Can domestically produced chips for smart cars break through technological barriers?

In the past decade, the booming mobile phone industry has been the main driving force for the rapid growth of the semiconductor industry. In the next decade, high-level autonomous driving, smart cockpits, in-vehicle Ethernet networks, and in-vehicle information systems will generate new semiconductor demand, among which automotive SOC, power semiconductors, automotive sensors, storage, multi-function MCUs, in-vehicle Ethernet, and advanced communication systems that support OTA upgrades are high-growth tracks in the sub-sectors.
ADAS/AD and intelligent cockpits drive the volume and price of the automotive SOC market to increase.
(1) Autonomous driving SOC: "Hardware pre-embedded + OTA upgrade" is the core factor driving the growth of autonomous driving SOC. It is estimated that the scale of China's autonomous driving chip market will reach 13.8 billion yuan in 2025 and 28.9 billion yuan in 2030, with a CAGR of 25.1%.

(2) Cockpit SOC: The upgrade of in-vehicle intelligent perception, interaction, and scenario applications is the core factor driving the cockpit chip from "single-chip single-screen" to "one-chip multiple-screen". It is estimated that the domestic cockpit SOC market size will reach 11.2 billion yuan in 2025, with a CAGR of 24.5%.

01

The electrification and intelligence of automobiles lead the new trend of SoC industry
The booming development of the mobile phone field has been the main driving force for the rapid growth of the semiconductor industry in the past decade. The electrification and intelligence of automobiles are expected to become the new growth level of the semiconductor industry. The industrial transformation will definitely give rise to new technology manufacturers and industry leaders. In the future, automobiles will become the main growth driver of the entire semiconductor industry, just like mobile phones and computers. The main reason is that more advanced autonomous driving, smart cockpits, in-vehicle Ethernet networks and in-vehicle information systems will give rise to new semiconductor demand. Horizon predicts that the global automotive chip market will be about US$100 billion in 2030, a year-on-year increase of 190% compared with the global automotive chip market of US$37.5 billion in 2017.
The number of chips equipped in new energy vehicles is about 1.5 times that of traditional fuel vehicles. It is expected that the semiconductor content of a single vehicle will double in 2028 compared with 2021. The higher the level of autonomous driving, the more sensor chips are required. L3 level autonomous driving is equipped with an average of 8 sensor chips, while the number of sensor chips required for L5 level autonomous driving has increased to 20. Similarly, the amount of information that the vehicle needs to process and store is positively correlated with the maturity of autonomous driving technology, further increasing the amount of control chips and storage chips installed. According to statistics, by 2022, the average number of chips installed in new energy vehicles will be about 1,459, while the number of chips installed in traditional fuel vehicles will be 934. Strategy Analytics expects the average silicon content per vehicle to double from $530 per vehicle in 2021 to more than $1,000 by 2028, and the silicon content of high-end manufacturing vehicles may exceed $3,000.

▲The number of chips installed in each car in China from 2012 to 2022 (unit: piece)

