With the development of new energy vehicles in recent years, the automobile "four (intelligent, networked, electrified, shared)" has become a major car company, zero Component manufacturers are in a position to fight, and they don't hesitate to invest heavily in them. There are also many startup companies. The entry of Huawei will undoubtedly make the competition that has already been fiercely competitive, and the competition will be even more fierce.
Only in the autopilot field has already brought together dozens of players, many of which are international giants. Looking at the pattern of autonomous driving, it is possible that all the companies in the chain will not be able to compete with Huawei.
In the era of fuel vehicles, domestic car "heart disease" is difficult to remove , Engine control technology relies on foreign suppliers, so that it becomes a pain that cannot be calmed down. The engine control system still relies on the supply of joint venture factories, and the new technology relies on the foreign side. In the era of intelligent networked cars,The chip is a high ground that the Chinese auto industry cannot lose.
Huawei Haisi series chips in the field of consumer electronics, although it is still not compatible with Qualcomm, Apple A series Counterbalance, but it is also a small achievement. With years of investment in the IC industry, Huawei's Shengtuo system's AI chip, from the public information point of view, up to 352TOPS super-power is already in the leading position in the industry. Based on this, Huawei has built an MDC (Mobi) platform to provide auto-driving full-stack full-service services for car companies.
Huawei Shengteng 910 chip, single chip computing density exceeds Nvidia
In high-level autonomous driving, the discrete electronic control unit (ECU) that the car can no longer use to match each function, but a partially centralized domain controller (DCU).
DCU needs to undertake multi-sensor integration, positioning, path planning, decision control, wireless communication, high-speed communication calculation. In centralized and hybrid architecture DCU also needs to handle the data processing of all or part of the sensor. Because of the large number of operations that need to be completed, the DCU generally needs to match a core computing processor, which is a chip, to provide support for different levels of computing power. The higher the support, the more functions are supported, so everyone is pursuing high computing power.There are many solutions in the industry such as NVIDIA, Infineon, U.S. Renesas, TI, NXP, Mobileye and so on.
Although chip development requires a lot of money, but the chip as a core component of autonomous driving, its high price and considerable profits, has attracted many giants to invest. Among the many players are consumer electronics semiconductor giants, traditional automotive semiconductor suppliers, and new players. I have selected a few representative companies to do a simple combing.
1. NVIDIA: NVIDIA has been working on the development of AI chips in recent years, and has complete solutions in the field of autonomous driving:
1) DRIVE PX2 open artificial intelligence vehicle computing platform;
2) NVIDIA DriveWorks Software Development Kit;
3) Digital Cockpit, High-Precision Map, Advanced Auxiliary Driving Solution.< /p>
2. Intel: The PC era in the past PC era CPU King Intel lost the smartphone era, fearing to miss again automatically Driving era. After launching the IntelGo platform, it took a lot of money to acquire Mebileye. The auto-driving open platform Eye Q5, which is dominated by the ADAS visual field, was also launched in December 2018.
3. Qualcomm: Qualcomm's layout in the automotive industry is not limited to its communication field. The 820A in the ADAS and cockpit entertainment fields also performed well. At the CES show in 2019, it was built on the basis of the Snapdragon chip. Scalable Autopilot - DriveAutomotive,It is available for car manufacturers to add modules according to their own needs.
But it’s just that the product-side layout doesn’t seem to show Qualcomm’s confidence in the entry industry, and NXP’s acquisitions are twists and turns, but it’s ultimately fruitless.
4. NXP: As a traditional supplier of automotive semiconductors, NXP targets multiple control domains, including: connectivity, body comfort, autonomous driving, infotainment, power Assembly. Among them, the BlueBox computing platform is built in the automatic driving domain, and its BlueBox2.0 supports the L4~L5 level automatic driving demand.
5. Renesas: Renesas Electronics' R-Car V3M high-performance image recognition system-on-chip (SoC), which greatly optimizes the smart camera, panoramic viewing system, u>Lidar and other applications, and Renesas also launched Renesas Autonomy open platform ADAS/autopilot.
6. TI: Texas Instruments has its own advantages, from the low-level ADAS field, its vision, ultrasound, millimeter wave radar and other sensors The chip layout, as well as the extension of the Jacinto series SoC from infotainment systems to ADAS, introduces the TDA series SoCs, which quickly and cost-effectively capture the market. In the field of subsequent automatic driving, TI will continue to exert its strength.
