On the morning of August 20, Beijing time, Tesla, Inc. AI Day was held as scheduled, and as its name indicates, the activity revolves around AI artificial intelligence. Unlike in the past, Tesla, Inc. did not invite the media or investors, but only engineers in related fields, so after putting aside the issue of commercialization, this event is more like a forward-looking report and demonstration on technology.

On AI, Andrej Karpathy, director of artificial intelligence and autopilot vision of Tesla, Inc., gave a detailed introduction to Tesla, Inc. 's "neural network." Tesla, Inc. executives also highlighted Tesla, Inc. 's progress in automatic marking and other technologies, and the previously closely watched supercomputer, the Dojo, was also officially unveiled.

On closer inspection, it is not difficult to find that all the technologies introduced by Tesla, Inc. AI seem to convey a faint signal: even if more and more car companies adopt lidar technology to achieve autopilot, Musk will still be a maverick and will follow a "pure visual route" to the end.

The "dojo" Dojo with super power of calculation

Five days ago, Tesla, Inc. officially released a poster to warm up the AI Day event.

The poster shows the structure of a large-scale chip module, including chip core, copper plate, radiator, metal shell and other elements, and there is speculation about whether this has anything to do with the new product to be announced by Admiral Tesla, Inc. AI. This has proved to be the case.

This is actually a training module with 25 D1 chips, and the self-developed AI training chip D1 can be said to be one of the most important technological breakthroughs in AI.

D1 chip is independently developed by Tesla, Inc., 7 nm manufacturing process, a single FP32 reaches the computing power of 22.6 topsBF16 computing power 362TOPs, which is almost the strongest chip on the market.

And D1 chips can be seamlessly integrated, 25 D1 chips form a training module, and more training modules combine each other to form a supercomputer Dojo with more powerful computing power.

This is not the first time the name Dojo has appeared. Andrej Karpathy introduced the supercomputer at the CVPR 2021 workshop in June this year. At that time, the Dojo was still equipped with NVIDIA Corp's chip, with a total calculation power of 1.8 EFLOPS, and was considered to be one of the fifth largest supercomputers in the world.

After the appearance of the Dojo with D1, Wu Ren, a computer game expert and chip expert, praised it as a "wonderful design" and even said in moments, "if Elon intended, maybe this is the biggest competitor of nVidia. In fact, the seats of chip giants may need to be rearranged." "

Musk has always believed that "the only way to solve the self-driving problem is to solve the AI problem in the real world, whether hardware or software. Unless a company has strong AI ability and super computing power, it is difficult to solve the autopilot problem." "

Therefore, Tesla, Inc. launched the powerful supercomputer Dojo this time, in fact, it is also to solve the problem of autopilot. Unlike other supercalculations, the supercomputer, named "Ashram", uses all its power to do only one thing: to train the entire autopilot system, including Autopilot.

Pure visual route

The "ashram" has been built, and the trained Tesla, Inc. AI neural network is the key.

Tesla, Inc. 's neural network is mainly used to deal with the data needed for functions such as object recognition and road planning, which are the basis for supporting Tesla, Inc. Autopilot/FSD.

On AI, Tesla, Inc. Autopilot Engineering Director Milan Kovac showed the audience how Tesla, Inc. 's AI neural network was applied.

In Tesla, Inc. 's car, each car is equipped with eight cameras. After each camera acquires the original input, it can create a different resolution for different functions and purposes. Eventually, this information will be input into a complex neural network to generate additional information useful for self-driving.

The problem, however, is that even with eight cameras, there is still not enough vector space for the neural network behind it. As a result, Tesla, Inc. developed automatic marking technology, even if the field of vision is obscured during driving, according to the marking of the data, the vehicle can navigate more safely and accurately.

Andrej Karpathy also said that Tesla, Inc. 's autopilot scene simulation system can now draw road conditions in real time while driving, combined with the drawing of multiple vehicles passing through the same place to get a complete map.

And this is precisely the biggest difference between competitors such as Tesla, Inc. and Waymo: based on functions such as human brain-like visual perception system, automatic tagging, and simulation, Tesla, Inc. is trying to get rid of lidar on Autopilot/FSD, trying to take the "pure vision" route more thoroughly.