▲Automotive semiconductor market size and single-vehicle semiconductor value trend (right axis unit: US dollars)
Automotive chips have the following main application areas:
The main control chip is used to generate the calculation and generation functions of the main control signals of the car. The main control chip receives the signals collected by various sensors, calculates the corresponding processing measures, and sends the drive signal to the corresponding control module. Therefore, the main control chip is equivalent to the "brain" of the car.
Power chips are the part with the largest increase in value of new energy vehicles. The demand side is mainly IGBT, MOSFET and IPM modules with multiple IGBTs integrated. The core is used in high current and high voltage environments.
CMOS chips are chips that convert photons into electrons for digital processing and convert image signals into digital signals. They include microlenses, photodiodes, processing chips and IO interfaces, and are key components of cameras. With the improvement of the level of autonomous driving, it is estimated that assisted driving above L3 will require about 18 cameras, mainly used in areas such as reversing rear view, surround view, front view, and turning blind spots.
RF receivers are important components of wireless communications. RF chips refer to chips that can convert RF signals into digital signals. They include power amplifiers PA, filters, low noise amplifiers LNA, antenna switches, duplexers, tuners, etc. In the future, RF chips will help the development of C-V2X technology like the ears of cars, organically linking traffic participation elements such as "people-car-road-cloud", making up for the shortcomings of single-car intelligence and promoting the development of collaborative application services.
Ultrasonic/millimeter wave/lidar are sensors for perceiving the body of the car. Smart cars obtain a large amount of data through sensors, and L5-level cars will carry more than 20 sensors. On-board radars mainly include ultrasonic radars, millimeter wave radars and laser radars. Among them, China's ultrasonic radars have developed relatively maturely and have low technical barriers; millimeter wave radars have high technical barriers and are important sensors for smart cars. They are currently in a stage of rapid development; laser radars have high technical barriers and are important sensors for high-level autonomous driving, but they are currently expensive, difficult to pass vehicle regulations, and difficult to land.
Memory chips are the "memory" of smart cars. The upgrade of autonomous driving technology will bring about a long-term trend of continuous high growth in the bandwidth of automotive storage. In the future, automotive storage will move from GB level to TB level.
Automotive panels are showing a trend of multi-screen. The acceleration of automotive intelligence and electrification will drive the number of panels per car, and automotive panels are also beginning to move towards standardization. From the perspective of demand, display screens are increasingly used in cars, and the demand is growing strongly. In-vehicle displays mainly include central control display screens, instrument display screens, windshield composite head-up display screens, virtual electronic rearview mirror display screens, rear seat entertainment display screens, etc. With the promotion of factors such as the Internet of Vehicles, new energy vehicles, and unmanned driving, people's demand for in-vehicle panels with functions such as navigation, vehicle status, and multimedia audio and video will continue to expand.

▲Main applications of chips in automobiles
Traditional functional vehicles use a distributed electronic and electrical architecture. Discrete ECU hardware and software are tightly coupled and each ECU is highly independent. Hardware resources cannot be shared and data islands are formed. The overall cycle of feedback on new user needs is more than 20 months, making it difficult to form a continuous and rapid iteration of software development. Therefore, the software-defined car development model drives the evolution of the vehicle's electronic and electrical architecture from distributed to centralized. The core is the centralized development of on-board computing, and highly integrated domain controllers and on-board central computing platforms are the key.

