Tesla’s supercomputer is so rough that the network can’t handle it

Elon Musk’s company has built one of the most muscular devices on the planet.

Developing Tesla’s self-driving features requires a tremendous amount of computing power. To serve this, the company’s headquarters in California built the A supercomputer called Dojo – it says Electrek.

According to the newspaper’s information, the main purpose of the machine is the training of Tesla’s artificial intelligence, including the development of machine vision. Elon Musk’s company needs the latter because – unlike other manufacturers – they do not use radar or lidar in their cars, they only operate their driver support systems known as Autopilot and Full Self-Driving (FSD) mode based on images from the cameras. There is still a problem with these at the moment: they don’t always follow the line of the road properly, and due to the lack of radars, it has already happened that the cameras saw a motorcyclist in front of the car as a much further vehicle, which ended in a fatal accident.

Tesla Dojo is one of the most powerful supercomputers in the world (Photo: Tesla)

With the application of Dojo, they presumably want to change this as well. This summer, Tesla presented a previous supercomputer with a 7360 Nvidia video card (GPU) works, and which, based on this, is the seventh most powerful similar machine in the world. However, according to the news, Dojo also ticks off this in terms of performance, moreover, the latter supercar was not assembled from parts from third-party manufacturers, but the entire infrastructure was designed by Tesla engineers.

Previously, all that was revealed about Dojo was that they had successfully developed the chips representing the soul of the system, but now the supercomputer was actually presented in full operation. According to the company’s claim, a single unit of Dojo can replace six of the previous processors that used GPUs. According to Tesla, a single such block is equivalent to the performance of 3-4 so-called racks used in supercomputers.

The high-performance cooling system had to be developed separately for the system, but a separate energy management solution had to be created as well, because Dojo’s machines use a lot of energy. The most eloquent example of this was during a system test this year, when the local electrical network could not even handle the load when the super-smoothing machine was connected to the system.

Earlier this year, we began testing the load capacity of our power management and cooling system, which we were able to pull up to 2 megawatts before the transformer station closest to us shut down.

– explained Bill Chang, senior systems engineer at Dojo.

The machine consists of modules built into separate housings, called Dojo Exapods. Tesla also published the most important metrics of each Dojo Exapod: according to this, such a unit is capable of 1.1 EFLOP (exaflop) performance, 1.3 terabytes of high-speed RAM and 13 terabytes of broadband DRAM. For comparison, a 2022 iPhone 14 processor has a performance of 1,500 teraflops, which is equivalent to 0.0015 exaflops. According to the company’s information, the first such Exapod can be put into operation at the beginning of next year, and Tesla wants to install a total of seven Dojo Exapods in its California headquarters.

Although for the time being, as we mentioned, the car manufacturer will train its own artificial intelligence with the help of Dojo, Elon Musk believes that it is possible that in the future the supercomputer can be rented in the same way as Amazon AWS is used by thousands of users worldwide.

We still have a lot to tell you, you can find everything interesting here!


Source: PC World Online Hírek by www.zoldpalya.hu.

*The article has been translated based on the content of PC World Online Hírek by www.zoldpalya.hu. If there is any problem regarding the content, copyright, please leave a report below the article. We will try to process as quickly as possible to protect the rights of the author. Thank you very much!

*We just want readers to access information more quickly and easily with other multilingual content, instead of information only available in a certain language.

*We always respect the copyright of the content of the author and always include the original link of the source article.If the author disagrees, just leave the report below the article, the article will be edited or deleted at the request of the author. Thanks very much! Best regards!