Surpassing Tesla's FSD has always been the goal of Chinese car companies.
As early as the end of 2020, He Xiaopeng publicly declared war on Musk on Weibo.
Starting next year, (Tesla's) autonomous driving in China must be mentally prepared to be beaten by us, and as for the world, we will meet. 」
In April 2023, Yu Chengdong used data to crush Tesla at the press conference.
Huawei ADS20 has reached 98 in terms of ramp merge success rate86%, far ahead in the industry, Tesla is 8816. Of course, don't laugh at Tesla, because Tesla is much better than other brands in the world. 」
After several years of verbal warfare, Tesla's FSD has finally accelerated its entry into China.
On November 20, Tesla's official website in China was updated to include FSD beta content in the Chinese owner's manual.
On the evening of November 23, Tesla China responded to the rumors of FSD entering China: It is indeed being promoted.
Tesla's soul is finally coming, can Chinese car companies hold it?
FSD enters China, and Tesla is prepared
Along with the news of FSD's entry into China, there is also the news that the FSD V12 version (end-to-end autonomous driving) will be released soon.
Combined with these two pieces of news, it is not difficult to see that Tesla's FSD entry into China is obviously prepared.
In August 2021, Tesla unveiled its Transformer-based BEV at the first AI Day.
Since then, Huawei, Xpeng, NIO, Li and many other manufacturers have quickly followed suit and launched intelligent driving solutions based on BEV+Transformer.
During this period, although Tesla has a certain first-mover advantage, it has not opened a significant generational gap with the intelligent driving solutions of Chinese car companies.
Both parties basically realize neural network computing through the BEV+Transformer architecture at the sensing end, but the downstream planning and control are still dominated by regular algorithms.
Even FSD v11, which introduces more neural networks, has about 300,000 rows of C++ control**.
As long as it involves the rule algorithm, Tesla FSD will have to rewrite the control ** after entering China to adapt to China's traffic laws and road conditions, and even rebuild a team of autonomous driving algorithms in China.
Under the condition of adapting to the water and soil, Tesla FSD will most likely fail on the battlefield.
However, the upcoming launch of FSD V12 reverses this situation, and through end-to-end autonomous driving, it has opened up a generational gap with Chinese automakers.
Each major iteration of FSD essentially hands over more subtasks to the neural network, reducing the degree of participation in human rules.
FSD V12 realizes that almost all subtasks are completed by neural networks, and the manually programmed C++ control** is reduced from 300,000 lines in V11 to 3,000 lines, realizing complete end-to-end autonomous driving, that is, after inputting images in the neural network model, control instructions such as steering, acceleration, and braking will be directly output, and no rules are required during the period**.
The core advantage of end-to-end autonomous driving is that it can greatly improve the performance ceiling and training efficiency of the model.
In traditional autonomous driving, the upper limit of the performance of the model depends on the quality of the rules, and the gap in the intelligent driving ability of each car company is essentially the gap in the rules.
However, finite ** cannot cover infinite scenarios in any case, and engineers can only keep patching the model after encountering edge scenes.
The so-called intelligent driving, in fact, has no intelligence at all, and cannot understand the rules, just driving according to the rules written by humans.
FSD V12's end-to-end autonomous driving is completely different, where the upper limit of model performance is no longer determined by rules**, but by data and computing power.
The end-to-end intelligent driving model can simulate human thinking, through a large number of training to learn to drive.
The more data used for training, the stronger the computing power, the better the performance of the model, and even the concept of "emergence" in large language models may appear, that is, similar to the "enlightenment" and accumulation of human beings.
FSD realizes end-to-end autonomous driving, and only needs to train the intelligent driving model with sufficient local driving data to achieve a better intelligent driving experience.
At the same time, thanks to the greatly reduced number of rules**, Tesla does not need to build a team of thousands of intelligent driving algorithms in China, only needs a local operation team of about 20 people and a data annotation team of hundreds of people to achieve the smooth implementation of FSD.
To catch up with Tesla, Chinese car companies will be at least 2 years
It will not be easy for Chinese automakers to keep up with Tesla's FSD V12.
From the perspective of FSD's own development experience, it took Tesla two and a half years from the large-scale model to the realization of end-to-end autonomous driving.
In September 2021, Tesla released the BEV+Transformer architecture, and large models began to be on the car.
In September 2022, Tesla's FSD V11 adopted occupancy technology to further improve 3D spatial recognition.
In early 2023, Tesla began officially training end-to-end models.
In early 2024, FSD v12 is expected to be officially launched.
However, at present, Chinese manufacturers, such as Huawei, Xpeng, Li Enterprise, NIO and other companies, generally only realize the BEV+Transformer architecture in 2023, which is nearly two years behind Tesla.
