It is the generation task of the automotive industry: autonomous driving. Tesla now wants to make the dreams of the future a reality with the help of a new supercomputer. The new system should be able to join the top ten most powerful computers in the world.
• Tesla presents new supercomputer and announces successor model
• New system will be one of the most powerful in the world
• It is to be used for autonomous driving according to the vision-only approach
Tesla’s new high-performance computer with NVIDIA technology is intended to represent a further step in the direction of autonomous driving.
Tesla’s AI boss presented computers at a virtual conference
The head of artificial intelligence (AI) at Tesla, Andrej Karpathy, recently presented the new supercomputer at a virtual presentation of the Conference for Computer Aided Image Recognition (CCVPR) 2021. The computer is supposed to be the solution to the generational task of autonomous driving for the automotive industry.
According to Tesla’s AI boss, the supercomputer with NVIDIA GPUs ranks around fifth in the top 500 ranking of supercomputers in terms of computing power. According to the online portal t3n, the graphics processor and chipset manufacturer NVIDIA currently has only 4480 instead of 5760 graphics processors. It recently replaced the Microsoft counterpart. The Elon Musk group is already working on a successor model under the project name “Dojo” , Karpathy explained in the video conference.
Powerful system to improve autopilot
As Karpathy describes, the new supercomputer is supposed to process incoming information from the Tesla vehicles and thereby continuously improve the AI of the autopilot. It is intended to serve the neural network of the autopilot and the system’s AI, which is still under development. The company is thus continuing its path of equipping data centers and vehicles with ever more powerful computers.
According to Tesla’s AI director, the 5,760 processors in the supercomputer work with a data transfer rate of 40 terabytes per second. The vehicles require such a high computing power to run the driver assistance software. Because the vast amounts of data go from the cars to the computing centers, which not only control the neural network, but also the AI’s learning function.
Autonomous driving based on the vision-only approach
“Computer vision improves the autopilot. So that we can implement that, we have to train a huge, neural network and experiment a lot,” Karpathy explained to the viewers in the virtual room. The vision-only approach used by Tesla means that lidar and radar sensors in vehicles are being replaced by high-end optical cameras. This immense data set of videos requires a correspondingly powerful computer. Finally, the system collects videos from eight cameras at 36 frames per second per vehicle.
Elon Musk had already stated several times in the past that he only wanted to rely on cameras in the future. Cameras work faster than lidar and radar and the Model 3 and Y in the USA are now only built without radar. In addition, the technology could theoretically work anywhere in the world, as no high-resolution maps are required, adds t3n.
- Renault wants to build affordable electric compact cars : Effecting Shares
- Big discount in valuation: are quality stocks on the verge of recovery?
Autonomous driving should be safe
In the video conference, Karpathy also emphasised the safety aspects of autonomous driving. Once fully developed, it could even be safer than driving manually. With a new function such as “Pedal Misapplication Migration”, for example, the system can interpret the mix-up of accelerator and brake pedals and intervene accordingly. The human delay caused by reaction time is then eliminated. At the moment, according to Karpathy, the systems still have some problems, especially in densely populated areas such as San Francisco. However, they seem to be working very well in rural areas.
According to the Chip.de e-portal, Elon Musk had previously indicated that the Dojo system could become the most powerful computer in the world. Musk has even announced that it will also give other companies access to Tesla’s data centers so that they can use them to train their own vehicle AI.