Wen 丨深途 (ID: shentucar), author 丨 liming, editor 丨 Wei Jia
In the spring of 2017, at the American Circuit in Texas, USA, an electric supercar called EP9 ran a score of 257 kilometers per hour in the driverless test, becoming the fastest driverless car in the world at that time.
The manufacturer of this car is a company from China, NIO. The news is a sensation in the racing circle, and it has also greatly encouraged domestic car companies.
In that era of PPT car manufacturing, many players' cars have not yet been built, and Weilai has taken the lead in getting tickets for automatic driving.
However, self-driving is very expensive, and it is not easy to feed a team of more than 300 people in China and the United States. In the serious financial crisis in 2019, the loss of personnel in the autopilot team of Weilai was serious, and the independent research and development pressed the pause button.
In the process, Xiaopeng Automobile, another new car-making force, suddenly caught up. With a clear positioning and aggressive strategy, Xiaopeng has created multiple firsts in autonomous driving. For example, the first landing with NVIDIA Xavier chip, the first lidar model to achieve mass production, and so on.
In March last year, He Xiaopeng, chairman of Xiaopeng Motors, said that it is necessary to be the first in the world for autonomous driving.
Two months after He Xiaopeng released these words, ideal car's only mass-produced model ideal one, finally ushered in a mid-term redesign. Ideal ONE added a chip, four millimeter-wave radars in the redesign, and the overall upgrade of autonomous driving hardware. Then, at the end of last year, Ideal officially pushed the NOA high-speed pilot auxiliary function.
Although it is later than Weilai and Xiaopeng, and the function is a little less interesting, the ideal is not absent.
In the five years from 2017 to the present, the appearance of the cars we have seen has not changed substantially, but the internal structure and function have long been very different. For example, the power source of the car is changing from oil to electricity, and for example, the amazing driverless five years ago is getting closer and closer to the real landing.
He Xiaopeng said that in the future, the core of judging whether a car is a smart car is whether the car has automatic driving functions.
Around automatic driving, China's new car-making forces have set off a vigorous scramble for positions. In 2022, the three companies of Weilai, Xiaopeng and Ideal will meet head-on in this arena. In addition to launching a smart car with the highest matching in history, they will also carry out a full range of PK in hardware, software, data, team and other dimensions.
Wei Xiaoli's autopilot competition is coming.
The hardware arms race is not just a race against time
On the road of autonomous driving, among the three companies of Wei Xiaoli, the earliest to emerge is Weilai.
Delivered in June 2018, the NIO ES8 was the earliest mass-produced model among the new forces and was equipped with the most advanced ADAS (Advanced Driver Assistance System) in Japan at the time.
The most core component of this system is an autonomous driving chip - EyeQ4.
Camera components of mobileye EyeQ4 Source / Mobileye official website
The chip came from an Israeli company, Mobileye, which at the time accounted for about 70 percent of the global ADAS market share. The EyeQ4 equipped with the Weilai ES8 is the latest achievement of Mobileye, which is 8 times higher than the EyeQ3 used by Tesla earlier. The ES8 became the world's first mass-produced model equipped with mobile EyeQ4 chips.
At the beginning, the top hardware was used, allowing Wei to achieve automatic driving.
The importance of hardware for autonomous driving is self-evident. This is similar to mobile phones and computers, CPU and graphics card performance is not up to standard, playing large games will be stuck.
But the problem is that under the impetus of Moore's Law, the performance of chips has been improving, and the most advanced at the moment does not mean that it will be advanced in two years.
The hash rate of the Mobileye Q4 chip is 2.5 TOPS, and The three production models of Weilai, ES8, ES6 and EC6, Xiaopeng's first car, the G3, and the ideal ONE before the redesign, are all used in this chip.
However, when Xiaopeng launched its second car P7 in 2020, it was replaced by NVIDIA Xavier chip, and the computing power was directly increased from 2.5 TOPS to 30 TOPS. This allows Xiaopeng to catch up in the hardware architecture. The Xiaopeng P5, launched in 2021, also continues this architecture.
Subsequently, the ideal ONE in the mid-term 2021 redesign, replaced by two horizon "Journey 3" chips, the computing power reached 5TOPS, although not as good as NVIDIA Xavier, but compared to mobileye Q4 has also doubled.
