Tesla, Inc. AI Day opens at 5: 00 p.m. on August 19, Western time.

After a few minutes of brief opening remarks, musk began to look for someone off the stage, "Andrej."

Soon, Andrej Karpathy, senior director of the AI department of young Tesla, Inc., appeared on the stage with a smile and officially opened the curtain of AI Day.

On this day, a number of Tesla, Inc. 's AI business executives appeared on the stage one after another, in addition to Andrej, including:

• Director of Ashok Elluswamy,Autopilot Software
  

• Senior Director of Ganesh Venkataramanan,Autopilot hardware, head of Porject Dojo
  

• Milan Kovac,Autopilot Engineering Director.

This AI team focused on showing Tesla, Inc. inAutopilot awareness algorithm、Regulation control algorithm、analogue simulation、Cloud training、Self-developed chip、Supercomputing platformAnd other "treasure" achievements in many fields.

This time, Tesla, Inc. AI Day has three core main lines:

• The Progress of Autopilot FSD algorithm and Neural Network training
  

• The hypercomputing system Dojo and the training chip D1 developed by Tesla, Inc.
  

• Humanoid robot Tesla Bot.

After Autonomy Day in 2019 and Battery Day in 2020, Tesla, Inc. once again shows us that Tesla, Inc. is a top AI company and even the largest robot company in the world, in addition to car companies, self-driving software companies, battery companies and energy companies.

Becoming the top seller of new energy vehicles in the world and the global market capitalization.firstAfter the car company, the AI team led by Andrej Karpathy and others seems to show us that Tesla, Inc. is still far from his own border.

While competing with luxury car companies such as BBA, Tesla, Inc. will also compete with top technology companies such as Alphabet and NVIDIA in the artificial intelligence battlefield.

1 Perception and decision-making: let cars have themselves

Perception, planning, decision-making and execution, Tesla, Inc. has wanted cars to drive on their own since 2013.

In order to achieve autopilot, Tesla, Inc. 's basic logic is to build pure visual perception technology similar to human beings, to build computing platforms close to human computing power (on-board FSD chips and cloud supercomputing platforms), and to train the brain with a large amount of data to make cars more and more good at driving-making cars ego.

The primary prerequisite for the realization of autopilot is perception and decision-making, which was explained by Andrej at the beginning of Tesla, Inc. AI Day.

Every Tesla, Inc. is carrying.8 颗Cameras, the data they collect can be formed.3D vector space(Vector Space), if you add the time stamp, it is what Tesla, Inc. called4D autopilot system。

Andrej describes the use of deep learning for perceptual recognition, like the multitasking model Hydra, which involves camera detection, cross-camera fusion and so on.

With the help of multiple cameras and neural networks, Tesla, Inc. can draw maps (SLAM) in real time through the algorithm.

In recent years, the accuracy of Tesla, Inc. 's use of camera for object detection and depth estimation has exceeded that of camera + millimeter wave radar, so Autopilot has removed millimeter wave radar from production vehicles.

The next step after perception is to do a good job of planning and control.

Ashok Elluswamy, another director of self-driving software, took the stage.

According to Ashok, there are two major challenges in planning:Non-convex surface(Non-Convex) andHigh dimension(High-Dimensional).

To put it simply, when a vehicle is driving on the road, its behavior and path are not the only optimal solution, but there are many different solutions, and it is highly dynamic.

To take an example of lane change, there are usually two kinds of paths, one is to try to change lanes early and slowly, but not very comfortable, and the other is to change lanes later and quickly, but there is a risk of missing the lane change.

The best solution is to maintain a balance of comfort and safety when changing lanes. When the vehicle is planned, not only the vehicle itself, but also each road traffic participant is taken into account.

Next, Tesla, Inc. showed a wonderful video.

In a narrow street, there are no lanes, all kinds of vehicles are parked on both sides of the road, and a white car suddenly drives out of the opposite direction of the narrow road. from the main point of view, we can see that Tesla, Inc. obviously dodged to the right, just like a human driver trying to make way for the opposite car.

However, the white car politely pulled over to make room for Tesla, Inc.. Tesla, Inc. continued to drive forward.

Ashok reveals how Tesla, Inc. FSD "thinks" behind the video.

