During the Spring Festival this year, will the NOA function of the new car be like the three-pronged star emblem of Mercedes-Benz, and become a "homecoming New Year's goods" worth showing off?After all, NOA can not only provide a "noble experience of autonomous driving", but also "expensive".
N**iGate on Autopilot (NOA) is often referred to as "pilot-assisted driving" or "high-level intelligent driving", and is currently divided into two categories: high-speed NOA and urban NOA.
Statistics from Zosi Automobile Research Institute in the first three quarters of this year show that the penetration rate of high-speed NOA in domestic passenger cars is 67%, a year-on-year increase of 25 percentage points ;The penetration rate of NOA in the city is 48%, an increase of 2 percentage points year-on-year. It is expected that the annual penetration rate of high-speed NOA will be close to 10%, and the urban NOA will exceed 6%.
However, it is worth noting that the main driving force for the increase in penetration rate is mainly from the increase in the proportion of new energy passenger vehicle sales. According to the data of the passenger association, in the first ten months of this year, the proportion of more than 300,000 models was 142%, a year-on-year increase of 33 percentage points. The models equipped with NOA are mainly concentrated in the 30 and 300,000 yuan or more.
Why is high-end intelligent driving so "expensive"?When will it enter the mainstream production passenger car market?
The Status: A Game for the Few
First of all, we have to admit that autonomous driving, including NOA, is still in the blue ocean state in the whole industry, so the premium will be very high. Dong Xiaohang, chief consultant of Neusoft Reach's ADS Innovation Division, told Cyber Auto.
Taking the vehicle of 300,000 yuan as an example, the proportion of the NOA system in the cost of the vehicle is about 10% for high-speed NOA and 15% for urban NOA. As a result, high-end intelligent driving is currently only a game for a few people.
Maxieye, a leading supplier of intelligent driving systems, said at a recent press conference that the current price of new L2 functional cars is concentrated in the range of 15-250,000 yuan, with a carrying rate of 40%, and the terminal price difference between non-functional models is 10,000 yuanThe price of high-speed NOA models is between 25-300,000 yuan, with a loading rate of 3% and a terminal price difference of 20,000 yuanThe price of the urban NOA model is more than 300,000 yuan, the installation rate is unknown, and the terminal price difference is more than 30,000 yuan.
At present, on the whole, the cost of high-end hardware accounts for 10%, which is a very big cost pressure for car companies. But if 10% of the cost doesn't bring more than 10% of the value, it's hard to deploy at a particularly large scale. This is also why the allocation rate of new forces is higher, because it is mainly aimed at the early stage of technology.
The marketing person in charge of a **chain player told Cyber Motors: ".Today's high-end intelligent driving is more competition-oriented, some players shoot, and other players can only roll.
The "volume" inside the player is already very intense.
Generally speaking, half of the price of the high-end intelligent driving option package, that is, its cost, the vast majority of car companies are currently sold at cost, or "loss".
For example, the buyout price of AVATR ADS high-end feature package is 320,000 yuan, the current price is 180,000 yuan;Robo Drive Max of Jiyue 01 has a buyout price tag of 4990,000 yuan, the current price is 1990,000 yuan;NIO NOP+ (currently excluding the urban navigation function) monthly subscription**380 yuan, which is 56% of the full-featured NAD, but so far there is no real charge, and the rights and interests continue to be given;Xpeng, ideal high-end intelligent driving functions are free, but there is a gap between non-intelligent driving models;The Zhiji LS6, which is equipped with high-end intelligent driving hardware as standard, provides lifetime free use rights during the launch periodThe Wuling Yunduo Lingxi version with high-speed navigation and memory parking is 10,000 yuan higher than the non-intelligent driving type**.
The automotive industry has also branched out. In the second half of the year, SAIC, Changan, Geely and other mainstream car companies launched 10-150,000 yuan of new energy products, have given up in the field of intelligent cabin, intelligent driving and 150,000 yuan above the "volume", but will invest more costs in energy consumption and control performance.
