Millimeter-wave radar, from 3D to 4D, from sparse point cloud to dense point cloud output close to the imaging level, the next stage of change will be around the upgrade and matching of the electronic architecture of the whole vehicle.
During this year's CES exhibition, TI, the world's leading millimeter-wave radar chip manufacturer, announced that it officially launched the first generation of chip solutions (named: AWR2544) that can be used for satellite radar architecture (middle and late data processing is centralized to the domain controller).
The AWR2544 is also the industry's first single-chip solution with LOP technology. This technology reduces the size of the sensor by 30 percent by integrating a 3D waveguide antenna on the other side of the printed circuit board, i.e., a waveguide interface mounted on the package.
The AWR2544 is manufactured on a low-power 45nm RFCMOS process, enabling high levels of integration (including PLLs, VCOs, mixers, and baseband ADCs) in a small package with ultra-low BOM counts.
As we all knowAt present, most of the 4D imaging radars that have been put on the car use FPGA as the end-side big data processing, which is costly. In contrast, the radar of the satellite architecture itself does not require this additional cost processing power.
For example, due to the significant increase in the number of point clouds, satellite architectures can take full advantage of the computing power redundancy of domain controllers for advanced algorithms for object detection, classification, and tracking.
In addition, at present, some car companies are only trying to upgrade some radars in the forward or blind zone to carry out 4D imaging verification tests. At the same time, due to cost considerations, 5R and 6R configurations will not all use 4D imaging.
The use of satellite architecture, from the perspective of quantity, can help radar manufacturers achieve faster scale of front-loading on the car; For car companies, it is also a way to reduce costs and increase efficiency.
Imaging radar has attracted a lot of attention from the market. At the same time, we are seeing a strong demand for imaging radars that achieve optimal performance. Valeo thinks. Judging from the progress of the industry last year, the cost of 4D radar has dropped significantly, which is expected to drive the market to really increase.
At this year's CES exhibition, Altos Radar, a 4D radar startup from China, brought two high-performance 4D imaging millimeter-wave radars, which are only 1 10 of LiDAR.
For example, the latest product in the Altos V series, the Altos V2 (Forward 4D Radar), is the world's first 4D radar to be mass-produced with a non-FPGA solution, and the chip cascade design is completed based on TI's TDA4. According to the current market price, it is expected to reduce costs by more than 20%.
In addition, the ALTOS RF series is a non-computing front-end radar module (commonly known as a satellite-based sensor architecture), which is based on the computing resources of the domain controller for data processing, greatly reducing the hardware cost on the device side, and is expected to achieve a 50% cost reduction.
At present, there are three main solutions in the industry to achieve cost reduction.
One is to develop custom ASICs to replace FPGAs; For example, NXP launched the industry's first dedicated 16nm imaging radar processor S32R45, equipped with 4 ARM A53 cores, 3 lockstep ARM M7 cores, and 8MB SRAM + LPDDR4 + external flash memory.
The second solution is mainly to reduce cascade and localize chip solutions; For example, the two-chip cascade solution launched by Chuhang Technology, based on the traditional 3D millimeter-wave radar process design and hardware structure, adds an additional RF chip, which can greatly improve product performance in algorithms, RF antennas and hardware upgrades, and achieve effective cost control.
The third solution is the distributed + first-class computing architecture, which co-integrates the data of multiple surround 4D radars on the car body in the first-class domain controller (for example, the intelligent driving domain controller), so as to reduce the data processing requirements on the end-side.
In particular, the dense point clouds brought by 4D imaging radar and the altimetry ability brought by the pitch angle bring the possibility of deep neural network model training. At the same time, with the help of the BEV feature map of the radar, it can be fused with the BEV feature of the image to further enhance the confidence of target detection.
By reducing the complexity of vehicle architectures, the industry will be able to continue to cope with the rapidly growing complexity of hardware and software due to feature-rich, highly automated applications, such as more sensor configurations. In the eyes of industry insiders, the key to the decoupling of software and hardware lies in continuously simplifying the complexity of the end side and concentrating more processing power in the first domain.