When the opposite car was first encountered, the possibility that the white car would continue to bypass other vehicles was predicted to be a high probability event, while parking was a low probability event.

Therefore, Tesla, Inc. chose to pull over and make way.

However, the car has always chosen a small probability of parking, Tesla, Inc. obtained and understood this behavior, and then chose to move on.

Make a brief summary:

The pure visual perception system is responsible for inputting data and formingVector space(Vector Space) andMiddle layer characteristics(Intermediate Features), and then output it to the neural network, and then go further throughTrajectory distribution model(Trajectory Distribution) output toPlanning and control system(Planing & Control), which finally outputs steering or acceleration instructions to the vehicle.

2 How does Tesla, Inc. digest huge amounts of data?

When self-driving vehicles hit the road, a large amount of data will be generated continuously.

How to analyze and process these data and train the algorithm?

Tesla, Inc. is probably the largest data processing company in this field.

Tesla, Inc. set upManual data annotation、Automatic data annotation、Emulation和Data scale(Scaling data geeneration) four teams.

At present, Tesla, Inc. has set up a manual data annotation team of 1000 people.

Andrej disclosed that four years ago, Tesla, Inc. was also an outsourced third-party tagging team like most companies in the industry, but later found that the "delay" was too high, so he turned to the self-built tagging team.

With the sharp increase in the amount of data, Tesla, Inc. introduced automatic tagging. Because Tesla, Inc. uses the perception of multi-camera fusion, its automatic labeling tool can realize the simultaneous annotation of multi-view and multi-frame pictures of all cameras at one time.

At the CVPR 2021 event held in June this year, Andrej disclosed that Tesla, Inc. had 1 million video data of about 10 seconds, and marked the60 亿The distance, acceleration and velocity information of each object, and the total amount of data is up to1.5PB。

As for simulation, Tesla, Inc. 's "simulation scene technology" can simulate "edge scenes" that are less common in reality, such as when a family is running on the highway, and use it for autopilot training.

In the simulation scene, Tesla, Inc. engineer can provide different environment and other parameters (obstacles, collision, comfort, etc.), which can greatly improve the training efficiency. At present, Tesla, Inc. has built more than 200 miles of roads in the simulator.

How can the huge data produced by collection and simulation be used for efficient training?

Tesla, Inc. 's plan isSelf-developed supercomputer Dojo。

3 Self-developed Dojo chip, won the crown of the highest computing power

As early as a few years ago, under the instruction of musk, Tesla, Inc. internally began to develop an energy-efficient and high-performance supercomputer Dojo for AI training.

By April 2019, musk first announced its Dojo R & D plan on Tesla, Inc. Autonomy Day, and has been recruiting talent for Dojo ever since.

The top killer on Tesla, Inc. 's AI Day this year.Should belong to Dojo。

According to Ganesh Venkataramanan, senior director of hardware of Tesla, Inc. Autopilot and head of the Dojo project, Dojo has three major R & D goals:

• Achieve the best AI training effect

• Can accommodate larger, larger hybrid neural network models

• Power consumption and cost optimization to achieve higher performance-to-price ratio.

Tesla, Inc. believes that it is easy to improve the computing power in the research and development of supercomputers, but it is very difficult to expand the bandwidth and control the delay.

Based on this, Tesla, Inc. began to design from the chip level.

It is not easy to develop Dojo from the chip.

The chip industry chain includes three major links: design, manufacturing and closed testing. how to solve the problems of energy efficiency and cooling is full of challenges in each field.

D1 adopts distributed architecture, 7 nm technology, each D1 chip carries 50 billion transistors and 354 training nodes, and the computing power of single chip BF16 is as high as362 TFLOPs, but the power consumption is only400WIt has both GPU level training ability and CPU level controllability.

In the closed test phase, the Dojo chip usesInFO_SoW technologyThe original WSE chip and Taiwan Semiconductor Manufacturing Co Ltd WLSI platform are all packaged by InFO_SoW.

InFo_SoW, also known asOn-wafer system"the technology is to design all the chips on the same wafer and turn the whole wafer into a super-large chip.

The advantage of this is that communication can be achieved with low latency and high bandwidth.

