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

04

Intelligent cockpit perception, interaction, and scenario application upgrades
Intelligence has gradually become one of the indicators that consumers care more about when buying a car. The intelligent development of the car cockpit is driven by three parts, namely, the perception of the inside/outside environment, multimodal human-computer interaction solutions such as vision and hearing, and the Internet of Vehicles with coordinated perception and computing. The intelligent development of the car cockpit is to achieve intelligent interaction with people, roads, and cars through the configuration of intelligent and networked in-vehicle products. It is an important link and key node for the evolution of the relationship between people and cars from tools to partners.

汽车座舱正成为具有拟人化交互能力的驾驶伙伴。目前智能座舱系统主要包括内饰、电子两大系统，像车内的座椅、空调、灯光、仪表盘、中控屏、车联网、语音识别、手势识别等。智能座舱目前处于智能助理的初级阶段，在硬件方面，座舱内部的实体按键被简化，大屏化、多屏化趋势显著；在软件方面，语音交互技术被广泛应用，人脸识别技术和手势识别技术也被尝试，座舱所实现的功能趋于多样化。

At present, the penetration rate of new cars equipped with smart cockpits in the world and China is 49.7% and 53.3% respectively. At present, the penetration rate of smart cockpits in Chinese cars has exceeded half, and it is expected that the growth of the penetration rate of smart cockpit products in China will lead the global market in the future. At present, China's smart cockpits are mainly equipped in mid-to-high-end models, and the equipment rate of low-end models is relatively low.

▲Assembly rate of smart cockpits in China
Compared with the traditional multi-core multi-screen solution, the high-computing power single-chip solution greatly reduces the system cost and can provide a multi-screen interactive smart interconnection experience. "One core, multiple screens" has become a development trend, and the chip itself is also developing in the direction of miniaturization, integration, and high performance. The technical barriers of cockpit SOC chips are high and the market concentration is high. Under the trend of domestic substitution, domestic cockpit SOC manufacturers are expected to usher in development opportunities.
▲Overview of the competitive landscape of the core market segments in the field of smart cockpit hardware
The automotive E/E architecture will evolve in the direction of "distributed" → "domain centralized" → "central computing". In sync with the automotive E/E architecture, the cockpit chip solution will also undergo the evolution of "single core, single screen" → "single core, multiple screens" → "integrated development" in three stages:
(1) Different cockpit electronic devices under the distributed architecture are controlled by different controllers, which is manifested as "single core, single screen". However, with the improvement of cockpit functions, the disadvantages of the "single core, single screen" form gradually emerge: 1) There is a delay in cross-chip signal transmission; 2) Cost pressure begins to rise.
(2) The centralized domain solution uses a system-level main control chip SOC to control all components in the cockpit. It not only realizes the separation of software and hardware at the software level, but also realizes centralization in hardware. The intelligent cockpit moves from "passive intelligence" to "active intelligence". With the improvement of cockpit intelligence, the diversified and personalized requirements for multi-screen human-computer interaction, voice and other AI functions, as well as the continuous update of OTA requirements by OEMs, the requirements for underlying hardware have increased, and the "single-core multi-screen" SOC solution in the cockpit has begun to enter the public's field of vision.
At present, the cockpit SOC is based on the CPU, and it took less than 7 years for the CPU computing power to increase from several K DMIPS in the past to more than 100 K DMIPS today. The cockpit main control SOC not only needs to handle the requirements of multi-screen scenarios from instruments, cockpit screens, AR-HUD, etc., but also needs to perform operations such as voice recognition and vehicle control. Therefore, the user experience indicators such as the response speed, startup time, and connection speed of the cockpit system directly determine the competitiveness of the car brand. The performance and computing power requirements of smart cars for cockpit SOC continue to rise.
At present, the CPU computing power of Qualcomm Snapdragon SA8155P is about 105K DMIPS, the CPU computing power of SA8195P is about 150K DMIPS, and Qualcomm's fourth-generation cockpit SOC chip SA8295 even reaches more than 200K DMIPS. For domestic manufacturers, the CPU computing power of Huawei Kirin 990 exceeds 75K DMIPS, the CPU computing power of CoreDrive's latest cockpit chip X9U reaches 100K DMIPS, and the CPU computing power of Rockchip's latest smart cockpit chip RK3588M also reaches 100K DMIPS.
Among them, Samsung's Exynos Auto V910, which has been mass-produced, has an AI computing power of about 1.9TOPS. Samsung plans to put the Exynos Auto V920 cockpit chip into mass production around 2025, and the NPU computing power will reach about 30TOPS; Qualcomm's SA8155P chip, which has been mass-produced, has an AI computing power of about 8TOPS, and its fourth-generation cockpit SOC integrates an NPU computing power of up to 30TOPS, which is the cockpit SOC product with the highest AI computing power currently released, and is scheduled to be put into production in 2023. In terms of domestic cockpit SOC, CoreDrive's cockpit products, from mid-level products to supreme products, are embedded with AI computing power, and its X9U product has an AI computing power of 1.2TOPS; Rockchip's latest cockpit SOCRK3588M has an AI computing power of 6TOPS; Geely's CorePower Technology's Dragon Eagle No. 1 has an AI computing power of about 8TOPS.
From the perspective of architectural evolution, in the past, cockpit SOC chips did not have separate NPU units, but as the demand for AI computing power increases, independent NPU units have begun to appear in cockpit SOCs. For example, the 8155 chip does not have an independent NPU core, and AI computing is mainly completed through an 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. The computing power of Qualcomm 8295 chip reaches 30TOPS, and its AI computing power is 7.5 times that of Qualcomm 8155, which is two hexagonal tensor DSPs.
The core essence of "software-defined car" is the decoupling of algorithm and application development from the computing platform. Software is no longer developed based on a fixed hardware, but has the characteristics of portability, iterability and scalability. As the framework and tool chain of artificial intelligence become more and more mature, the accuracy and maturity of the algorithm depend more on the amount of data and the quality of annotation. The iteration speed of the algorithm in the later stage is getting faster and faster, but the iteration speed of the hardware is not so fast. Therefore, software-defined cars are more about utilizing the characteristics of rapid iteration of algorithms or software. After sales, they can expand the functions and performance of the car through OTA to improve the driving experience:
based on the hardware of the chip platform, kernel systems such as Hypervisor and Linux are installed to manage software and hardware resources and complete task scheduling.

