Tesla builds the world’s fifth most powerful supercomputer

Tesla continued its radar sensor program last month. After a while, he decides to launch a camera-based system within his “Model Y” and “Model 3” cars.

At the Computer Vision and Pattern Recognition Conference 2021, Testa’s Chief AI Officer Andrzej Karpathy announced the shutdown of a supercomputer. As for FLOPS – the number of floating point operations per second – it would be the fifth most powerful supercomputer ever built.

What are Tesla’s plans for a supercomputer?

These devices will be used outside vehicles to design popular ones Tesla self-driving software. To be more precise, it is the brand’s automation pilot and its FSD or full self-driving IA. At the same time, the California company needs powerful engines to operate within its electric vehicles.

This is what Andrzej Karpathy says “Computer View” fiancee “Bread and Butter” Tesla activities. At the same time, it allows the automated pilot to function properly. According to him, we are currently facing a massive demand for data from the Navy. Since Tesla invests a lot in computing, it is necessary to build massive neural networks.

Note that these neural networks contain more than 1.5 petabytes of 4D data from Tesla Autonomous Driving Software. The latter contains videos of cameras fitted inside cars. They are then used to train the software so that it can move autonomously through cameras and radars.

However, in May, the US company dropped the radar. He decides to focus on his “Model 3” and “Model Y” cameras.

What are the features of Dojo Supercomputer?

The supercomputer program is not recent. In 2020, Elon Musk lets go into a tweet: “Tesla is developing an NN training system called Dojo to process large-scale video data. Join our AI or PC / Chip teams if this is a beast!

This suggests that the company Goes down to the race for the fastest supercomputer In this world. However, it should be noted that the speed of 415 petaflops is higher than Japan’s supercomputer Tesla.

For Dojo, its 1000 betaflope capacity will allow it to achieve standard levels of training that are not overseen on video data. It will be presented to the Central Tesla organization and will contribute to the training of its autonomous AI. Super computer game Total switching capacity of 640 dbps, 720 8xA100 80 GB nodes, 10 ppm NVME storage 1.6 dbps Finally 1,8 EFLOPS.

The dojo function, according to Muskin, will be accessible to anyone who wants to train for neural networks.