just! Tesla released the fastest computer in history, and there is also a mysterious robot. Is the real “Iron Man” coming?

A safety accident and the death of a car owner brought Weilai and the entire new energy automobile industry to the forefront.

As a leading company in the new energy vehicle industry, Tesla is no exception. A large number of safety accidents have caused it to be criticized. In August of this year, the U.S. automobile safety regulator also opened a safety investigation on Tesla's assisted driving system AutoPilot. .

How should the different levels of autonomous driving technology be named, how safe it is, and how far Tesla's autonomous driving technology has evolved.

As the accident continues to ferment, people's doubts about "autonomous driving" will only increase.

The Tesla AI Day held in this situation will naturally become the center of new topics. What did Musk release at the event, and did he respond to relevant queries?

Come and see with us.

Tesla's self-developed chip Dojo is coming

Tesla has a new one, this time it is a self-developed chip-Dojo D1.

Earlier, Musk had mentioned Dojo on Twitter many times. This is a supercomputer chip developed by Tesla, and it is also a super AI chip.

Compared with common CPUs or GPUs, Dojo has abandoned a large number of functional modules and added more computing modules in exchange for higher computing power and efficiency. It is more suitable for relatively fixed computing types and low computing accuracy, but the amount of computing is very high. Huge AI field.

The special design brings special rewards. According to Tesla, "We have achieved GPU-level computing power in a CPU-sized figure."

Moreover, Dojo's computing power can still be superimposed, which is related to its other feature-splicing.

After a while, Tesla’s staff came up with a Dojo chip model called Plus++++++++, which is actually a chip system composed of 25 Dojo D1s.

Why did Tesla design this way?

The most intuitive benefit is to reduce the communication distance. As shown in the figure, the distance between each Dojo D1 is very close (almost to be pasted together), which can greatly accelerate the transmission and flow of data between various chips.

Tesla stated that the data transmission speed of this Dojo system can reach 9TB/S.

On the other hand, compared with traditional supercomputers, the use of splicing can also save a lot of connection cables, and it also means that the computing power of the Dojo chip system can be almost doubled, and it can more flexibly match different usage needs.

In terms of actual computing performance, the Dojo D1 chip model displayed at the event has a computing power of 362 trillion floating-point operations per second. If more Dojo D1 chips are spliced ​​together, the result can be imagined.

Of course, when it comes to chip performance, it is inevitable to compare with industry products. This time Tesla chose Google's self-developed AI chip-TPU v3. According to the display, Dojo's performance far exceeds TPU v3.

TPU v3 is a product released by Google in 2018

You must know that AlphaGo, which defeated many of the world's top chess players such as Li Shishi and Ke Jie, used only the original TPU chip that was several times weaker than TPU v3.

And if 3,000 Dojo chips are spliced ​​together, its computing power will reach 1.1 EFLOP, surpassing the Japanese supercomputer "Futake", which was previously ranked number one in the world.
Similar to Google's TPU, strong computing power often means high heat dissipation pressure. For this reason, Tesla added a full-layer water-cooled module and a copper structure to the Dojo D1 for two heat dissipation designs.

No wonder Musk dared to say before that "Dojo will be the best supercomputer in the world."

With such a powerful computing power, its application field must have reached the industrial level. Google's search results, Street View and other services rely on TPU calculation optimization, while Dojo D1 is mainly used in Tesla's visual perception system to help car identification The environment brings a better autonomous driving experience.

As for the actual effect, Musk said that Dojo will be used next year, and we will wait and see.

Towards a more complete automated assisted driving

Tesla released the FSD Beta 9.0 version in July this year . The new version of FSD supports urban road assistance, which can realize actions such as merging, turning, and merging into the main road.

The new version of FSD uses Tesla Vision, a vision system that only relies on optical images. The huge data it collects requires strong data analysis capabilities and computing power.

Dojo can receive a large amount of video data from the car terminal, send it back to the cloud, complete fully automatic large-scale algorithm training, and then push it to the car terminal to complete the closed loop of the entire process.

In this process, the most important link is undoubtedly algorithm training.

Musk said in an interview, Tesla was initially carried out training on artificial intelligence algorithms via video to.

Before owns Dojo, Tesla Autopilot team has more than 500 members tagging data, specifically for high-value data to carry out manual annotation.

Manual labeling

However, this method seriously slows down the learning progress of AI because the amount of data is too large.

As of April this year, Autopilot-based mileage has reached 4.8 billion kilometers, accounting for more than 99% of the industry's total road test data.

The "son" encountered difficulties in studying, and Tesla, the "old mother" must be anxious.

So, Dojo came.

Tesla uses Dojo to simulate a world that is very close to the real world in the cloud for training automatic assisted driving technology.

Moreover, Tesla also said that the traffic situation in this virtual world is much more complex than the real world.

Tesla joined the virtual world of a lot of extreme road conditions, you may see a moose was crossing the zebra crossing on the road, and even see a couple of morning run on high-speed roads.

A couple running in the morning on the highway

It is worth saying is that Tesla is not just for space was marked for the point in time build model, Dojo is also "the rain descends."

By constantly collecting new road information in the real world, this training model will be constantly updated, the new data will overwrite the original information.

For the body the amount of data accumulated so far, Tesla said they used one billion different images, 300 million different scenarios to build the training model.

That is why, Tesla himself dubbed the "data labeling factory."

So have these trainings achieved the desired results?

In this Tesla AI Day, Tesla showed us the tremendous progress of its autonomous driving technology.

