Multi-billionaire Elon Musk announced on July 19 that the automobile manufacturer Tesla, of which he holds 23% of the capital and is general manager, will invest more than 1 billion dollars (930 million euros) in Dojo, a supercomputer whose Production began this summer. Zachary Kirkhorn, who was then CFO of Tesla (he announced his departure in early August), immediately clarified that this billion dollars was going to be invested in R&D and capital expenditure over three years.
Elon Musk justified this large sum to take advantage of the “staggering quantity” of videos and maps at its disposal, thanks to the use by Tesla customers of the driving istance software called Autopilot and the Full Self-Driving (FSD) autonomous driving program.
This represents more than 300 million kilometers of cumulative data that only a supercomputer is capable of fully exploiting. This m of information will allow Dojo to “train” the Austin (Texas) manufacturer’s autonomous driving systems on a large scale.
“Our commitment to being at the forefront of AI development has entered a new phase with the start of production of the in-house designed Dojo Intensive Training Computers. The greater the training capacity of the neural network, the more opportunity our Autopilot team has to “iterate” on new solutions.”explained Tesla management.
In computer science, iteration is a repetitive calculation process that loops until a particular condition is met. The Dojo supercomputer will thus train its machine learning models (machine learning And deep learning) to perfect advanced autonomous driving – a taste of the autonomous car.
Tesla also aims to make Dojo – from 2024 – one of the five most powerful supercomputers in the world, thanks to its own chip boosted with artificial intelligence and Nvidia graphics processors (A100), thus entering directly into the “ exascale.” Objective by October 2024: reach 100 exaflops, or 100 billion billion operations per second!
The FSD system will be offered under license to other car manufacturers. A Musk challenge, when we know that the current largest “exascale” of the Top500.org (HPE Cray Frontier with AMD chips) “only” displays less than 1.2 exaflop… The machine learning of Dojo will then be able to train the humanoid robot Tesla’s Optimus, in development for two years and also equipped with… FSD. Initially envisaged by Elon Musk for 2023, the production date has not been specified.