7. Tesla: Tesla is not willing to just build a car. Recently released its self-driving chip FSD, which will be carried in Tesla in the future. u>AutopilotHardware platform. From the perspective of hardware integration, Tesla completed the integration from chip to board level, system integration, and vehicle integration, and opened up the autopilot industry chain layout.
8. Horizon: The horizon that was not established for 4 years is quite popular with the capital market, and its valuation has soared.In 2018, the company relied on its combination of soft and hard AI processor technology, has released the Matrix autopilot computing platform and the horizon XForce edge AI computing platform.
9. Cambrian: Cambrian depth learning chip Cambricon-1M, support for personalization Deep learning, which can be used for multi-channel video real-time processing, including auto-driving.
10. Xijing Technology: Xijing Technology started from chip research and development. In October 2017, it joined Zhenhua Heavy Industry to release its autonomous driving brand Qomolo for unmanned heavy truck ports. surroundings. Its autonomous driving solution is based on its DeepWell deep learning brain chip, which completes the industrial chain layout from chip development to vehicle production.
In the field of autonomous driving chips, domestic companies are also emerging, except for the above three.There are also companies like Zero Run Technology, Feibu Technology, Jiefa Branch, Shenjian Technology, etc., which are not listed here. Of course, it is not excluded that some enterprises have quietly invested a lot of manpower and resources. For example, Pingtou, who is dedicated to the field of intelligent networked chips, has not yet released its related products, but it is certain that the autopilot AI chip will definitely be involved in the future.
The technical barriers and investment of the chip are extremely high. The field of automotive chips has always been the domain of traditional semiconductor manufacturers. Even Intel and Qualcomm can not attack for a long time. Use the acquisition method to seek a qualitative breakthrough. For startups and even domestic semiconductor companies, it is a great challenge to break through in this field. However, in the well-known background, AI chip startups are also favored by the capital market, which is considered to be good, at least to ease the pressure on capital. Whether the domestic AI chip company can make a breakthrough is worth looking forward to.
The road environment of the vehicle is complex and variable, which makes the perception unable to rely on a single sensor. In high-level self-driving cars, sensors such as positioning, radar, and vision are cooperatively integrated, and can input the collected environment in the form of images, point clouds, and the like. Data, and through the extraction, processing and fusion of algorithms, further form a complete driving situation map of the car to provide a basis for driving behavior decision-making.
Multiple sensors cooperation is essential
In this round of autopilot industry entrepreneurship,The field of environmental perception is dominated by visual, millimeter-wave, and laser radars. Due to the maturity of cost and maturity, the visual and millimeter-wave radar is the first to be mass-produced, from the L0 level warning to the advanced assisted driving iteration.
Although the introduction of the Huawei Auto Show, its camera, millimeter wave / ultrasonic radar, laser radar, T-Box, GPS, The executive agency and so on are provided by the partners, but the author believes that in the future, Huawei will develop independently at least on the sensor algorithm level of vision, millimeter wave radar, and surround vision system, and does not rule out the possibility of launching its own sensor products.
I sort out startups by sensor type and function.
(1) ADAS forward vision camera
Forward ADAS vision camera currently has a single Projects such as eyes, binoculars, and trinoculars,Due to the mass production cost factor, the main flow production plan is mainly monocular, and binocular cameras are equipped on some models.
With the mass production of the ADAS (Advanced Assisted Driving) function, the domestic market share has gradually become clear. The passenger vehicle sector is dominated by international suppliers such as Mobileye and Bosch; domestic startup products are mainly distributed in the commercial vehicle field with the use of monocular solutions, following the promotion of commercial vehicle regulations, and early warning in the passenger vehicle field ( Mass production has been achieved with FCW, LDW) and AEBS functions. Of course, this is not to say that domestic companies have no layout in the passenger vehicle sector, but only from the perspective of market share.
Tesla's autopilot interface, the company insists on using visual sensors for autopilot
Compared to the monocular solution, the binocular or multi-mesh scheme is superior to the former in terms of detection accuracy, target detection type, detection range, etc. However, due to binocular (multi-head) The high cost of the program makes monocular mainstream, and in the next few years, it may still be dominated by monocular solutions. In the technical realization, the single-phase cooperation with the 77 GHz forward millimeter-wave radar is combined with data fusion to make the two complementary advantages. Technical requirements.
