At the highly anticipated AI Day, Tesla revealed many details about a new supercomputer called Dojo. The company talked about the team's achievements in the development of hardware and capability, as well as its plans to make Dojo available to other AI developers in the future.
Tesla cars process data on a massive scale utilizing in-house-built chips. Its vehicles need a powerful computer on board, and so the company had to develop its own chip capable of processing hundreds of neural tasks and thousands of images per second. Basically, the chips need to be able to work in real-time, processing extremely complex information.
At Tesla AI Day, which took place on August 19 in Fremont, the company unveiled its Dojo supercomputing technology. Dojo's goal is to increase the speed and accuracy of training. Dojo is a Neural Network (NN) chip being developed by Tesla's hardware team to increase the learning speed of neural networks at the server-side.
Tesla's new development was presented by Ganesh Venkataramanan, Senior Director of Autopilot Hardware at Tesla and Dojo Project Manager. He unveiled the Dojo D1 chip, which uses 7nm technology and delivers breakthrough bandwidth and computing performance.
The D1 chip was completely developed by the Tesla team internally. "This was entirely designed by the Tesla team internally. All the way from the architecture to the package. This chip is like GPU-level compute with a CPU level flexibility and twice the network chip-level IO bandwidth," Venkataramanan said.
Each of the D1's 354 chip nodes reportedly has one teraflops (1,024 gflops) of compute. Venkataramanan said the entire chip was capable of up to 363 teraflops of compute as well as 10tbps of on-chip bandwidth/4tbps of off-chip bandwidth.
To house the chips, the company has developed what it calls 'training tiles' with which to build its computing systems. Each tile consists of 25 D1 chips in an integrated multi-chip module, with each tile providing nine petaflops of computing and 36tbps of off-tile bandwidth.
To create Dojo, the company plans to install two trays of six tiles in a single cabinet, for 100 petaflops of compute per cabinet. Once complete, the company will have a single 'Exapod' capable 1.1 exaflops of AI compute via 10 connected cabinets; the entire system will house 120 tiles, 3,000 D1 chips, and more than one million nodes.
According to Tesla CEO Elon Musk, the system will be launched next year. The company also plans to make Dojo available to other AI developers in the future.
© 2021, Eva Fox. All rights reserved.
We appreciate your readership! Please share your thoughts in the comment section below.