Automatic drive chip with surging demand

With the continuous improvement of laws and regulations, medium and high-level automatic driving is expected to be gradually implemented. In the past, due to the difficulty in the development and testing of automatic driving software and algorithms, and the imperfection of relevant policies and regulations, the overall market maturity of automatic driving was not high. In the era of vehicle intelligent transformation, the intelligent cockpit can integrate more information and functions, and bring more intuitive and personalized experience to users. Therefore, it has become the pioneer of vehicle intelligent. Since 2020, countries have successively issued policies related to automatic driving or high-level automatic driving operation licenses:

1. United States: in February 2020, the National Highway Traffic Safety Administration (NHTSA) approved the production and release of unmanned electric delivery vehicles by nuro, a autonomous vehicle start-up.

2. Japan: in March 2021, the Japanese government approved Honda L3 level automatic driving legend to be listed in Japan.

3. Germany: in December 2021, the German Benz L3 autonomous vehicle EV EQS was approved by the German Federal automobile transport administration, which can replace human drivers to control the vehicle status in some sections.

4. China: in September 2021, the General Administration of market supervision (Standards Committee) officially issued the national recommended standard for classification of automobile driving automation (GB / T 40429-2021). According to the national standard, starting from level 3 automatic driving, the object of target and event detection and response is changed from the driver to the system, and the dynamic driving task backup is also changed from the driver to the dynamic driving task backup user.

With the active layout of major manufacturers, there is a huge space for the future automatic driving industry. At present, automatic driving at L3 and above is expected to take the lead in the application of closed, semi closed and low-speed scenarios. As a low-speed complex scenario of automatic driving, autonomous parking will provide data training and accumulation in the low-speed domain for the evolution of automatic driving technology. Although there is still a long way to go before the commercialization of the high-speed scene of automatic driving, Tesla, Google, NVIDIA, Qualcomm and other manufacturers still focus on high-level automatic driving in order to take the lead before the industry turning point. According to the prediction of IHS, autonomous vehicle will start a round of explosive growth around 2025. By 2035, half of the vehicles running on the road will be self driving, and the total revenue scale of self driving vehicles and related equipment and applications will exceed 500 billion US dollars.

In the process of the development of automobile E / E architecture from distributed architecture to centralized architecture, automatic driving chip, as the carrier of computing, has gradually become the core of intelligent automobile era. Under the trend of "software defined vehicle", the chip, operating system, algorithm and data together form the calculation ecological closed loop of intelligent driving vehicle, and the chip is the core of the ecological development of intelligent driving vehicle. The pioneer of automotive electronic and electrical architecture reform, represented by Tesla, took the lead in adopting the centralized architecture, that is, a computer is used to control the whole vehicle, and the domain controller gradually integrates the early sensor, data fusion, path planning, decision-making and other computing processor functions.

The upgrading of the demand for computing power drives the growth of the vehicle chip market. In 2020, the demand for chips in the automotive field has accounted for 11.4% of the global chip market. The continuously rising demand for computing power will drive the growth of the market scale of on-board computing chips, and the on-board computing chip market will usher in a period of rapid development. According to the EU billion think tank, the market scale of China's automotive computing chips will reach US $1.51 billion in 2021, and the market scale will rapidly increase to US $8.98 billion in 2025.

For car enterprises, computing power and power consumption are the two main factors for choosing on-board chips / computing platforms:

1. Computing power: for high-level intelligent driving systems, the increase in the number of sensors and the improvement of the resolution have brought huge data processing requirements, and the complexity of the algorithm model has also been greatly increased. With the gradual centralization of automobile E / E architecture, the computing capacity of intelligent vehicles will be mainly realized by a few domain controllers or central computing platforms, which also puts forward higher requirements for the computing power of a single vehicle chip.

"Hardware embedding and software upgrading" has become the mainstream strategy of car enterprises. Intelligent head car enterprises preset large computing power chips in the new generation of vehicles. Automotive products have a long life cycle, generally 5-10 years. The upper limit of computing power of the on-board computing platform determines the upper limit of software service upgrade that can be carried in the vehicle life cycle. Compared with the software iteration, the software iteration cycle is faster, so the iteration cycle of intelligent driving software and the hardware replacement cycle are misplaced. In order to ensure the continuous software upgrading capability of the vehicle in the whole life cycle, the main engine factory adopts the strategy of "hardware preset and software upgrade" in intelligent driving, and provides sufficient development space for subsequent software and algorithm upgrading and optimization by presetting large computing power chips.

