiveness.

The current version of the D1 chip is not as powerful as its Nvidia alternative. However, Tesla 10 times performance on Dojo 2 compared to Dojo 1. We need to wait until the EV maker reveals details of the D2 chipset.