02

Automotive SoC Core Components of Smart Car Functions
Currently, MCU is the largest sub-category in automotive chips. According to data released by IC Insights, in 2021, the top three global automotive chips in terms of product segment share are microprocessors, analog chips and sensors, accounting for 30%, 29% and 17% respectively. The full name of MCU chip is microcontroller unit, also known as single-chip microcomputer or single-chip microcomputer.
It is a chip-level computer that appropriately reduces the frequency and specifications of the central processing unit, and integrates peripheral interfaces such as memory, counter, USB, A/D conversion, UART, PLC, DMA, and even LCD drive circuits on a single chip. Usually MCU can only complete a few tasks, such as turning on smart wipers, or automatically locking after getting off the car, etc. Therefore, there may be hundreds of MCUs in luxury cars to realize various intelligent functions.
▲Composition of automotive chips in 2021
System-level chips (SOC) were born in the era of artificial intelligence. In the era of artificial intelligence, computing architecture has evolved from a single chip model to a fusion heterogeneous multi-chip model. System-level chips (SOCs) that heterogeneously integrate CPUs with general/special chips such as GPUs, FPGAs, and ASICs, and integrate AI accelerators, have emerged.
In a broad sense, MCUs with slightly stronger computing power (above 2K DMIPS) in the automotive field can be considered SOCs. Arteris predicts that the number of single-vehicle SOCs in the future will be 23, and high-computing power SOCs are mainly aimed at two fields on the vehicle side, namely smart cockpits and smart driving. The
automotive field is replicating the evolution of the mobile phone field from "feature phones → smart phones". SOC plays an important role in the era of smart cars, among which the "xPU" that realizes artificial intelligence computing power is crucial:
PC field: mainly focusing on general computing, so the chip architecture is "logic computing CPU + GPU";
mobile phone era: the most important applications are payment, playing games and taking pictures. Therefore, the chip architecture adds ISP, and the mobile phone SOC chip structure is "CPU + GPU + ISP (image processing)".
Smart car era: Chip architecture is more complex, and the large screen in the cockpit pays great attention to graphics processing, so GPU is needed. Cameras are the main sensors for realizing autonomous driving, so ISP is needed. In addition, neural network NPU with brain-like functions is needed to realize autonomous driving. Therefore, the structure of automotive SOC chips is "CPU+GPU+ISP+NPU".
CPU is responsible for logical operations and task scheduling; GPU, as a general accelerator, can undertake neural network calculations and machine learning tasks such as CNN, and will undertake the main computing work for a long time; FPGA, as a hardware accelerator, has the advantage of programmability, and performs well in sequential machine learning such as RNN/LSTM/reinforcement learning, and plays a prominent role in some mature algorithm fields; ASIC can take into account performance and power consumption, and as a fully customized solution, it will become the final choice after the autonomous driving algorithm matures.
▲Comparison of CPU, GPU, FPGA and ASIC (NPU, TPU)
From the perspective of chip process, different automotive chips have large differences in process requirements. MCU mainly relies on mature processes, and 70% of MCU production in the world comes from TSMC; while cockpit, autonomous driving SOC and AI chips and other main control chips continue to pursue 7nm and below advanced processes.
▲Comparison of wafer size and process of key chip products
Automotive MCUs closely follow the development of automotive electronic and electrical architectures. SOC chips will integrate some low-end MCU functions. Therefore, the use of MCUs per vehicle will decrease in the future, and the use of distributed domain control will gradually increase from the current 30-40 to 70-80. However, in the future, as the centralized architecture is implemented and computing power is concentrated on the vehicle computing platform, the use of automotive MCUs will gradually decrease to about 50-60.
SOC chips cannot replace all MCUs. On the one hand, not all MCUs need to be connected to SOC chips. For example, if the MCU solution is not used for the "control method of making the turn signal flash", all connected to SOC chips will form a star network, which will not only increase the number of wires, but also increase the difficulty of management. On the other hand, some MCUs are also needed as alternative solutions for SOC chip safety redundancy.
Split the autonomous driving SOC structure, including CPU, GPU and other types of custom chips (such as NPU, deep learning accelerators (DLAs) and computer vision processors (CVP)). In addition, a typical autonomous driving SOC structure also includes the following parts:
at least one microprocessor (MPU) or digital signal processor (DSP), but it can also have multiple processor cores; the memory can be one or more of RAM, ROM, EEPROM and flash memory;
oscillators and phase-locked loop circuits for providing time pulse signals;
peripherals consisting of counters and timers, power supply circuits;
connection interfaces of different standards, such as USB, FireWire, Ethernet, universal asynchronous receiver and transmitter, and serial peripheral interface; voltage conditioning circuits and regulators.
After splitting the cockpit SOC architecture, we found that behind the multiple high-resolution screens and smooth systems, it is not only the computing power and video processing capabilities of the car chip that are competing, but also performance indicators such as AI capabilities. For example, the Qualcomm 8155 chip is Qualcomm's third-generation Snapdragon automotive digital cockpit flagship platform. It is a heterogeneous architecture chip that includes CPU, GPU, DSP, ISP and AI engine:
In the CPU part, the 8155 chip adopts an 8-core design of 1+3+4, with the core being Qualcomm Kryo485. The main frequency of the large core is 2.96GHz, the main frequency of the three high-performance cores is 2.42GHz, and the main frequency of the four low-power small cores is 1.8GHz.
As for the GPU, both the 8155 chip and the Snapdragon 855 use Adreno640. At the same time, the Hexagon690DSP and Spectra380ISP used by the 8155 chip are exactly the same in name compared to Qualcomm Snapdragon 855 and 855+.
In addition, unlike the autonomous driving chip, the 8155 chip does not have an independent NPU core, and AI computing is mainly completed through the AI ​​engine composed of DSP, CPU and GPU. Among them, Hexagon690 has an AI computing power of 7TOPS, and the sum of the AI ​​computing power of CPU and GPU is 8TOPS.