2. Power consumption: in order to support and be compatible with the requirements of sensors and actuators of L3 and above intelligent driving systems with a large number and types, the on-board computing platform adopts heterogeneous chip hardware solutions to meet the system interface and computing power requirements. Compared with the traditional ECU, the complexity of the on-board computing platform has increased several times, facing multiple challenges such as power consumption, heat dissipation, electromagnetism and quality, and there is a physical upper limit. Therefore, although the current industry generally uses "tops" as the unit to evaluate the theoretical peak computing power of autopilot chips, and the major chip manufacturers also constantly refresh the peak computing power, the effective utilization rate of computing power in the actual scenario is not high, and the theoretical peak computing power of autopilot chips may not be fully released in the actual operation, which is mainly limited by factors such as power consumption and environment.

The absolute computing power of the chip is of course important, but for the main engine factory to develop mass-produced models, the chip selection needs to take into account multiple factors such as computing power, power consumption, cost, ease of use and isomorphism. Therefore, how to help customers realize the most efficient operation of algorithm software under limited computing power is the core standard to measure the competitiveness of chip manufacturers.

From the development trend, the automatic driving SOC chip will develop to the heterogeneous architecture of "CPU + XPU". In the long run, the CPU + ASIC solution will be the mainstream in the future. SOC is a system level chip. Compared with MCU, it adds audio processing DSP, image processing GPU, neural network processor NPU and other computing units to the architecture. It is commonly used in fields with complex functions such as ADAS, cockpit IVI, and domain control. With the development of intelligent automobile, the structure of automobile chip has evolved from MCU to SOC. At present, the mainstream SOC architecture schemes of automatic driving chips in the market are divided into three types: (1) CPU + GPU + ASIC, (2) CPU + ASIC, (3) CPU + FPGA. In the long run, customized and mass-produced low-power, low-cost dedicated automatic driving AI chips (ASICs) will gradually replace high-power GPUs, and the CPU + ASIC solution will be the mainstream architecture in the future.

The consumer electronics chip giant entered the market and actively explored the opportunities in the smart car market. As the penetration rate tends to be saturated, the smart phone market enters a bottleneck period. The high growth rate and high profit brought by the smart phone chip market in the past can not be sustained. Therefore, the consumer electronics chip giant urgently needs to find new market opportunities to expand its profit space. Since 2014, Qualcomm and NVIDIA, two major consumer electronics chip giants, have taken the lead in laying out smart car computing chips to seize the market opportunity. Chip is an industry that relies on high R & D investment and achieves scale effect through large-scale production to spread the cost. Therefore, manufacturers who grasp more competitive advantages in the early stage of the market will gain cost advantage through scale effect after mass production.

In terms of the competition pattern of on-board computing chips, NVIDIA, Mobileye, Qualcomm and other manufacturers have obvious advantages. The traditional automobile chip market has long been occupied by traditional chip manufacturers such as Ti, NXP and Rissa, while the blue ocean market of on-board computing chips brought about by the development of automotive intelligence has attracted many parties to enter, forming four camps: consumer electronics chip giants, emerging chip technology companies, traditional automobile chip manufacturers, and self-developed / joint-venture chip manufacturers of main engine manufacturers. The pattern of the automobile chip market is gradually being reshaped:

(1) In the field of autopilot computing chips: NVIDIA and Mobileye, backed by Intel, are in the first echelon, while Qualcomm, Huawei Hisense and horizon are in the second echelon, showing a strong upward trend.

(2) In the field of intelligent cockpit chips: Qualcomm has an absolute leading edge in product strength and high-end market share, followed by Intel, Renesas, Samsung and other manufacturers, while NXP and Ti are the main players in the middle and low-end vehicle market.

(3) In the Chinese market, the domestic emerging chip technology companies represented by Huawei Hisense, horizon and Xinchi technology also showed strong competitiveness.

NVIDIA, Mobileye and Qualcomm have advantages in the field of automatic driving SOC:

1. NVIDIA: automatic driving at L3 level and above puts forward higher requirements for computing power. NVIDIA is the king of big computing power chips. Since entering the field of automatic driving in 2015, it has been leading the transformation of computing power of on-board chips.

Since NVIDIA put forward the concept of GPU in 1999, it has been iterating relevant technologies. Automatic driving requires large-area image processing. Therefore, NVIDIA is also leading the industry in the era with the "CPU + GPU + ASIC" SOC solution as the mainstream. NVIDIA has established a good customer base through Xavier and Orin generation SOCS, and the layout of large computing power chips / platforms has also enabled NVIDIA to establish a generation difference advantage. At present, it has obvious advantages in automatic driving at L3 level and above.

2. Mobileye: automatic driving at L2 level and below requires a small amount of data and simple algorithm. Mobileye is the leader in the field of assisted driving, and can be said to be the main founder and leader of automotive ADAS technology in the past two decades.