In Wei Xiaoli, the ideal is the role of the catcher. The late launch of the car, the hardware is not optimal, so the early hardware competition, mainly occurred between Weilai and Xiaopeng.
However, Nvidia Orin chips will not be in production until 2022. Xiaopeng's latest model, the G9, and the ideal next-generation model, the X01, also chose this chip. They all step into the same river.
Starting from the computing chip, Wei Xiaoli launched a direct arms race on hardware, 2021 is the first year, and 2022 will be more intense.
Hardware-level racing also occurs on perception devices, such as various radars.
In the process of achieving automatic driving, environmental perception is the first step, and the carriers are cameras, ultrasonic sensors, millimeter-wave radar, lidar and so on. The type, quantity, and performance of these hardware used by automobiles directly affect the degree of realization of autonomous driving.
2021 can be regarded as the first year of lidar on the car. After the Xiaopeng P5 was mass-produced with lidar, more and more car companies joined the lidar camp. Weilai ET7, Xiaopeng G9, and the new model M7 released by WM this year are all equipped with lidar, and the ideal X01 may reserve an interface for lidar.
In the hardware architecture of automatic driving, the new forces have accelerated the iteration, and no one dares to relax in the slightest.
Algorithms and data, the game below the surface
If you just pile up material on the hardware, then automatic driving is too simple. Behind the hardware, there is also a need for sufficient intelligent algorithms and software support, which is the key to really pulling the gap between driving companies.
From the above, we can see that Wei Xiaoli followed the same path in the choice of autonomous driving chips - Mobileye as a transition, the next generation of models evolved to the NVIDIA Orin platform.
The reason for choosing Mobileye in the early days was that Mobileye's EyeQ series products were really easy to use, and second, there was really no better choice at that time.
But there is a problem here, Mobileye's system is completely closed, the algorithm is bundled with the chip, it is like a "black box", it is difficult for cooperative car companies to design functions according to their own needs. What's more, there is no way to master algorithms and data.
Data is too important for autonomous driving. CITIC Securities believes that automatic driving is a technology based on machine learning algorithms, data is the basis for algorithm modeling and software landing, and a large number of data collection is the premise of automatic driving technology development.
Mobileye cooperates with car companies, which is equivalent to "teaching them to fish", but there is no "teaching them to fish", and car companies themselves do not have real automatic driving capabilities. According to CHEN Hong, chairman of SAIC, "lost his soul."
In contrast, NVIDIA is more open, it provides core chips, the rest is open, and car companies can do algorithms and data customization development and implantation based on its computing platform.
Experts from Third Bridge's High-Lin Consulting said that in the Chinese market, from the demand side, NVIDIA is more suitable for car factories that already have certain algorithm capabilities, and Mobileye is relatively friendly to auto factories in the early stage of autonomous driving.
This is the fundamental reason why Wei Xiaoli gave up Mobileye after his wings grew.
But there is a cost to giving up.
Tesla's example is a lesson for the past. In 2016, Tesla abandoned mobile Q3 and switched to NVIDIA DRIVE PX2, due to the corresponding perception algorithm to be redeveloped, resulting in a decline in the use of Autopilot experience, until the algorithm optimization has improved.
The above experts believe that considering that domestic OEMs, especially new forces, are in order to differentiate between products, the mainstream trend of autonomous driving chips in the future may be more inclined to cooperate with NVIDIA.
Today, Wei Xiaoli is taking this step. The question is, are they ready?
Xiaopeng is still relatively successful. Equipped with NVIDIA Xavier chips, and on this basis, the Xiaopeng P7 and P5, the landing models that have been developed by algorithms and software, first realized high-speed NGP navigation assisted driving, and then tested urban NGP. This is the lead among the new forces in the country.
Moreover, Xiaopeng is the first player in Wei Xiaoli to publicly claim to be full-stack self-developed. CITIC Securities believes that the full-stack self-research ability is the core advantage of Xiaopeng Automobile. In the first quarter of 2021, Xiaopeng disclosed for the first time the revenue in the automotive software business - 30 million yuan, which has exceeded half of the annual revenue in 2020.
