Tesla set up a supercomputer project team to help improve artificial intelligence algorithms.

Che Dongxi (WeChat official account: Che Dongxi)

Make up? |? Liumao

On August 18, elon musk, CEO of Tesla, recently recruited talents for developing Dojo neural network training computer on Twitter, and claimed that the computing power (i.e. the computing speed) of this computer can reach trillions of times per second in FP32 format.

Musk tweeted to recruit talents for the Dojo.

Tesla first confirmed the existence of Dojo computer on autopilot day last year. However, at that time, Musk only briefly mentioned this plan.

During the World Artificial Intelligence Congress in 2020, elon musk mentioned this plan again. He said that Tesla developed the Dojo training system, which was developed to quickly process a large amount of video data to improve the training of artificial intelligence systems.

1. Tesla launched the Dojo Plan? In order to improve the training of artificial intelligence system.

During the 20 19 Tesla Autopilot Day, Musk and other Tesla executives introduced what Tesla is doing to achieve full autopilot capability by the end of next year.

At that time, most people's focus was on Tesla's fully automatic driving hardware? In the release of 3.0, we ignored another computer that was simply mentioned-Dojo computer. Tesla was already studying this.

According to foreign media electrek, Musk said at the event, "We do have an important plan in Tesla, but we don't have enough time to talk about this plan called ‘Dojo' today. That is a super powerful training computer. Dojo's goal is to receive massive data and train at the video level. Dojo plans to conduct large-scale unsupervised training on massive video data.

Electrek believes that Dojo means "Dojo" in Japanese, and this word is usually used to refer to places where meditation or martial arts are practiced. Therefore, Dojo computer may also become a place where Tesla trains its autonomous driving artificial intelligence system.

Musk once responded to inquiries about Dojo on Twitter.

Shortly after this, someone asked Musk on Twitter: "Eric, can you tell us about the Dojo project?" .

In this regard, Musk did not disclose any new news, but his reply seems to imply that Dojo may prompt something different to happen.

During this year's 2020 World Artificial Intelligence Conference, Musk once again mentioned the Dojo plan.

"We also developed the Tesla Dojo training system, which aims to quickly process a large amount of video data to improve the training of artificial intelligence systems. Dojo system is like an FP 16 training system, which is mainly limited by the heat output of the chip and the communication speed. Therefore, we are also developing new buses and cooling systems to develop more efficient computers to process video data more effectively. " He said in a video speech.

Now, the latest news about the Dojo project comes from Musk's Twitter.

He said in a tweet on August 15: "Tesla is developing a neural network training computer called Dojo to handle a large number of real video data. ? This is a wild animal! ? If it sounds interesting, please consider joining our AI or computer/chip team. "

More importantly, he also said that the computing power of FP32 format can reach trillions of times per second. From the perspective of the same industry, such a level of computing power will make Tesla Dojo computer a truly cutting-edge supercomputer.

Second, Musk's madness continues? Tesla plans to complete the basic functions of L5 autopilot this year.

Tesla began to lay out autonomous driving very early, and now, the American electric car manufacturer is developing autonomous driving on the track.

During the 2020 World Artificial Intelligence Conference in July, Musk, as one of the invited guests, introduced Tesla's latest progress and planning in autonomous driving in his speech.

Musk said that he is full of confidence in L5 autopilot technology, and Tesla will complete the development of basic functions of L5 autopilot system this year.

In Musk's view, the L5 autopilot system currently has no underlying challenges, and some details have not been fully handled. Although the existing technology can deal with most scenario problems, there are still some strange and abnormal scenarios from time to time, so a system is needed to find and solve these strange and abnormal scenarios.

In this regard, Tesla is constantly optimizing its software. The CEO of Tesla believes that Tesla's existing hardware has been able to achieve L5 autonomous driving, and this goal can only be achieved by continuously upgrading the software.

Musk participated in the World Artificial Intelligence Conference online.

In response to the development of AI chips, Musk said that the Autopilot autopilot chip promoted the development of AI chips. Tesla developed its own chips because the chips with strong computing power on the market have high power consumption and low power consumption, and the computing power is really not good.

At present, in the aspect of autopilot software, Tesla has developed Dojo training system, which can quickly process a large amount of video data and improve the training of artificial intelligence system.

Besides, in order to adapt to the different traffic conditions in various countries and regions, including China, Tesla has started to set up an autonomous driving engineering team in China.

"What we have to do in China is to carry out a lot of original engineering development. Therefore, instead of simply copying American things directly to China, original design and original engineering development are carried out in China. So, if you consider working, please consider working in Tesla China. " Musk said in the video speech of the day.

Conclusion: Dojo computer or Tesla autopilot system has evolved to 4D.

Tesla is planning to upgrade its autopilot computing system from the current 2.5D to 4D.

The so-called "4D" simply means adding time information on the basis of three-dimensional space information, thus forming four-dimensional information and sending it to the automatic driving computer for calculation.

At present, Tesla can only use the camera to generate two-dimensional image data, and the autopilot computer can identify, classify and mark the objects in the two-dimensional image, and finally form "2.5" data for autopilot calculation.

Compared with the previous three-dimensional information, the information richness of four-dimensional information has been improved, which can make the perception, recognition and decision-making process of Tesla autonomous driving more accurate.

Dojo computer can quickly process massive video data information. From this perspective, it can also provide some help for Tesla to realize the "4D evolution" of autonomous driving system.

This article comes from car home, the author of the car manufacturer, and does not represent car home's position.