Although it is suppressed by NVIDIA and Qualcomm in the L3 / L4 market, Mobileye still has an absolute advantage in the L2 market with a market share of more than 75%. In 2021, the shipments of mobileyeeyeq chips reached 28.1 million, and by the end of 2021, the cumulative shipments of eyeq series chips exceeded 100 million. However, with the gradual evolution of automatic driving to a high level, Mobileye's products and solutions gradually lose their advantages.

3. Qualcomm: Aiming at the middle and high-end automatic driving market, the king of the intelligent cockpit will march into the driving field.

Snapdragon ride, an autopilot chip platform launched by Qualcomm in 2020, has a computing power of 10-700tops and supports the full scene autopilot of L1-L5. Although it is not as powerful as NVIDIA Atlan chip in terms of computing power, it has also been significantly ahead of other autopilot SOCS such as eyeq.

Since L3 automatic driving is now gradually landing, there is still a certain distance from L4-L5 level, and the requirements for computing power are not so strict. Therefore, Qualcomm has successfully entered the market by virtue of ride platform and won the fixed points of great wall, general motors, BMW, Volkswagen and other important customers. On the other hand, since Qualcomm is the absolute leader in the cockpit field, it has created a huge automobile ecology so far. Through the cooperative relationship established with the engine manufacturer in the cockpit field, Qualcomm can more conveniently promote its own driving domain products.

The automotive electronic and electrical architecture is gradually centralized, and multi domain integration is the general trend. With the continuous development of intelligent and networked vehicles, the traditional distributed electronic and electrical architecture with single-chip microcomputer as the core can hardly meet the development needs of intelligent vehicle products in the future. Therefore, the automotive electronic and electrical architecture is changing from the traditional distributed architecture to the domain architecture and the central computing architecture, and the centralized EE architecture is also an important hardware foundation for realizing software defined automobiles. During the evolution from the coexistence of several domain controllers in the vehicle to the high-performance computer HPC, the integration of multiple domain controllers and the integration of the driving domain and the cockpit domain has become a necessary trend. Finally, the central computing platform with strong central computing power will coordinate the realization of the functions of the intelligent cockpit and automatic driving of the vehicle.

The centralized E / E architecture also puts forward new requirements for automotive software architecture. With the gradual centralization of the automobile EE architecture, the domain controller or central computing platform is deployed in a layered or service-oriented architecture, and the number of ECUs is greatly reduced. The underlying hardware platform of the automobile needs to provide more powerful computing support, and the software is no longer developed based on a fixed hardware, but should be portable, iterative and expandable. Therefore, at the software architecture level, the automotive software architecture is also gradually upgraded from signal oriented architecture to service-oriented architecture (SOA) to better realize software hardware decoupling and software rapid iteration.

Compared with the past, the complexity of automotive software has significantly increased. According to our previously released report "smart car depth Series 1: star sea of Automotive Software", currently, automotive software can be divided into system software, functional software and application software from bottom to top in the software and hardware architecture of smart cars:

(1) System software: composed of hardware abstraction layer, OS kernel (operating system in a narrow sense) and middleware components, it is the core part of the generalized operating system;

(2) Functional software: mainly the core common functional modules of automatic driving, including library components such as automatic driving general framework, AI and vision module, sensor module and related middleware. System software and functional software constitute the operating system in a broad sense;

(3) Application software: mainly including scene algorithm and application, it is the core of differentiation between intelligent cockpit (HMI, application software, etc.) and automatic driving (perceptual fusion, decision planning, control execution, etc.).

Under the new architecture, the development links that software manufacturers participate in are increased, and the difficulty of software development is also greatly increased. Manufacturers with profound knowledge how accumulation of automotive software and full stack capability (bottom development capability - Middleware - upper application) are expected to benefit. Specifically:

1. The integration of cockpit domain and cockpit domain needs the support of hypervisor technology

In the automotive electronic and electrical system, different ECUs provide different services and have different requirements for the underlying operating system. According to the ISO 26262 standard, the automotive instrument system and the entertainment information system belong to different security levels and have different processing priorities. The automotive instrument system is closely related to the power system, which requires high real-time, high reliability and strong safety, and is mainly based on QNX operating system; The infotainment system mainly provides a control platform for human-computer interaction in the vehicle and pursues diversified applications and services, mainly Linux and Android.

After EE Architecture tends to be centralized, the emergence of hypervisor technology makes "multi system" a reality. In the context of the evolution of electronic and electrical system architecture from distributed to domain centralized, various functional modules are concentrated in a few domain controllers with strong computing power. At this time, applications with different security levels need to share the same computing platform. Traditional