In Aptiv's view, in the traditional automotive era, a small number of sensors are designed to process data independently, which is in line with realistic needs, and can be adapted to different platforms and different positioning models. A typical representative is the forward-looking all-in-one solution.
However, with the increase in the number of sensors and the advancement of functions, the current distributed processing architecture of sensors means independent development and fragmented software algorithm function deployment and additional chip costs.
That's how we are todaySatellite processing architecture
The architecture separates intelligence from the sensors and concentrates it in a powerful** domain controller, keeping the "satellite" sensors containing only the necessary hardware for data acquisition (e.g., transceivers for millimeter-wave radars), while processing and decision-making take place in the domain controller.
At the same time, the trend of decoupling software and hardware is also affecting the architecture design of sensors, and once standardized point cloud processing software development is realized, plug and play can be realized, which greatly shortens the entire development cycle of the system.
In addition, from the traditional 3D to 4D imaging technology upgrade, the software algorithm ability has become a shortcoming of millimeter-wave radar companies that relied more on hardware design, development and manufacturing in the past. (Especially from 4D to imaging, the requirements for software are also a qualitative change).
For example, in the case of increasing point cloud density, how to increase the probability of success of detection, reduce echo noise, and realize the detection of objects with weak signal reflection at a distance, it is also necessary to solve the problem of multi-signal interference.
At the same time, compared with the detection of traditional 3D radar (due to insufficient angular resolution, inability to distinguish static targets, and inaccurate classification), 4D radar also needs to add additional capabilities from point cloud to clustering, accurate target recognition and classification.
Previously, Arbe said bluntly: We are providing the software stack as much as possible to help our end customers (car manufacturers) adapt to this technology as quickly as possible. For car companies and Tier1s, it also takes time to optimize software and hardware performance.
Similarly, Tesla has previously confirmed that at present, the company is only carrying out relevant verification work on the Model S and X models to evaluate the actual performance of the products, and there are no plans to install 4D radar on the Model 3 and Y models.
However, the 4D imaging market itself is also stratified.
For example, last year, the delivery of the standard version of the 4D imaging radar** has dropped to 500-1000 yuan, and the lowest has even fallen below 500 yuan, ** is rapidly approaching the 3D forward radar. According to a number of business sources, it is expected that the ** will decline by 30%-40% in 2024.
The demand for this product mainly comes from the direct replacement of traditional radar. However, the performance itself will not have much impact on the computing power of the whole vehicle (for example, increasing the altimetry capability and outputting more point clouds than before).
The reason is, "At present, the biggest problem faced by 4D millimeter-wave radar as the main sensor is that 4D imaging radar is still in the early stage of pre-installation, and the future performance needs to be greatly improved, and at the same time, OEMs and Tier1 do not know what 4D point clouds can do?" “
For example, the dual-chip cascade scheme can achieve a good balance between cost and performance, and will be the mainstream solution for large-scale 4D imaging radar in the next few years. However, there are also companies looking for technological breakthroughs.
Last year, Aptiv launched the seventh-generation 4D millimeter-wave angle radar, which was developed by a local Chinese team and equipped with the first integrated radar chip in China and the industry's first air waveguide antenna solution.
In addition, Aptiv is also one of the first companies in the world to propose the concept of sensor satellite architecture. According to the data given by the company, the above architecture can reduce the size of the radar sensor by 70% and the weight by 30%, and at the same time, it is also more conducive to the future OTA upgrade of the sensor.
At the same time, the advantage of this architecture is that it can better adapt to the concept of pluggable hardware in the future vehicle. Because the real separation of software and hardware means that the increase or decrease of hardware does not require car companies to redevelop software.
**Originated from the Internet, not for any commercial use, if there is any infringement, please contact to delete. The content is for reading purposes only and does not constitute investment advice, so please treat it with caution. Investors act accordingly at their own risk.