Tesla’s Optimus robot prototype was not the only item shown at the company’s AI Day presentation on Friday evening. During Tesla’s AI day, Autopilot director Ashok Elluswamy spoke about how the company’s Full Self Driving software has progressed. He claimed that 160,000 clients are already using the beta software, up from 2,000 at this time last year.
According to Tesla, there have been 35 software versions of FSD. In a Q&A session at the conclusion of the presentation, Musk made another forecast — one of many — that the technology would be ready for a global deployment by the end of this year, but he recognized the regulatory and testing challenges that remained.
Following that, Paril Jain, Tesla’s engineering lead for Autopilot motion planning, demonstrated how FSD has improved in particular interactions and can make “human-like” judgments. For example, when a Tesla makes a left turn into an intersection, it may pick a path that avoids colliding with impediments like as pedestrians crossing the street.
Every Tesla is known to be able to give datasets for the models that FSD employs, and according to Tesla’s engineering manager Phil Duan, Tesla will soon begin generating and processing comprehensive 3D structures from that data. They claim that automobiles improve decision-making in various environmental scenarios such as darkness, fog, and rain.
Tesla’s AI software is trained on its supercomputer before being sent into customers’ cars through over-the-air software upgrades. To do this, it analyses video feeds from Tesla’s fleet of over 1 million camera-equipped cars now on the road, as well as having a simulator created in Unreal Engine that is utilized to enhance Autopilot.
The company already has a huge Nvidia GPU-based supercomputer and a data center that can house 30PB (that’s 30,000,000GB) of data. Tesla is also developing Dojo, a new custom-built computer based on Tesla-designed semiconductors that the firm claims can replace 72 GPU racks containing 4,000 GPUs with just four Dojo cabinets.
Executives showed Dojo’s initial chip and training tiles during last year’s AI day, which would ultimately mature into a complete Dojo cluster or “ExaPod.” The first ExaPod is projected to be finished by Q1 2023, according to the company’s announcement today. The business intends to construct a total of seven in Palo Alto. Tesla claims that the Dojo ExaPod, a 10-cabinet system, will break the ExaFlop computation barrier and feature 1.3TB of high-speed SRAM and 13TB of high-bandwidth DRAM.
Since last year’s AI day, Dojo development has reached many milestones, including the installation of the first Dojo cabinet, the testing of 2.2MW of load testing, and the company’s current construction pace of one tile per day. Dojo was also shown running a Stable Diffusion model with 25 Dojo dies, resulting in this AI-generated artwork inspired by the prompt “Cybertruck on Mars.”
All Tesla cars now come standard with Autopilot, a driver-assist technology. Owners may pay an extra $15,000 for the Full Self-Driving option, which Musk has consistently claimed would one day provide fully autonomous capabilities to Tesla car owners. FSD is still classified as a “Level 2” advanced driver-assistance system, which means that the driver must be fully involved in the operation of the vehicle while in motion.
FSD, which is now accessible to around 160,000 drivers in the United States and Canada, enables users to employ Autopilot’s partly automated driving aid system on city streets and smaller roads. The system claims to be able to accelerate and decelerate, turn (even unprotected left turns, which are particularly difficult for automated systems), and identify traffic lights and other road indicators.