Develop and expand various functional software under the AUTOSAR framework, call and process sensor and actuator data, execute autonomous driving algorithms, and realize various application functions such as perception fusion, decision planning, control execution, and HMI.

Based on the SOA software service architecture, the vehicle bottom layer is decoupled and reused to achieve rapid iteration of software functions, and through personalized OTA interaction with car owners, a personalized and differentiated cockpit product experience is created. In addition, in order to cope with the iterative and changeable characteristics of cockpit software requirements, reusability and scalability must be emphasized in the design of the SOA service architecture. At present, many technology companies such as Continental EB, Thundersoft, Neusoft Reach, Huawei, ArcherMind Technology, and Banma Smart Driving have laid out smart cockpit software platforms.

Thundersoft released the smart cockpit platform TurboX Auto 4.5, which is based on SOA architecture and realizes the decoupling of scenarios and services. It can quickly complete the development, change and upgrade of scenario services.
Neusoft Rich has built a universal and standardized software architecture and software platform, which can quickly adapt to the hardware platforms of mainstream SOCs in different markets, and realize the mass production of high-end, medium-end and low-end multi-platform smart cockpits to meet the positioning and needs of different car manufacturers and different models.
At present, the number of cockpit screens is generally one or two, and some slightly more models will use three or four. However, with the increase in the number of vehicle screens and the increase in electronic components in the car (audio, monitoring, etc.), a single chip may become difficult to process this amount of information. At this time, there are two ways to deal with it:
(1) Use chips with higher computing power. However, this method will lead to an increase in procurement and development costs. For example, Jidu uses Qualcomm 8295, and the corresponding chip value will also be higher;
(2) Use a multi-SOC mode to divide the work of the chips. Although one chip with multiple screens is achievable, a large amount of data is accumulated together, which requires more complex algorithms. For example, the current Ideal ONE adopts this method of multiple smart cockpit chips. The Ideal ONE is equipped with a Snapdragon 820A chip and a Texas Instruments Jacinto6 chip. The Snapdragon 820A chip is responsible for driving the underlying system of Android Automotive used by the 16.2-inch central large screen and the 12.3-inch co-pilot entertainment screen, and the Jacinto6 chip is responsible for driving the Linux system used by the LCD instrument panel and auxiliary driving display service.