Tesla Unveils Dojo Supercomputer to Process Massive Fleet Video Data
Tesla introduces its custom-built Dojo supercomputer and D1 chip, claiming the new system is the world's fastest AI training machine. The technology processes immense amounts of video data from Tesla's vehicle fleet to train neural networks.
Tesla unveils its highly anticipated Dojo supercomputer technology at its AI Day, showcasing a system designed specifically to handle the massive amount of video data coming from its fleet of over one million vehicles. Because existing hardware options fail to meet the company's needs for training computer vision neural networks, Tesla develops the Dojo system entirely in-house to process this immense data load more efficiently.
At the center of this new supercomputer is the D1 chip, a 7-nanometer processor designed completely internally by Tesla's growing chip team. The company claims this new chip delivers GPU-level compute with CPU-level flexibility and achieves a breakthrough in bandwidth by seamlessly connecting 500,000 nodes without any glue.
By combining 25 of these D1 chips, Tesla creates a single training tile that fits in less than one cubic foot and delivers 9 petaFLOPS of computing power with 36 terabytes per second of bandwidth. While the company only shows off this single tile on stage, Tesla claims the full Dojo system has the potential to reach over an exaflop of capacity, making it the most powerful AI training machine in the world.