SAIC Roewe's P1 coaxial motor and five-in-one PICU power domain integrated controller, Changan Qiyuan, which uses Infineon's 3-series MCU control chip and 10-layer hairpin flat wire oil-cooled motor, all devolve the advanced technology of models of 20-300,000 yuan and above to entry-level models.
On the demand side, the gap between product experience has not yet been crossed. After the experience breaks through the base point, the price of the model will definitely go down. Lv Peng, general manager of Horizon Intelligent Driving Product Planning and Marketing, said.
How does NOA eat 10% of the cost?
When will NOA become standard?
Respondents invariably said,Tesla HW3The proportion of hardware cost of 0 is the proportion of standard configuration. hw3.The cost of 0 is about 7,000 yuan, according to the model3 entry version 22 at the beginning of 2023990,000 yuan of ** calculation,The high-end intelligent driving system accounts for 3% of the vehicle price and 4% of the vehicle cost6% (calculated at cost at two-thirds of the selling price).
According to the data of the Passenger Association, in the first ten months of this year, passenger cars in the range of 5-200,000 yuan accounted for 661% (5-100,000 yuan accounted for 15.)4%, 10-150,000 yuan accounted for 339%, 15-200,000 yuan 168%），Therefore, if the median is 100,000 yuan, the standard cost of the NOA system should be 3,066 yuan, and if the median is 130,000 yuan, it will be 3,987 yuan.
This year, DJI, Qingzhou, Momo and other high-end intelligent driving companies have released a series of "thousand-yuan machines" starting from high-speed NOA, which are about yuan, reaching or approaching the most mainstream passenger cars. Unfortunately, the cost of getting on the NOA isn't just the cost of hardware.
Dong Xiaohang divided the hardware cost of NOA into direct cost, platform sharing cost and customized cost.
Among them, various hardware such as sensors and computing platforms and corresponding logistics and after-sales costs are direct costs. “These are all direct costs, but whenever the OEM buys a similar set, the related products will incur expenses. This part accounts for the lion's shareDong Xiaohang said. Lv Peng also said,At present, the proportion of hardware cost in high-end intelligent driving systems is relatively high.
Platform sharing cost refers to the pre-R&D cost of a set of high-end intelligent driving system platforms, which is apportioned according to various factors such as supporting scale, platform function level, and ecological side.
The customized cost mainly refers to the targeted investment in the implementation of the project.
Advanced intelligent driving is a complex system in which algorithms work in tandem with sensors, computing platforms, and communication systems inside and outside the vehicle. Different car companies, and even different models, have different performance requirements for electronic and electrical architectures, computing platforms, and sensor selection. When a set of intelligent driving systems that run through the laboratory are mass-produced and put on the car, they need to be engineered for these "differences", which may incur unpredictable costs.
HD maps and stockpiles are also part of the cost that is currently eating up.
At present, high-precision maps are generally charged according to the ** of 100 yuan, and a one-year subscription fee for one car. If you don't use HD maps, the hardware cost will increase accordingly.
Stockpiling is a process in which value does not outpace time. Currently, perception and computing platforms are iterating at a rate of 2-3 years. Stockpiling hardware that is geared toward the next decade may become "obsolete" before it reaches its limit, resulting in wasted vehicle costs and user spending.
The high-speed NOA began to approach the stage of good use, and the technical line began to converge. However, NOA in urban areas is still very early, and with the emergence of new technologies and new chips in the future, there will definitely be higher cost performance. Therefore, we do not recognize the behavior of stacking, and we must combine the model and the target user to choose the most suitable hardware. Lv Peng said.
In addition, Dong Xiaohang believes thatAt present, there is no healthy business logic in the industry, which is an important reason for the high cost of each link of the NOA system. "Before, everyone still had to spend more money and people to smash this thing out. This process is definitely more expensive. ”
How to reduce costs?
Despite the difficulties, "standard configuration" is the inevitable goal of the development of high-end intelligent driving. Because as data-driven has become the technical consensus in the industry, penetrating into 60% of the mainstream market is the only way to obtain data and promote high-end intelligent driving from usable to easy-to-use.
