An in-depth look at the upcoming Tesla Dojo AI Supercomputer at Hot Chips 34
Tesla’s Dojo is the company’s custom supercomputer platform constructed by Tesla for AI machine learning and video training by utilizing the video data delivered from its line of vehicles. An extensive backbone of the Dojo’s microarchitecture is run on video data and is more advanced than merely looking at static images or basic text. This advancement is crucial for Tesla as the company requires a hefty amount of computing for artificial intelligence and standard autonomy for self-driving vehicles. The most significant technological advancement for the Dojo is its System-On-Wafer solution. There is an integrated D1 die processing at 15 kilowatts for every twenty-five dies. Included with the twenty-five D1 dies, Tesla is also incorporating forty I/O dies that are much smaller. Cooling and power are integrated into the training tile on the Tesla Dojo. The tiles are scaled between additional tiles with a link total of 9TB/s and must be attached instead of relying on servers. Tesla’s first version of the Dojo Interface processor offers high bandwidths of memory on a PCIe card which can be either standalone or on a server. The Dojo uses the in-house Tesla Transport Protocol (TTP) interface offering extreme levels of bandwidth. The Tesla Transport Protocol also uses Ethernet by the included software to locate a consistent address space across the Tesla system or be bridged to the interface card. The Tesla Dojo custom AI supercomputer is packed with as many as five cards utilizing a PCIe host server and provides 4.5TB/s bandwidth on each training tile. News Source: Serve the Home