(2) Millimeter wave radar
Vision camera is subject to target detection The influence of environmental factors is large, and the target object or recognition effect may be unrecognizable under special circumstances. Compared with the visual camera, the millimeter wave radar is less affected by environmental factors;The vision is better than the millimeter wave radar in detecting the type of target. The technical solutions of ADAS functions such as AEBS and ACC combine the two to make up for the perceived defects of a single sensor.
The domestic millimeter wave radar band is mainly 24GHz and 77GHz, and 79GHz is still in the research and development stage. The detection distance is only 50~70m for the 24GHz (short range) millimeter wave radar, and the 77GHz millimeter wave radar (medium and long distance) can reach 150~250m.
With the cost reduction of 77GHz millimeter wave radar and the demand for millimeter wave performance of ADAS function, 77GHz millimeter wave radar gradually becomes forward radar in mass production scheme Mainstream. 24GHz millimeter-wave radars tend to be used for lateral detection applications, such as BSD (blind zone detection), LCA (change lane assistance), etc. Since the early warning system is only a reminder to the driver and does not involve vehicle control, the performance requirements are lower, the price It is also easier for manufacturers to accept.However, the future 24GHz wave radar may also be replaced by 77GHz or 79GHz radar.
From the perspective of domestic passenger vehicle market share, 77GHz millimeter wave radar is mainly based on Bosch, Continental, Denso, and Anbofu; 24GHz millimeter wave radar It is mainly based on Veoneer (automatic driving company after the split of Autoliv), mainland China, Haila, Anbofu and Valeo. Tier1 provides system solutions for car companies in the passenger vehicle market (including sensing, control, execution and other components). Due to the inferior position of domestic suppliers in system integration capabilities, additional product reliability, stability, consistency, cost control, There are also disadvantages in terms of quality control, etc. It is more difficult for domestic component companies to enter the passenger vehicle market than the commercial vehicle sector. Of course, there are also a number of millimeter wave radar companies in cooperation with the passenger car period.
Different from passenger cars, commercial vehicles are production materials, end users have low demand for comfort and active safety, and the promotion of ADAS functions depends on regulations. . At the present stage, in the field of commercial vehicles,The mass production assembly of the millimeter wave radar is mainly to meet the AEBS function, so the mass production time node in the commercial vehicle front loading market will be affected by regulations. In accordance with traffic regulations, the passenger car needs to be equipped with an AEBS system in April 2019. The truck needs to be equipped with an early warning function from May 2020 and an AEBS system in 2021. As mentioned above, the performance requirements of the early warning system for visual and millimeter wave radar are slightly lower, and more are the cost factors to be considered, so this also gives the domestic millimeter wave radar enterprise the opportunity to mass production.
On the commercial vehicle market, the mainland and Fujitsu days are also powerful competitors, domestic millimeter wave radar companies such as: Wuhu Sensi < u>Tektronix, Yi Laida, Mu Niu Technology, Zhibo Technology, Suzhou Millimeter Wave, Cheng Tyco technology and other companies will also face a lot of competitive pressure.
With the increasing penetration of ADAS functions, millimeter waves are important sensors and the market prospects are expected. The domestic millimeter wave has made a breakthrough in the future and it is worth looking forward to.
Lidar is indispensable for high-level automatic driving An important sensor that accurately obtains three-dimensional position information, can determine the position, size, speed, posture, external topography and even material of the object, and can provide accurate information for target recognition, target tracking, and obstacle detection. Compared with Vision and Millimeter Wave Radar, it has excellent ability in ranging and resolution, and can independently build 3D models, but the penetration capability of Lidar is relatively weak.
Lidar has the strongest perception of the environment
Lidar can be divided into: mechanical laser according to the presence or absence of mechanical machinery. There are two types of radars and solid-state laser radars. Mechanical laser radars have rotating parts that control the angle of laser emission. They are large and expensive, and the measurement accuracy is relatively high. Solid-state laser radar relies on electronic components. Control laser emission angle, no mechanical rotating parts, small size.