In the second half of the year, the emergence of the high-end intelligent driving system "thousand-yuan machine" has reflected the thirst of the industry. “It is really difficult to reduce costs, and since the beginning of the epidemic, the pressure (of cost reduction) has been passed on to the first-class businessmen layer by layer, and also to the second-level first-class businessmen. Dong Xiaohang said.
First, of course, scale. In Dong Xiaohang's view, the direct BOM cost of the NOA system is directly related to the system requirements, scale dimensions, technology platform and production efficiency. “It is no different from mobile phone manufacturing and home appliance industry. Platformization, volume, you can reduce costs.
From 10,000 (vehicle support) to 1 million (vehicle support), it is estimated that the overall cost reduction should be at least about 20% of the hardware cost reduction. Lv Peng said.
But 1 million units is no easy feat. China Automobile Association**, China's passenger car sales this year are 23.8 million, and according to the high-speed NOA penetration rate of 10%, it is only 2.38 million, which is distributed to various players and system platforms, and the amount is pitiful.
However, before scaling up, technological progress can provide some help to reduce hardware costs.
One is the cost per unit of computing power.
With the increasing proportion of AI models in intelligent driving algorithms, the demand for AI computing power in the system increases, while the demand for CPU rule computing power decreases. In the past few years, chip architecture design and advanced manufacturing processes have increased the density of AI computing power, and the scale caused by the increasing demand for large computing power chips in data centers and devices has driven a significant reduction in the cost per unit of AI computing power. In contrast, CPU computing power does not have such a cost reduction process.
The single-chip computing power of Horizon Journey 6 reached 560 tops, an increase of 4 compared with Journey 537 times.
The second is the cost of sensing and domain control hardware.
For example, Qingzhou's 11V solution can be "** become a 7V solution, and the fisheye lens, which is usually only used in the parking function, can be reused on the driving to reduce costs." In addition, the same single-journey 5 scheme, the high-speed NOA that landed in a light boat uses BEV perception technology to improve the performance of the system in dealing with truncation and occlusion scenarios. At the same time,".The selection of the chip, whether the chip itself is liquid cooled or passive heat dissipated, and even the ability of the business chain will also affect the cost. Hou Cong, co-founder and CTO of Qingzhou Zhihang, said.
System platform technology and engineering capabilities are the two driving forces for NOA systems to improve performance and reduce costs.
The system platform technology not only includes the pure vision, multi-modal fusion, large model, small model, neural network, rule algorithm and other technical route selection familiar to the terminal, but also includes the efficiency of R&D organization and technological innovation, which is the embodiment of the full-stack capability of the intelligent driving business.
Among them, the selection of technical routes is both technology and art. “A lot of times, technology leaders need to have a little bit of industry intuition to choose the route, and the key is to get it rightDong Xiaohang said. Organizational efficiency is a reflection of the management mechanism and corporate culture.
Technological innovation is the key to reducing the cost of "growth" of the system.
This includes innovations in software architecture. Neusoft Reach recently released the OpenVOC open technology framework, which not only decouples software and hardware, but also realizes the decoupling of software and software, clearly cutting software modules that have been intertwined in the past such as AI, vehicle control, vehicle cloud, and big data, and promoting all technical forces to engage in their own good parts within the framework and give full play to the creativity of upper-level applications. As a result, the feature development cycle was shortened by 50 percent and more than 80 percent of engineering tasks were solved.
Including the construction of efficient data closed loop:Including rapid data collection and mining capabilities, automatic annotation capabilities, effective data classification and distribution, training methods based on variables only, rapid deployment of quantitative vehicles, and evaluation methods for whether the effect after deployment meets expectations
According to Cyber Automobile, the current leading intelligent driving players in the industry have data centers with an investment of hundreds of millions of yuan. Driven by massive real vehicle data and efficient data closed-loop, Tesla can only use 8 2 million pixel cameras and 144TOPS computing power (HW3.).0) Explore the city NOA.
The engineering capability determines how much "cost" the system platform will spend on the car.