For more accurate recognition lane

As can be seen from the above figure , Tesla's accurate identification of lane lines was limited to the surrounding area of ​​the vehicle.

But now, not only is the lane in which the vehicle, system for the entire intersection are well known, and therefore better route planning.

Predict the behavior of other vehicles

Accurately identifying vehicles parked on both sides of the road is a piece of cake for today 's Tesla, and it is its true ability to predict the behavior of other vehicles on the road.

Such as during the oncoming vehicle, the system will analyze all the possible scenarios, get a different line in the path of the driving route. Therefore, it can respond flexibly regardless of whether the other party gives up or not .

Perhaps due to the pressure of public opinion brought by the U.S. automobile safety regulators, Tesla's attitude towards automatic assisted driving on the AI ​​day is still relatively conservative, and there is no lack of emphasis on the safety of new technologies.

In addition, when talking about the significance of visual recognition for automatic assisted driving, Tesla engineers showed this picture:

Tesla perceives and maps the surrounding road conditions like a "fog of war".

Obviously, this function has touched the relevant domestic policy red line, and how it will be implemented in the country is not yet known.

Robot Tesla Bot

May also explain the relatively dull because of technology, Tesla quickly arranged for a humanoid robot came to power, also danced, the whole event site instantly happy again.

Tesla built a robot Tesla Bot that can dance?

In fact, this is just an actor wearing a robot-like costume and dancing. Musk made a very "Tesla" joke with everyone, but this robot plan is real.

Detailed attributes of Tesla Bot

According to Musk's expectation, Tesla Bot will inherit Tesla's car and machine system, including assisted driving, etc., and can make different actions according to the surrounding environment, and finally complete the manual labor that could only be done by humans.

Yes, this time Musk wants to liberate the "workforce" and let people do what they want.

Of course, Musk also said that it is still up to people to decide whether to accept the help of Tesla Bot.

It sounds like cyberpunk, but when Tesla Bot will produce energy and come to consumers, Musk only said that it will release a prototype of the product sometime next year.

How far are we from fully autonomous driving?

The International Association for Automated Engineering (SAE) divides a car's driving mode into 6 levels according to the degree of automation, from L0-L5.

And Tesla’s AutoPilot driving mode is classified as L2 and still belongs to assisted driving, which means that the car still needs to be driven under the driver’s attention, and the driver is responsible for driving behavior. This is also the current automatic The level of automation of the driver assistance system.

The well-known car critic @不只评车的38 has made an intuitive and vivid evaluation of this:

At this stage, assisted driving requires humans to be ready to take over at any time. It is more like a human-assisted car than a car-assisted human.

Musk is obviously not satisfied with assisted driving. His goal is the more automated L5-fully autonomous driving, where the car system can autonomously complete driving operations under all conditions.

According to Tesla's vision, the visual sensing system of an AI-based car should be like a human body system.

When the camera (eyes) sees the content of the picture, it judges the 3D structure and shape of the object in the picture, and calculates the distance between the vehicle and the object, the height and size of the object, etc., and the car (human driver) also automatically makes it accordingly. Corresponding driving changes to avoid collisions, this time corresponds to a stationary object;

Picture from: syncedreview

In the face of moving objects, such as a moving car, Tesla's AI sensing system should be able to recognize the height and size of the 3D object while also measuring its movement and speed.

Of course, these are just assumptions. The Dojo supercomputer can indeed accelerate AI calculations, help cars recognize objects, and bring a better assisted driving experience, but as Musk said, "Realize L5 fully autonomous driving", I am afraid it will not work. .

As early as 2019, Tesla stated that it would remove the radar and move towards a pure visual perception system, but the AI ​​that visual perception relies on is only simulating humans. At this stage, the mystery of the human brain has not been fully revealed, and AI can only make Part of the decision-making, it is not a substitute for a real human driver.

▲ AI can't replace humans

The data training completed by Tesla is far from enough, even if its data volume is leading in the new energy vehicle industry.

It is important to know that 95% of the data in the field of autonomous driving is invalid. Repeated road conditions are of little value to AI training. Most of Tesla's training data previously came from the United States. It still needs to do more in more countries and regions. Training in this situation.

The combination of extreme situations caused by different regions and different traffic conditions is almost infinite, which undoubtedly poses greater challenges for Tesla's choice of visual perception technology route.

In addition, safety accidents in recent years have aggravated people’s concerns about autonomous driving. In this case, will the number of people willing to use AutoPilot driving mode or participating in Tesla FSD (Fully Automated Driving System) tests be reduced? Put a big question mark.

Moreover, the supervision of the "assisted driving" mode of new energy vehicles by relevant agencies is also becoming more and more perfect. In addition to the US automobile safety regulatory agency mentioned at the beginning of the article, the safety investigation of Tesla's assisted driving system Autopilot has also been carried out by the Ministry of Industry and Information Technology. The "Opinions on Strengthening the Management of Intelligent Networked Automobile Manufacturers and Product Access" was issued .

Later, the Ministry of Industry and Information Technology and related departments will further refine the relevant specifications for autonomous driving and assisted driving.

In this way, it may be a bit slower to realize fully automatic driving, but in the face of life safety, it is not a good thing.

At the end of the event, Musk also expressed his vision for autonomous driving technology, which is no longer so radical:

I believe that cars in the future will certainly have autonomous driving capabilities, but do they need drivers?

Probably it is still needed, just like in the automobile age, horse-drawn carriages also exist.

This article was jointly completed by Zhou Yu and Li Hua.

#Welcome to follow Aifaner's official WeChat account: Aifaner (WeChat ID: ifanr), more exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo