Tech & Auto
Here’s What You Should Know About Tesla's Dojo Supercomputer
By NADEEM SARWAR
Tesla has repeatedly stressed that it’s not just a carmaker but a technology company, so it's no surprise that it developed the world's fifth-most powerful supercomputer in 2021.
Tesla began laying the foundations for a more powerful version called Dojo in 2023. Dojo can help the company with various projects, such as Autopilot.
The company worked on several ideas to speed up the neural network processes and create a scalable system that could keep up with AI models as they evolve in complexity.
Dojo uses a training tile that integrates a 5x5 array of 25 self-designed D1 chips. Thanks to its design and components, the chip can deliver 362 TFLOPs of computing power.
Project lead Ganesh Venkataramanan shared at the AI Day event, "This chip is like a GPU-level computer with a CPU-level flexibility and twice the network chip-level I/O bandwidth."
The D1 tile includes a custom voltage regulator module with an area matching that of the silicon die, delivering an impressive power density close to 1A per millimeter square.
To help with their scalable system, Tesla created an interface processor with a memory bandwidth of 800 GB/s that runs a custom protocol called Tesla Transport Protocol (TTP).
Other elements include a PCIe Gen 4 interface and an ethernet port with a throughput of 50 GB/s, which can connect to a Dojo Host Interface for video-based training.
The assembly is then put together in the tower cabinet. These cabinets can be connected to form the Dojo Accelerator, which can deliver ExaFLOP-class computing power.
Tesla says the Dojo will become the fastest training computer, delivering roughly four times the raw performance output at the same cost with a footprint that's five times smaller.