This is indeed Musk's consistent insistence.

In 2019, Musk said, "it's really stupid to put a lidar on a car," and directly split the autopilot route in two, causing thousands of waves.

The route of re-algorithm, represented by Tesla, Inc., which only uses cameras and various sensors as hardware, stands on the opposite side of the technical route of paying attention to lidar represented by Waymo.

Up to now, there has not been a decisive conclusion about which of the two can go to the end.

Lidar, as its name implies, uses the laser as the signal source, and the pulsed laser emitted by the laser hits the surrounding object to cause scattering, and part of the light wave will be reflected to the lidar receiver. Calculated according to the principle of laser ranging, the distance from the lidar to the target point is obtained. If the laser scans the target continuously, all the data can be obtained, and the accurate three-dimensional image can be obtained after imaging processing.

In Musk's view, "pure visual perception is the road to the real world AI". Since human beings drive through visual information collection and brain processing, autopilot should also be able to achieve safe driving through visual perception and algorithm processing.

When the external environment becomes more and more complex, more and more sensors are installed on the automatic car. If the information from the radar and the camera contradicts each other, the autopilot system will be more difficult to choose.

Musk refused to "pull" and chose to directly play the role of the camera to the extreme.

Implement the "first principle" to the end?

People who are familiar with Musk know that he has always adhered to the "first principle", that is, to return to the most basic conditions of things, to split them into various elements for structural analysis, so as to find the best way to achieve the goal.

At present, "perceiving and judging the driving environment" is undoubtedly the core difficulty in the field of self-driving. Musk still returns to the basis of "perception" and spares no effort to evolve AI to a level comparable to human visual perception.

Musk believes that it is almost impossible to make up for the defect that millimeter wave radar is not good at describing objects, while cameras are not good at detecting distance. In order to "get there in one step", Musk did not consider installing the lidar on Tesla, Inc..

In fact, Tesla, Inc. 's "pure visual route" requires more in-depth study of massive data and in-depth training of neural networks, which is by no means easy for other automatic car companies. Musk's persistence has something to do with his early start and the amount of data he can accumulate.

According to Goldman Sachs Group's estimate, the current number of Tesla, Inc. fleet in the world is more than 1.5 million, which provides Tesla, Inc. with a large, diversified and real-world database. This means that even if the data collection efficiency is the same, the number of Tesla, Inc. running on the road is dozens of times that of the competitor, and the amount of data is dozens of times that of the competitor.

Coupled with the fact that Tesla, Inc. 's sales soared to more than 180000 vehicles worldwide in the first quarter this year, with an increase of more than 200%, Tesla, Inc. 's profits as a whole are on the rise, and there is no need to worry about R & D funds. With sufficient funds, Tesla, Inc. dares to "throw money to open the way" for pure visual autopilot.

It is understandable that Tesla, Inc. wants to maintain his unique advantage in the technical route. according to the analysis of relevant people, if Tesla, Inc. can open up a road in "pure vision" and form a unique closed loop of self-driving technology, it will undoubtedly make Tesla, Inc. one step closer to "Apple Inc of the emus".

Of course, the business risks behind this decision can not be ignored.

Because on the other hand, with the development of technology, lidar continues to improve and optimize, the price is also gradually falling.

Lidar head company Velodyne announced this week that the price of its most popular lidar system, VLP-16, has halved from 2016. The decline in lidar prices has benefited from factors such as production capacity, equal sharing of R & D costs, and auxiliary autopilot promotion.

Nowadays, more and more car companies have chosen lidar, which means that the future "lidar" technology route may have more economies of scale and cost advantages than the "pure vision" route.

And if the future vision + lidar school does not differ much from Tesla, Inc. in autopilot ability, it means that Tesla, Inc., as a "wind breaker," the cost of exploring pure visual autopilot will be out of proportion to the benefits. This will be a lot of pressure for Tesla, Inc., who needs long-term investment.

But for Musk, he doesn't care. It seems that under the idea of the first principle, he is going to fight to the end with the "pure visual route" in the field of autopilot with Yu Gong's drive away from the mountain.