Tesla has unveiled the Dojo supercomputer for the first time even though Elon Musk has been talking it up on Twitter for almost a year now. It unveiled it during its AI day conference where the world's most capitalised automaker has claimed that it is the fastest in the world for training machining learning algorithms. This is particularly important for computer vision which is a branch of AI that helps develop intelligence for self-driving cars using cameras. Tesla has been handling tons of data from its active fleet of cars that have cameras and its autopilot ADAS software. It has over 1 million vehicles out there that it uses to train the neural network using the Dojo supercomputer.
Also Read: Tesla Bot unveiled
Last year it was teased that the Dojo supercomputer had more than an exaflop of capacity which is equal to 1,000 petaFLOPs. This could make the Dojo supercomputer potentially the most powerful one in the world. It is powered by its custom chip which Tesla calls the D1 which is based on a 7-nanometre manufacturing node. This is the second chip apart from the FSD chip that the company has developed in-house.
It is a 7nm meter chip which is the 2nd one Tesla has deisgned in-house
"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 bandright," said Ganesh Venkataramanan, Tesla's senior director of AutoPilot hardware and the leader of the Dojo project.
Uniquely, Tesla has designed the chip in such a way that they can connect without any glue to each other. This way it was able to connect 500,00 nodes which add 9 petaflops just with the training tile with 36TB per second bandwidth is less than 1 cubic foot format. Tesla can also combine 2 by 3 tiles in a tray and 2 trays in a computer cabinet of over 100 petaflops per cabinet. Tesla believes it can link all these to create the ExaPod.
Tesla hasn't put the entire system together but Musk believes that it will be fully operational next year. This way it becomes one of the fastest supercomputers in the world, which is the fastest for training AI while also being very power efficient and that too in a relatively compact package. Tesla will not only train its neural networks on it but will eventually open it up to AI developers. Tesla believes that it can at least aim for 10x improvement in the next iteration of the computing appliance.
