Tesla introduced the latest version of a supercomputer called Dojo, which is the company's own development and optimized under artificial intelligence algorithms, in particular for machine learning using video data from the car manufacturer's electric vehicles. It is reported that the system is so powerful that it even shut down the power grid at Palo Alto during the presentation.
Tesla's arsenal already has a large supercomputer based on the NVIDIA graphics processors, which are among the most powerful in the world. However, Dojo is manufactured on the basis of an individual car manufacturer's order, using the chips developed by Tesla's own forces and imprisoned for its tasks.
It is expected that the new supercomputer will enhance the capacity to train neural networks via video data, which is crucial for computer vision technology, which is the backbone of Tesla's autonomous driving system.
Tesla Dojo used "plastic-on" systems. Each plate contains 25 D1 accelerators and 40 input-output modules. In addition, there are power and cooling subsystems. According to Tesla, one Dojo plate can replace six blocks of graphic processors, and one plate is cheaper to produce than a box of graphic processors.
One system tray contains six silicon plates. According to Tesla, one tray is equivalent. The company is currently continuing to test the infrastructure needed to combine several cupboards and build Dojo ExaPOD with a capacity of 1.1 Eflops with 1.3 Tbyte SRAM and 13 Tbyte DRAM with high capacity. Tesla plans to build seven Dojo ExaPOD in Palo Alto.