Jiyue, which focuses on pure vision solutions like Tesla, has only just implemented the BEV+Transformer + Occupancy architecture on the just-released Jiyue 01, which is one year behind Tesla.
As a catch-up, Chinese automakers may be able to shorten the time for technology research and development.
However, if you want to surpass Tesla, you must not ignore its advantages in data and computing power - after all, it is data and computing power that determine the upper limit of end-to-end model performance.
In terms of data, at the beginning of the FSD v12 version of training, Tesla fed about 10 million Tesla owners with driving** clips, and this is far from the upper limit of Tesla.
Musk's Biography mentions that Tesla has access to 160 billion frames** per day from car owners for FSD training. To date, Tesla's FSD has accumulated more than 500 million miles and Autopilot has used more than 9 billion miles.
Behind the huge data is Tesla's global sales of 4.5 million. In the Chinese market, Tesla also currently has 1.6 million sales, which is enough for model training.
On the other hand, Chinese car companies, Wei Xiaoli's cumulative sales are 430,000, 380,000 and 580,000 respectively, and Huawei's Wenjie and AVATR have cumulative sales of 160,000, which is a big gap with Tesla in terms of data scale.
BYD, which surpasses Tesla in sales, is not currently taking intelligent driving as its core business development.
Due to the obvious scale effect of the automotive industry, it is easier to have the Matthew effect of "the strong will always be strong", and it takes a lot of time to surpass Tesla in terms of sales volume and data scale.
In terms of computing power, catching up with Tesla is also not an easy task.
As early as 2022, the computing power of Tesla's computing center has reached 2eflops.
In August 2023, Tesla launched a computing power cluster consisting of 10,000 NVIDIA H100GPUs, which can provide 10EFLOPS of computing power.
At the same time, Tesla's self-developed Dojo Supercomputing Center also began mass production in July this year, and it is expected to invest $1 billion by the end of 2024, when Dojo's computing power will reach 100EFLOPS.
Among domestic manufacturers, Huawei currently has the highest computing power of 28eflops。
It is followed by Geely with a computing power of 810 pflops (1eflops=1000 pflops), and Li Auto, Momo Zhixing and Xpeng Motors have a computing power of 750 pflops, 670 pflops and 600 pflops respectively.
Even Huawei, which has the highest computing power, has a large computing power gap with Tesla, and it requires a lot of capital investment to catch up with Tesla.
It is understood that the official price of an H100 chip is 350,000 US dollars, which is even speculated to 30-400,000 yuan on the black market, and 10,000 H100 chips will cost a minimum of 2.5 billion yuan.
Combined with the $1 billion invested in Dojo computers, Tesla's investment in computing power will reach 10 billion yuan this year and next, while Wei Xiaoli's R&D investment in 2022 will only be 10.8 billion, 5.2 billion and 6.8 billion.
For Chinese car companies, if they want to catch up with Tesla in the field of intelligent driving, technology, data (sales), computing power and capital are indispensable.
But judging from the current situation, no car company can meet all the requirements.
Catfish Tesla, or will lead to another wave of price reductions
Although Tesla FSD V12 has advantages in terms of technology, whether it can occupy the market depends on **.
For policy reasons, the data collected by Tesla in China must remain in China, which means that if Tesla wants to train FSD in China with similar capabilities to the U.S. version, it will have to build data centers and Dojo-like supercomputing centers in China.
According to Tesla's rhythm in the United States, it will take at least a year and 10 billion yuan for FSD to land in China.
The landing time of more than a year has undoubtedly given Chinese car companies a chance to breathe and even overtake, and the 10 billion repeated construction has significantly increased the cost of Tesla's FSD.
Up to now, Tesla's cumulative sales in China are 1.6 million units, and if these 1.6 million vehicles are purchased with FSD, Tesla's intelligent driving research and development cost per car will only be 6,250 yuan.
But apparently getting every Tesla owner to buy FSD is an impossible thing to do.
In North America, Tesla FSD penetration is 5%-7% on the Model 3 and 12%-13% on the Model Y, according to CITIC**.
If the penetration rate in China is calculated at a rate of 10% similar to that of North America, then the cost of intelligent driving research and development per vehicle is 6250,000 yuan, which is less than the FSD in North America 1$20,000 (8.)40,000 RMB).
However, compared to Chinese car companies, each car is 6The cost of 250,000 yuan is slightly expensive.
