Visual China.
Text |Tram pass.On January 1, He Xiaopeng announced that XNGP has covered 243 cities across the country, and on February 2, Hongmeng Zhixing announced that the intelligent driving system of the Wenjie model has been upgraded, and the high-end intelligent driving has covered 99% of the country's road sections, which can be used in large cities to small villages.
During the Spring Festival, I experienced the intelligent driving function of Xpeng G6, and the available road sections almost did not need to be taken over, and the lane markings and traffic lights could be correctly identified. Intelligent driving, which can greatly improve the travel experience, is growing much faster than we imagined.
The highway conditions are simple, and car companies such as Wenjie and Xiaopeng have achieved basic coverage, and the focus of the next competition of car companies will be the pilot assisted driving in urban areas. How car companies are preparing to deal with the next competition, there have been a lot of rumors in the industry.
Balancing the level and cost of intelligent driving is an issue that car companies and autonomous driving companies need to be cautious about, such as Xpeng, which plans to continue to reduce lidar in the future due to cost considerations. However, at this stage, it is still difficult for high-end intelligent driving to leave lidar.
The M9, released on December 26 last year, is equipped with the industry's first mass-produced 192-line lidar, with a recognition distance of up to 250 meters, an imaging capability of 1.84 million points and seconds, and a vertical resolution of 01°, the radar scanning frequency of 20Hz, compared with the mainstream 96-line 156-line lidar on the market, it can be called far ahead.
However, shortly after the official announcement of Huawei's 192-line lidar, Beiwake Photon officially announced that its 256-line lidar was put into production, and at the CES2024 conference, the lidar giant Hesai Technology launched a 512-line lidar - AT512, which has a detection distance of up to 300 meters, a maximum range of 400 meters, and an imaging capability of 128880,000 points per second.
The more lidar beams, the richer the details of the captured objects, the more complete the modeling outline, the clearer the point cloud map will be, and it has higher redundancy and anti-interference capabilities, which can improve the safety of intelligent driving. However, the wiring harness is too high, followed by the rise in cost and power consumption, I am afraid that only a luxury car of about 500,000 yuan like the M9 can use the lidar with 192 lines and above wiring harness. Considering that the power consumption of vehicle-mounted lidar is generally within 50W, the impact on battery life will not be obvious.
In terms of volume hardware, Chinese companies have rich experience, and at this rate of development, it is believed that in a few years, thousands of beams of lidar will be available.
Zhijia also has certain requirements for chips, and the domestic horizon company will release the Journey 6 chip in April, with a computing power of up to 560TOPS, and the computing power of the THOR chip released by NVIDIA has reached 2000TOPS. However, compared with the parameter improvement of other hardware such as radar, it is not difficult to improve the computing power.
Lidar + high computing power chip can undoubtedly improve the safety of intelligent driving, but the high cost is not willing to accept by every car company. Of course, even Tesla needs to be based on high-computing power chips, and some domestic companies have chosen a more extreme route, with low-computing power chips + cameras + a small amount of radar, which can achieve L2+ level autonomous driving.
The intelligent driving solution jointly launched by SenseTime and Nezha Automobile only requires 16TOPS computing power chip and 5R11V (5 millimeter-wave radars and 11 cameras) sensors to realize pilot assisted driving in high-speed and urban road sections.
The solution launched by DJI Automotive only requires a minimum of 32TOPS computing power chips, with 7 9 cameras, and can achieve regional memory driving without any radar. If you want to achieve L2+ level pilot assisted driving, you can upgrade the chip to 80TOPS, and the cost of this set of hardware is not as good as a lidar.
Automotive intelligent driving hardware has shown two major directions, high-end models pursue hardware performance improvement, cost is a secondary consideration, and the number of lidar wiring harnesses and chip computing power are the main competitive points. Mid- and low-end models pursue the ultimate squeeze of hardware potential and low cost in order to benefit consumers with lower car purchase budgets, which inevitably puts forward higher requirements for software algorithms.
