On August 15th, Elon Musk tweeted that Tesla is recruiting AI or chip talents for a neural network training computer project named “Dojo”. The goal of Dojo supercomputer is to process the vast amount of video data generated by more than 750,000 Tesla cars around the world, each equipped with 8 surround cameras.
Tesla has always used cameras to do visual detection instead of LiDARs. If the average user drives for about an hour a day, the entire Tesla fleet will generate about 20 million hours of 360-degree video per month. With 8 cameras per vehicle, the fleet will generate about 170 million hours of video per month. It is impossible to manually label all videos, let alone being economically feasible. Therefore, having the data learn to self-supervise will greatly improve data processing efficiency.
In November 2019, Tesla’s senior AI director Andrej Karpathy revealed in a speech that the goal of Dojo training computers is to increase performance by an order of magnitude at a lower cost. At present, Dojo is still under development, starting from version 1.0, which will be released in about a year. Elon Musk explained on Twitter that the team had already simulated the operation of this supercomputer with FPGA at about 0.01 percent capability.
The Dojo is another “beast” in the field of autonomous driving after Tesla’s FSD chip. On August 4, 2019, Musk first mentioned the project on Twitter, officially announcing it on the Tesla Investor Day five days later. At the World Artificial Intelligence Conference in July 2020, Musk once again referred to Dojo and its current bottlenecks in chip heating and computing speed. To these ends, Tesla is also developing a new bus and cooling system to serve as the upgraded auxiliary to computing units. If all goes well, “Dojo” can improve the way Tesla Autopilot works —— the system is currently operating in “2.5D” and Musk plans to upgrade to a “4D” environment.
At present, Tesla Autopilot has 300 talented engineers plus more than 500 highly skilled data annotators (with plans to enlarge the latter team to 1,000 people). Musk admits during an interview that data tagging is an arduous job, and it does require skills and training, especially when it comes to 4D (3D plus time series). Moving forward, the company has plans to enlarge the data annotation team to 1,000 people, to back its camera solutions for fully autonomous cars. (Source)
We know you don’t want to miss any story. Subscribe to our popular Synced Global AI Weekly to get weekly AI updates.
very good thanks