Source / Visual China
Ideal has been trying to catch up. In fact, it was not until after the IPO in 2020 that Ideal began to exert efforts in research and development and vigorously recruit talents. In September 2020, Wang Kai, former global chief architect of Visteon, parachuted into Ideal as CTO and was fully responsible for R&D work, which is an important symbol of Ideal's increased investment in R&D.
By 2021, Ideal will rely primarily on third-party hardware and software vendors to implement assisted driving capabilities. Unlike WEILai and Xiaopeng, Ideal inserted a Horizon Journey 3 platform in the process of switching from Mobileye to NVIDIA.
Some people believe that this has not undergone an essential change in the fundamental architecture, but has to spend energy on the re-adaptation of the algorithm. Ideal ONE originally planned to launch the NOA "Navigation Assisted Driving" system in September 2021, but it was actually a full 3 months later than planned.
On the next generation of models X01, the ideal narrows the gap in hardware, but whether it can make good use of the powerful computing power of Nvidia Orin chips depends on the ideal level of self-research.
The evolution of autonomous driving: upgrading and fighting monsters cannot be ended
In the path to the realization of the auxiliary driving function, from the final effect, Wei Xiaoli is roughly similar - through the upgrading of hardware, the evolution of software, and the increase in driving data brought about by the increase in sales of new cars, the iteration of algorithms and the unlocking of more functions are realized.
For example, the ideal ONE, the positioning of this car is a family car, and assisted driving performance has not become the core selling point. In the original version, it had some obvious deficiencies in automatic driving, such as the inability to achieve automatic lane change functions, which Tesla, Weilai and Xiaopeng all had.
This is because the ideal ONE is used to do forward-looking detection hardware, only 1 front-view monocular camera and 1 millimeter-wave radar, the rear of the vehicle lacks millimeter-wave radar, can not detect the long distance behind the vehicle.
Ideal ONE hardware configuration before the makeover source / CITIC Securities
This lack of hardware directly limits the space for assisted driving ability. Later, ideal ONE added four millimeter-wave radars in the mid-term rebrand, and then upgraded through OTA to solve this problem. After the hardware upgrade, the ideal ONE only pushed the NOA high-speed pilot auxiliary function to the owner at the end of the year, which kept up with the pace of Weilai and Xiaopeng.
Xiaopeng adopts the method of continuously superimposing functions from L1 and L2, and finally realizing L3 and L4. Specifically, Xiaopeng dismantled the entire automatic driving scenario into a single point of functional breakthrough, such as autonomous parking, automatic lane change, high-speed NGP, urban NGP, etc., and finally through OTA, to achieve a full closed loop of point-to-point assisted driving (from garage to garage).
At the beginning, Xiaopeng focused on the best breakthrough of the autonomous parking scene, and He Xiaopeng publicly promoted Xiaopeng's parking function more than once. In P7, the hardware and software of the vehicle have gone up to a higher level, Xiaopeng's self-developed algorithm has really exerted its power, and the high-speed NGP function is also unlocked at this time. The P5, which has the support of lidar, takes the urban NGP function a step further.
In general, Xiaopeng unlocks one function after another through one OTA after another, forming an advanced driver assistance system that is getting closer and closer to L3 and L4. The length of time this process takes depends on the comprehensive impact of hardware, software, algorithms and other dimensions mentioned above.
After the efforts of the past few years, Wei Xiaoli and these three companies have pushed their own high-speed navigation assistance driving functions to car owners, and the difference in the coverage of functions is not too big, but the experience will be different. Experts from Third Bridge's High-Level Consulting believe that in terms of automatic driving functions, high-speed navigation and urban navigation are the focus of attention of new car-making forces.
One variable is that The ideal CTO Wang Kai, who has been in the job for less than two years, may leave in the near future. Wang Kai was previously responsible for the research and development and mass production of ideal smart car related technologies. This will more or less have an impact on the development of ideal autonomous driving.
For Wei Xiaoli, exploring the process of automatic driving is a real long-distance run. There are too many uncertainties in the process, and there are too many detours, and if you are not careful, you may fall into the pit and fall behind your opponents.
But in another dimension, they are close comrades-in-arms. In the face of the tide of the times, they explore the unknown together, always fighting side by side, riding the wind and waves.
Finally, post a picture of the three big guys from He Xiaopeng's Weibo in June 2020. The caption is: Three bitter, in the memory of the change...