Since the second half of 2023, many car companies have announced that they will move AI models and generative AI to cars, in order to link cars with intelligent assistants and productivity. For example, BYD's Xuanji architecture is divided into two parts: cloud AI and vehicle-side AI, and the vehicle-side AI model does not need to upload data, has better confidentiality, and can execute the instructions issued by users without a network.
Compared with AI large models and generative AI, the prospect of large models in the field of intelligent driving can be seen with the naked eye. The scene is discretized, divided into hundreds of thousands of small scenes, and matched with the actual driving environment, and the system can make intelligent driving deductions and judgments based on similar scenarios.
Autonomous driving companies and car companies are trying to deploy intelligent driving models on the vehicle side, but even if the intelligent driving model with hundreds of billions of parameters is reduced to tens of billions of parameters, due to the computing power of the on-board chip, the temporary response speed is not as fast as the cloud large model. Stacked chips can solve the problem of computing power, and replace it with the problem of cost.
One of the main advantages of the large model is that it can reduce the cost of intelligent driving ecological construction. As of November 2023, the length of domestic roads has reached 5.35 million kilometers, and it is still changing, and the cost of all high-precision mapping is unimaginable. At present, although we have gotten rid of the dependence on high-precision maps, we still need lightweight maps or high-precision maps.
Huawei has broken this tradition by not opening one city at a time, but directly covering the whole country. Huawei's approach is similar to that of MomozhixingThe intelligent driving model collects enough data and discretizes it to become the basis for road topology reasoning, which can be combined with ordinary navigation maps to realize pilot assisted driving.
Car companies are getting rid of the dependence on high-precision maps, but they still need lightweight maps, which leads to the lack of coverage speed of pilot assisted driving. In addition, domestic highways have been changing, and some areas may need to be drawn repeatedly, and backward areas such as small towns may not be able to wait for the mapping team of car companies in 2030.
Through large models, discrete scenarios, and road reasoning, pilot-assisted driving and even autonomous driving can be realized, which requires higher computing power and data volume, and the overall cost is lower. With Huawei as a precedent, we may be able to see other car companies adopt the same solution to achieve nationwide intelligent driving in 2024.
In recent years, we can see that the relevant departments are promoting the commercial use of autonomous driving in an orderly manner.
On February 8, 2024, the country's first intelligent vehicle intelligent driving performance data open platform was unveiled in Suzhou. The platform is built by Chek, the purpose of which is to solve the problem of information asymmetry of smart cars, and many car companies such as Wenjie, Xiaopeng, and Weilai have joined the platform, and will disclose data such as the number of intelligent driving takeovers and obstacle avoidance capabilities through the platform.
The establishment of the platform, on the one hand, can reflect the shortcomings of car companies, point out the direction of technical optimization for car companies, on the other hand, it can be used as a reference for consumers when buying a car.
The continuous research and development and upgrading of software and hardware, and the continuous reduction of costs, have made models within 150,000 yuan see the dawn of popularizing high-end intelligent driving. We can see the data on the platform of high-end cars and low-cost models, and if the frequency of takeover of high-end models is higher than that of low-cost cars, I am afraid it will have a certain impact on the reputation of car companies. However, at present, the cost of intelligent driving hardware is high, and we cannot see too many results in a short period of time, and it is difficult to make comparisons.
High-speed NOA is relatively mature, and it is possible to achieve full coverage this year, while urban NOA is the focus of competition among car companies this year. The application of higher beam lidar and higher computing power intelligent driving chips can strengthen the intelligent driving ability of automobiles. ZEEKR 001 facelift was revealed or will become the first model officially equipped with NVIDIA Thor, Beixing Photonics 256-line lidar has been put into production, I believe that the first car equipped with this product is not far from the market.
The addition of large models can greatly improve the efficiency of intelligent driving training, accelerate the improvement of intelligent driving algorithms, and make the driving logic of the intelligent driving system more like that of old drivers.
He Xiaopeng once said that 2024 is the first year of autonomous driving. Judging from the actions of car companies, autonomous driving companies, and ** chain companies, the smell of gunpowder in this first year is particularly sufficient.