In addition to demonstrating the prototype of the Optimus robot, Tesla also shared new developments in other areas during AI Day, and Tesla Ashok Elluswamy, the autopilot software director, explained that the beta version of the Full Self Driving driver ' s support system was currently used by 160,000 people, compared with only 2,000 drivers last year.
A total of 35 updates of the FSD software were released. The head of Ilon Mask reiterated that the technology would be ready for global deployment by the end of this year, while recognizing that there were still problems with testing and regulatory approval.
The head of the autopilot planning team, Peryl Jane, showed improvements to the FSD, which is said to be capable of making "human" decisions. For example, when Tesla's electric car turns left at the intersection, the help system selects a trajectory in which the vehicle moves away from pedestrians crossing the street.
All Tesla electric vehicles are supplied with the Autopilot driver ' s help function. For an additional $15,000, you can get an FSD option, which is now an improved level 2 driver ' s help system. This means that the driver must keep his hands on the wheel and be ready to take over the car at any time.
The FSD system is designed, among other things, to accelerate and stop, to change lanes, to make bends, including left turns, to be extremely complex for automated systems, and to recognize traffic lights and other road signs. According to Tesla Technical Director Phil Duan, using II, the decision-making process has been improved in various environmental conditions, such as night, fog and rain.
It uses its own supercomputer Tesla based on the GPU NVIDIA, which processes information from more than 1 million electric vehicles with cameras on the roads. 30 Pbytes of footage are stored in the company's CRS. Tesla is also working on a new custom-made computer using Tesla D1 Dojo chips. According to the company, only four Dojo server cabinets are capable of replacing 72 counters with 4,000 GPU.
Tesla announced that the first cluster of Dojo ExaPod supercomputer would be completed by the first quarter of 2023. The company plans to build seven such clusters in Palo Alto. According to the company, a system with 10 Dojo ExaPod cabinets would cross the 1 exaflops line and contain 1.3 Tbyte high-speed SRAM and 13 high-capacity DRAM Tbyte. After the first AI Day event, the company has achieved a number of achievements, including the installation of the first Dojo cabinet. Tesla currently collects one "pit" per day with 25 Dojo chips per day.