Tesla's Dojo supercomputer tripped the grid.
According to reports, Tesla's newest iteration of the Dojo supercomputer was so powerful that it knocked down the electricity supply in Palo Alto at its unveiling.
Known internally as "Dojo," Tesla's supercomputer platform was developed expressly for artificial intelligence machine learning and, more significantly, for video training utilizing data collected from its fleet of automobiles.
The company already has a massive, one of the world's most potent, NVIDIA GPU-based supercomputers. Still, the new Dojo custom-built computer employs processors and an entire infrastructure constructed by Tesla.
Tesla's computer vision technology, which is powering its self-driving endeavor, relies heavily on training neural nets using video data, and the custom-built supercomputer is intended to improve the company's ability to do so.
Tesla introduced its Dojo supercomputer during last year's AI Day, although the business was still in the early stages of developing AI. It was still developing a complete Dojo cabinet and training tiles, which it called an "Exapod," It just had its initial chip and training tiles.
Last night, at its AI Day 2022 event, Tesla revealed the year's worth of Dojo program advancements.
A system tray and complete cabinet were added to the company's previous chip and tile setup, which they validated.
According to Tesla, one Dojo tile can do the work of six GPU boxes at the same price as only one GPU box. Each tray has six such tiles.
According to Tesla, "three to four fully-loaded supercomputer racks" may be represented by a single tray.
The business is building a massive, all-host assembly by incorporating its host interface into the system tray.
Two of these system trays, together with the host assembly, may fit inside a single Dojo cabinet from Tesla.
The closed and opened Dojo cabinet is seen here:
As announced by Tesla, the primary specifications of a Dojo Exapod are 1.1 EFLOP, 1.3 TB SRAM, and 13 TB high-bandwidth DRAM.
Even though the company's primary goal in holding the event was to attract new employees, it also utilized the opportunity to announce that its first complete cluster, or Exapod, is on track to be ready in the first quarter of 2023.
Seven Dojo Exapods are in the works for Palo Alto right now.
Exactly what use does the Dojo supercomputer serve for Tesla?
Your inquiry is valid. For what purpose is the world's most powerful supercomputer being developed by a car company? To hear Tesla tell it, the business is a technology firm creating goods to hasten the advent of a sustainable economy, not simply a carmaker.
According to Musk, Dojo as a service makes sense, compared to a "service that you can use that's accessible online where you can train your models considerably quicker and for less money." This might be an indirect shot at Musk's friend Jeff Bezos and his Amazon Web Services (AWS).
To develop its autonomous driving system, Tesla needs Dojo to identify fleet training footage and educate its neural networks automatically.
Tesla has recognized the need for substantial computational capacity to implement its strategy for building a self-driving system based on neural nets trained on millions of films originating from its client fleet. And so it set out to develop its supercomputer to meet that need.
That's the immediate objective, but Tesla has far more significant long-term AI program development aspirations so that the supercomputer will be put to good use anyway.