For example, the M5 intelligent driving version is 30,000 yuan more expensive than the standard version in terms of hardware, and Huawei ADS 20 Smart Driving Pack is priced at 1 for a limited time80,000 yuan (original price 3.)60,000), the cumulative cost of intelligent driving is 480,000 yuan;
The hardware cost of the Zhijie S7 intelligent driving version is 40,000 yuan, and Huawei ADS 20 Smart Driving Pack is priced at 1 for a limited time80,000 yuan (original price 3.)60,000), the cumulative cost of intelligent driving is 580,000 yuan;
The Max version of the Xpeng G6 is 20,000 yuan more expensive than the Pro version, and the XNGP is free, with a cumulative cost of 20,000 yuan
The Max version of the Ideal L7 is 40,000 yuan more expensive than the Pro version, and the city NOA is free, with a cumulative cost of 40,000 yuan
and Tesla, which also uses a pure visual solution, the original price of the Jiyue 01 intelligent driving package is 490,000, during the listing period** 1990,000.
At present, on Tesla's official website in China, the price of FSD is 640,000 yuan, in the current hot war, Tesla's FSD's competitiveness is very limited.
In the North American market, Tesla** sells FSD's core secret to its own commercial insurance system.
The premium of Tesla's own insurance is directly linked to the dangerous degree of the owner's driving behavior, and the safer the driving behavior, the lower the premium.
In Tesla's rating system, using FSD is the safest driving behavior. By using FSD, car owners can save more than $5,000 in premiums per year, which is much higher than FSD's $2,388 one-year subscription**.
However, at present, Tesla's annual premium in the Chinese market is about 7,000 yuan, which is lower than the FSD's annual premium of 170,000 subscriptions**. If Tesla doesn't reduce its FSD subscriptions**, even if it moves its North American insurance model to China, it won't be able to boost FSD sales.
The price reduction is the best choice for Tesla FSD to enter the Chinese market.
Since Tesla's FSD adopts a pure vision solution, the main cost of intelligent driving is in research and development, and with the increase in FSD sales, the average cost per vehicle will also decrease.
Therefore, Tesla can reverse the operation and increase the penetration rate of FSD by reducing prices, thereby reducing costs.
In the first half of 2023, the penetration rate of L2 autonomous driving in China's auto market has reached 40%, and if Tesla's FSD penetration rate can reach this level, the R&D cost of intelligent driving of a single car can drop to 250,000 yuan.
2.50,000** is lower than the average level of intelligent driving in the current market, and it is likely that it will continue to increase the penetration rate of FSD, thereby further reducing the cost of intelligent driving for Tesla's single car.
Reducing ** - increasing penetration - diluting costs - reducing ** again, Tesla's series of price cuts on cars caused by the ** war, may be in the field of intelligent driving again.
For Chinese manufacturers using lidar, it is not an easy task to reduce the cost of intelligent driving.
Although the first level of a single lidar has been reduced from hundreds of thousands of yuan a few years ago to a few thousand yuan now.
However, considering the diminishing marginal effect, the sales scale required to further reduce the cost of lidar will be much higher than before, and it is destined to be difficult to achieve in the short term.
If the unit price of lidar cannot be reduced, car companies can only start by reducing the number of lidars.
For example, Huawei's intelligent driving solution starts with ADS10 to ads20, the number of lidars decreased from 3 to 1.
In the short term, reducing the number of lidars can indeed quickly reduce the cost of intelligent driving.
However, in the long run, the cost of Tesla's FSD will continue to decline as sales increase.
However, the cost of single-vehicle intelligent driving of the LiDAR solution is always linked to the cost of hardware, and the potential for hardware cost reduction is far less than that of software.
Once Tesla FSD is the first to achieve scale in China, it will have both technical and ** advantages.
Therefore, the window period for Tesla to establish data centers and supercomputing centers in China is a life-and-death moment for Chinese car companies.
Although no Chinese car company can meet the requirements of technology, data (sales), computing power and capital at the same time, each car company has its own advantages.
In terms of technology, Jiyue has implemented the BEV+Transformer+OCC architecture, which has the smallest gap with Tesla's technical route.
In terms of data, BYD achieved sales of 5 million new energy vehicles in August this year, surpassing Tesla.
At the same time, the development of generative AI has also alleviated the problem of insufficient data for car companies to a certain extent, although generative AI may not be comparable to Tesla in terms of data volume, but its advantage is the ability to generate specific driving scenarios with low computing power, so as to train models in a targeted manner.
In terms of computing power and funds, Ren Zhengfei said that Huawei invests 3 billion to 5 billion US dollars every year in basic theoretical research, and the foundation of the upcoming fourth industrial revolution is large computing power, and Huawei will be committed to building a solid computing power foundation in China and building a second choice for the world.
From another point of view, FSD is like another catfish for Tesla, accelerating the promotion of Chinese car companies to make up for their shortcomings in intelligent driving.