On the one hand, it depends on the applicability, completeness, and reusability of the system platform.
The highly applicable platform has been adapted to the mainstream chips and peripheral sensors at the bottom of the construction process, and will not be subject to the different hardware selections of car companies when getting into the car
A high-integrity platform that maximizes turn-up at the first adaptation. For example, the chip itself supports 10-channel cameras, so all 10-channel cameras are turned on. If only five channels are transferred for the first time, and seven channels need to be landed, the underlying communication mechanism needs to be completely readjusted, which is basically equivalent to overturning and starting over
The highly reusable platform cloud model will be very complete, covering the needs of various functions, and only need to adopt different compilation and pruning strategies according to the needs of different scenarios when getting on the car, and only need to adjust the input and output parts to adapt to new projects.
On the other hand, it depends on the player's own experience.
Look at the player's own mass production experience. Have low, medium, and high-order functions been done?The degree of understanding of the current chip tool chain, underlying security mechanism, and algorithm support that is currently available on the market or used by oneself. “It can make your engineering less detours, and once the engineering takes a detour, it will be very costly. Dong Xiaohang said. The accumulation of projects without a platform is 2-3 times more than the accumulation of projects.
For example, Neusoft Reach, as one of the first domestic manufacturers to mass-produce intelligent driving functions on the SOA architecture, has made a lot of explorations on how to improve performance, reduce load and latency, improve the richness of atomic services, and ensure their rationality and performance. "When we reused the platform for the second SOA project, it went very well. Dong Xiaohang said.
This empirical advantage can even be refined down to specific perception hardware. Hou Cong said that the reason why Qingzhou was able to do the pre-fusion algorithm was that the experience it accumulated in the L4 stage was very important. ”The main difficulty of pre-fusion is that the spatio-temporal synchronization in the sensor should be done well, the accuracy of calibration, and the trigger mechanism to ensure that different sensors see the same object at the same time. The main thing is that the configuration on the sensor needs to be good enough, and the technology itself is not very difficult
Hou Cong said: "To configure the sensor well, you need to be very familiar with it. How does the camera work line by line, what is the time of each line, if you do motion compensation, etc. We've long since figured this out. ”
There are no shortcuts to technological advancement
From technology convergence to system cost reduction, the well-known "NOA" is actually still in the early blue ocean stage.
Dong Xiaohang believes that at present, the first-class business above the middle level, whether it is pure vision or heterogeneous integration of high-end intelligent driving solutions can be done,".But what kind of scenario is it suitable for, how many sensors are there, and what is the takeover rate?What are the judging dimensions of the system?Is the choice whether to have a function name on the configuration sheet to pass, or to surprise consumers?This is a different requirement. ”
In his opinion, since the 1999 Xiaomi mobile phone was born, the functions of smart phones have been continuously enhanced in the past ten years, and the sales have grown rapidly, but there has been no much lower than 2000 yuan**. On the other hand, the field of autonomous driving is more complex and is still being explored, and its significant cost reduction and widespread popularization are bound to be a challenging process.
In the opinion of the respondents,High-speed NOA is expected to enter a large number of models in the range of 15-200,000 yuan within one or two years. In the next three years, it is expected to achieve a cost reduction of 30-40%.
The reason for this is that the high-speed NOA is close to being easy to use in terms of performance. The improvement of software capabilities has promoted the decline in hardware demand, the convergence and homogenization of technical direction and performance requirements have led to the reduction of customization costs, and the ecological synergy has begun to form in the industry to reduce the system on a large scale.
And the urban NOA is far from convergence.
For example, respondents generally believe that a computing power of about 100 tops can achieve better high-speed NOA performance. And the urban NOA has hundreds to thousands of TOPS computing power, which is possible;In terms of sensors, sensors for strong point clouds such as lidar or 4D millimeter wave are considered to be a part of the urban scene that is difficult to cancelIn terms of algorithms, in addition to the current BEV and Transformer, whether large models such as GPT should enter the car side are also being discussed.
Therefore, in the next two years, urban NOA will still focus on models of 300,000 yuan and above for adaptation.