Tesla unveils new Dojo supercomputer, so powerful that it overloads the city’s electricity grid

During its new AI Day 2022 event, Tesla just introduced the latest version of its Dojo supercomputer, and it was so powerful that it brought down Palo Alto’s power grid. Instead of outsourcing, Dojo is a custom supercomputing platform that Tesla built from the ground up for machine learning and AI training tasks through videos sent in by Tesla cars.

Before that, Tesla had one of the most powerful supercomputers in the world using Nvidia GPUs, but the new Dojo supercomputer uses chips and entire infrastructure designed by Tesla. Introduced by Tesla during the AI ​​Day 2021 event, Dojo has been continuously improved by the company in its strength and capabilities during the past time and in this year’s event, Tesla is showing improvements in the project. his supercomputer.

With its power, Dojo is expected to enhance Tesla’s ability to train AI through video data – a particularly important element for computer vision technology applied to self-driving cars.

The company confirmed that it has moved from a chip and connector design to a full system tray and server cabinet. With this design, Tesla claims they can replace 6 GPU boxes with a single Dojo tray for less than a GPU box. Tesla said that a tray of these processors would be equivalent to “three to four full prices of supercomputers”.

Besides, the system tray containing these processors is also integrated directly into the server interface to form a full server cluster as shown below:

Tesla says these two system trays can fit in a cabinet of the Dojo supercomputer. Below is a picture showing what the Dojo’s locker will look like when closed and opened.

When these cabinets are put together, they form a larger cluster of servers that Tesla calls “Dojo Exapod”:

Currently, Tesla is still in the process of developing and perfecting the infrastructure for its own custom superserver. “We knew we would have to re-examine every essential aspect of our data center infrastructure to support this capability,” said Bill Chang, a key systems engineer at Tesla during the Dojo rollout. unprecedented cooling and power density.”

“Earlier this year, we started testing the load and cooling capacity of this infrastructure, we pushed power consumption to 2MW before overloading the substation and receiving calls from the city. city.” Mr. Chang added. That’s why they had to develop high-performance power and cooling systems specifically for their Dojo server cabinets.

Here are some specs of each Dojo Exapod cluster: 1.1 EFLOP processing capacity, 1.3 TB SRAM memory, and 13 TB high-bandwidth DRAM.

Owning one of the most powerful supercomputers in the world, why does Tesla need the new supercomputer Dojo?

Why would an automaker like Tesla focus on developing such a powerful supercomputer? Simply because they are not just a conventional electric vehicle manufacturer but also a technology company developing products that accelerate the transition to a more sustainable economy.

Mr. Musk said he will offer the Dojo server as a service, similar to Amazon AWS’s cloud server service, claiming it is ” a service that you can use online to train new players. ” faster and less expensive path models .”

But more importantly, Tesla needs the Dojo supercomputer to automatically label the videos of data it collects from its driving cars and then train the artificial neural network to build systems. self-driving system.

Tesla has long realized that its approach to building self-driving capabilities through millions of videos submitted by Tesla customers will require enormous processing power, so it decided to develop its own supercomputer to do that ability more efficiently.

However, this is clearly only a short-term goal and it is likely that Tesla will deploy many other applications on this supercomputer platform in the future when they are having great ambitions in developing programs. other artificial intelligence programs.

Refer to Electrek, Gizmodo