To put it simply, because the physical distance between the chip and the chip is very short, and the communication structure can be arranged directly on the wafer, all cores can use a unified2D mesh structureInterconnection, realizing the communication between chips.Ultra-low latency和High bandwidth; and because of the structural advantages, the lowerPDN impedanceTo achieve the improvement of energy efficiency.

In addition, because it is composed of multiple small chips in the array, the problem of "good product rate" can be avoided through redundant design, and the flexibility of chip processing can be realized.

Finally, a high-power chip D1, which can be used for AI training, was unveiled.

Based on the D1 chip, Tesla, Inc. integrates the self-developed chip into a training module, which is divided into seven layers:

• The first and fifth copper structures are water-cooled heat dissipation modules.
  

• The second layer structure is composed of 25 chips in the 5'5 array.
  

• The third layer is BGA package substrate with 25 array cores.
  

• The fourth and seventh layers are physical load-bearing structures with some thermal conductivity properties.
  

• The sixth layer is the power module and the black strip standing on it, passing through the interconnection module that radiates heat and communicates with the chip at a high speed.

Tesla, Inc. made 25 D1 chips into one. "Training moduleAbout 60 training modules, that is, 1500 D1 chipsMore than 530000The training node forms the Dojo supercomputer.

In theory, the performance expansion of Dojo has no upper limit and can be expanded indefinitely.

In practical application, Tesla, Inc. assembled 120 training modules intoExaPOD supercomputerExaPOD contains 3000 D1 chips and more than 1 million training nodes, and the computing power can reach1.1 EFLOP。

The score of 1.1 EFLOP gave Dojo a direct concession.The fastest in historyThe throne of AI training computers.

Next, the next generation of Dojo chips will be improved.10 倍Performance.

If the FSD chip solves the problem of the car side, then what Dojo wants to solve is the problem of the cloud.

Musk revealed that Dojo will be available next year.

With the debut of D1 chip, Dojo will walk quickly on the road of AI training, and Tesla, Inc. 's FSD autopilot software capability will also be greatly improved.

4 Robot? The Future of Tesla, Inc.

The D1 chip will not stop at Dojo, and Tesla, Inc. may enter the robot track based on D1 in the future.

On AI Day, Musk made it public for the first time: Tesla, Inc. wants to doHumanoid robot Tesla BotThe robot, which is 1.72m tall and weighs 56.6kg, will carry Tesla, Inc. 's advanced technologies, such as Autopilot cameras, FSD and Dojo artificial intelligence algorithms, and can be linked with Tesla, Inc. 's car.

According to the plan, Tesla, Inc. may launch the first robot prototype next year.

Today, Tesla, Inc. 's boundary is constantly expanding.

First of all, it must be a car company. Tesla, Inc. delivered the whole car in the second quarter of this year.201304 辆It hit an all-time quarterly high.

Tesla, Inc. expects his car delivery to grow by more than 50% year-on-year in 2021, to about 750000, up from 499500 last year. This sales volume is still at the top of the global list of new energy vehicles.

Tesla, Inc. has been branded as a technology company since he began developing autopilot in 2013.

It can be predicted that Tesla, Inc. 's next stage will be a genius game that makes efficient use of data, supercomputers, deep learning and other elements.

Regardless of whether robots can be commercialized, the first result of this game is to achieve fully autopilot one day earlier.

On the one hand, Tesla, Inc., as a car company, is becoming more and more scientific and technological, and technology companies are also stepping into the ranks of car makers, such as Baidu, Inc. in China.

Just two days ago, Tesla, Inc. AI Day, Baidu, Inc. also released a self-developed chip at his World Congress, which is also based on7nmKunlun II, which is mass produced by process.

The peak calculation power of Kunlun 2 is256 TopsIt is said that it may be used on a mass-produced Jidu car in 2023.

This year Tesla, Inc. AI Day does not have the technology that can be realized immediately, Dojo will be put into use next year, and Tesla Bot does not have a clear time for mass production.

Interestingly, Tesla, Inc. FSD was first mass-produced and released in 2019, but on the day of AI Day, there was no mention of FSD second-generation chips.

At a normal pace, we guess that Tesla, Inc. should release it next year.Next generation FSD chip了。

With the emergence of Dojo, Tesla, Inc. FSD will have a leap from quantitative change to qualitative change, and what changes will take place in the second-generation FSD chip is also very much expected.