Article reprint source: AIGC

Source: Cailianshe

Author: Zheng Yuanfang

On September 11, local time, Tesla's U.S. stocks surged by more than 10%, expanding its year-to-date gain to 122%. Its market value increased by $79.6 billion in a single day to $868.341 billion.

The trigger for the surge came from a report from Morgan Stanley, which pointed out that the Dojo supercomputer will bring Tesla a market value increase of up to $500 billion, while raising Tesla's base target price to $400. For reference, Tesla's stock price closed at $273.58 after a sharp rise on Monday, which is 46% higher than Morgan Stanley's expectations.

Analysts Adam Jonas and Daniela M Haigian pointed out in the report that the autonomous driving system is called the "mother of artificial intelligence projects". In the process of seeking to solve the problem of autonomous driving, Tesla developed the Dojo supercomputer, which can open up "new potential markets" for Tesla.

What is Dojo?

Dojo is a supercomputer developed by Tesla itself that can use massive video data to complete "unsupervised" data labeling and training.

Literally speaking, Dojo means "dojo, martial arts hall", which is exactly what it means - the training ground created by Tesla for AI.

At the AI ​​Day in 2021, Tesla released the Dojo supercomputer, but at that time it was still "fledged", with only the first chip and training block, and the company was still pushing to build the complete Dojo Exapod.

Tesla also said that in theory, Dojo ExaPod will be the world's fastest AI training supercomputer. Later, Dojo Exapod was finally unveiled. Each Dojo ExaPod integrates 120 training modules, built-in 3,000 D1 chips, has more than 1 million training nodes, and a computing power of 1.1 EFLOP* (quadrillion floating-point operations per second). In terms of microarchitecture, each Dojo node has a core and is a mature computer with CPU-specific memory and I/O interfaces.

Currently, Dojo is used for artificial intelligence machine learning and computer vision training. Tesla has started producing Dojo supercomputers in July this year, taking an important step towards faster and cheaper neural network training; the company plans to invest more than $1 billion in Dojo by the end of 2024; in the future, Tesla plans to use the computing power of both Nvidia and Dojo.

According to the computing power development plan released by Tesla in June, Dojo will become one of the top five computing facilities in the world in the first quarter of next year and will reach 100 EFlops computing power in October next year.

▌The cornerstone of Musk’s “AI empire”

It can be said that Dojo has become the cornerstone facility of Musk's "AI empire".

Looking back at what Musk said when he first introduced Dojo to the public in 2019:

Tesla does have a major project, which we call Dojo, which is a super powerful training computer whose goal is to take in a huge amount of data and train it at the video level… With the Dojo computer, we can do unsupervised training at scale on a large amount of videos.

Indeed, if the nourishment of artificial intelligence is data, then the nourishment of Tesla's autonomous driving is video data.

In order to achieve full neural network control rather than code control, FSD V12 obtains about 160 billion frames of video from the Tesla fleet every day for training, but less than 1% of them are the most useful. Musk said that the neural network envisioned by Tesla must be trained on at least 1 million videos before it can take shape. By the beginning of this year, FSD V12 had completed 10 million video analyses.

Just at the end of August, Musk showed off FSD V12 through live broadcast.

In this live broadcast, the vehicle was able to give way to pedestrians, avoid roadblocks, and turn at intersections on its own on non-preset roads. Musk repeatedly emphasized that there is no corresponding line of code in FSD V12 to manually set the vehicle to perform these actions - FSD 12 completes these actions entirely as a result of a large amount of video training. Through video training data, AI can learn to drive by itself and "do things like humans."

Of course, mediocre and random data is not enough, and the data supplied to the neural network needs to be carefully selected. Musk also emphasized that high-quality data from excellent drivers is the key to training Tesla's autonomous driving.

"A lot of mediocre data doesn't improve driving, and data management is quite difficult. We have a lot of software that controls what data the system selects and what data it trains on."

▌Towards General Artificial Intelligence

Today, for Musk, the value of Dojo is no longer limited to the autonomous driving business. In fact, Dojo has become the computing power infrastructure for the development of Tesla's entire AI business system.

About 70% of human information is acquired through visual perception, which is also the solution Tesla plans to prepare for cars and robots.

Tesla has previously revealed that the head of Tesla Bot "Optimus Prime" will be equipped with the same smart driving camera as its own cars, and will share the AI ​​system with the cars. In other words, Tesla's humanoid robot continues the vision-based sensing technology route.

Musk revealed in June this year that Dojo has been online and running for several months, and it is not only suitable for Tesla's fully autonomous driving. In addition, he said that Dojo V1 is highly optimized for large-scale video training, not for general purpose AI; but Dojo V2 will break this limitation.

This also means that the upgraded version of Dojo is more likely to target general artificial intelligence (AGI).

This is also mentioned in the Morgan Stanley report mentioned above. After comparing it with other technology companies' supercomputers, Morgan Stanley found that Dojo's future looks brighter. Considering Tesla's upcoming autonomous robot taxis and network services, Dojo's landing seems to be more clear and may promote a major upgrade of Tesla's ecosystem.