In the field of laser radar, foreign companies Velodyne, Quanergy, IBEO are at the forefront. US Quanergy and entrepreneurship The company Ouste uses mechanical rotary laser radar, which is relatively mature on the 16/32/64 line; Germany Ibeo's laser radar products are mainly low-beam 4-wire and 8-wire. In addition, Ibeo and Valeo cooperate to launch hybrid solid-state laser radar; German startup Quanergy and Israeli startup Inoviz are based on solid-state laser radar.
Velodyne's Lidar products
Domestic companies Sagitar Juchuang, Radium Technology, Beixing Photon, Hesai Technology, Han's Laser, and Beijing Branch are currently focusing on the development of laser radar, which is involved in both mechanical rotary laser radar and solid-state laser radar.
Because the current cost of laser radar is too high, it is more used in high-level automatic driving in the automotive field, but with the future development of solid-state laser radar technology Reduce costs,Make bulk loading possible.
At present, the hardware suppliers of laser radar are mainly start-up companies, and some suppliers are cutting in from the relatively mature laser field. Domestic and foreign companies have relatively small gaps in technology, and even have advantages over foreign products of the same grade in terms of accuracy and price. This also gives domestic companies the possibility to chase foreign companies.
3, domain controller and full stack autopilot
With the development of automotive electronics technology, there are more and more electronic control units for cars, and there are hundreds of them. In order to solve the problem of a large number of automotive ECUs, the concept of domain controllers came into being. The Autopilot Domain Controller (DCU), as an automatic driving data processing platform above the L3 level, naturally attracts many players, including many startups, Internet companies, and traditional parts suppliers.Taking into account the complexity of the DCU, personally feel that car companies, system vendors, semiconductor manufacturers jointly developed, to play their respective advantages may be the best choice. Of course, this combination is not uncommon.
Huawei MDC platform is similar to DCU. In essence, in order to solve the problem that the number of automotive ECUs increases, the vehicle control system becomes complicated and the capacity reaches the upper limit. The following is an introduction to the DCU.
Try out what capabilities DCU needs.
The above figure is the structure level of the intelligent networked car. In the distributed architecture, after the sensor-side data processing is completed on the sensor side, the detection result is sent to the DCU. After the DCU completes the multi-sensor data fusion process, the planning and decision-making are performed. After the decision command is issued, the execution unit completes the control. Compared with the distributed architecture, after the centralized architecture obtains the original identification data from the sensor side, the DCU completes the sensing layer processing, which of course requires not only the chip. High computing power, and also requires DCU suppliers to have sensor development capabilities directly or indirectly.
DCU must at least include path planning and determine the best path (from security , Convenience, environmental perspective) Finally to the ability to coordinate multiple actuators.
The complexity of DCU determines that it is difficult for a single startup to compete with giants or giants. This is not only reflected in the financial strength, but also in resource integration capabilities, technology reserves, and talent reserves. The startup companies are deeply involved in a certain field or start from a specific scene and iterate to the industry chain. Strategy,It is a wise choice. At present, the automatic driving is still in the L2 promotion stage. It takes time to realize the L3 automatic driving and the above, and the startup company is not without the chance to win.
Personally feel that the core of automatic driving above L3 level is DCU (perception and execution layer depends on partner or partial self-research), DCU supplier and full stack automatic The difference between a driving solution provider is that the latter provides other hardware including the DCU, as well as scene customization algorithms. So I categorized the full-stack autopilot program with DCU, and I have selected some representative companies among many companies.
Automatic driving above L3 level and There are many companies in the DCU field.On the domestic trunk logistics unmanned program, it has brought together Ali, Jingdong, Tucson Future, Zhijia Technology, Mainline Technology, Yuchao Technology, Xijing Technology, Wuhan Huanyu and so on. From the perspective of cooperation with OEMs, the competition is also extremely fierce. Only Audi has announced that it is an auto-driving partner above L3 level. In addition to its subsidiaries AID and zFAS platforms, there are Huawei and Horizon.
With the advancement of chip technology, the future DCU architecture will tend to be centralized. Accompanied by this, existing sensor companies are bound to face the choice of transformation, personally feel that there is nothing more than the algorithm and DCU or full-stack solution in two directions, such as pure visual solutions, laser radar advanced automatic driving program. Inevitably, it will compete with existing DCU and full-stack system solutions. It is not known that several people can win, but what is certain is that there will not be so many enterprises in the future, market integration is inevitable, and there may be survival opportunities in product segmentation and market segmentation.