Tesla and its dream for autonomous robots is still in works. The company has announced that it is building a foundation model for autonomous agentic robots. The company already has multimodal neural networks in vehicles that take inputs from maps, cameras, and GPS.

And will be trained on enormous amounts of compute pic.twitter.com/BsmG9Vse6I

— Tesla AI (@Tesla_AI) June 21, 2023

For training this foundation model, the company is now starting production of Dojo supercomputer from next month to enable large amounts of computation. According to the graph shared by Tesla, the GPU clusters built by the company will be ten times bigger and the production rate will be as close to, if not higher than NVIDIA. By October 2024, the company expected to produce 100 Exa-flops of compute.

Tesla is now able to reconstruct the entire scene with just multiple videos from the location shot during trips. This enables Tesla vehicles to take steps to avoid collisions by occupancy prediction. To combine this with auto-labelling capabilities of Full-Self Drive of Tesla, the heavy compute capabilities is a necessity for the company. 

Moreover, the company is now also building generative modelling techniques for predicting possible future outcomes based on past observations. This is done by linking multiple cameras across the body of the car. Tesla says that these foundation models that are generating the world would be used for both its cars and Optimus robots. 

Though it is not yet confirmed if Tesla would sell its chip or keep it for itself, this could be one of the biggest revenue generators for the company in the foreseeable future, as we can clearly see how NVIDIA is rising in the AI world. Tesla is clearly not a car company anymore. 

The post Tesla to Contend NVIDIA with Dojo appeared first on Analytics India Magazine.