On March 27, 2022, the Bosch Global Intelligent Automotive Frontier Summit of the Chinese Electric Vehicle Hundred People Society was held.
Bosch: Cross-domain integration is the trend, preparing for intelligent ecology
User experience and orientation continue to drive the development of intelligent connected vehicles. The current domain-centralized electronic and electrical architecture has developed into a cross-domain fusion architecture, which has stimulated the demand for cross-domain fusion central controllers in integrated vehicles, especially the fusion of the two domains of intelligent driving and intelligent cockpit.
Bosch established its Intelligent Driving and Control Business Unit on January 1, 2021. The Bosch Intelligent Driving and Control Business Unit will reduce system complexity and improve the updating speed of vehicle functions through cross-domain software and electronic solutions. Intelligent driving, intelligent cockpit, intelligent connected vehicles and intelligent vehicle controls are our four main business sectors. The Chinese region has over 1,100 employees and R&D centers in Suzhou and Shanghai.
Bosch’s intelligent cockpit domain controller is completely self-developed locally and provides service to customers with Chinese speed. Bosch started pre-research on the cabin-driving integrated platform many years ago. At the software platform level, Bosch is actively developing a cross-domain SOA software platform to support different application requirements of intelligent driving and intelligent cockpit with one middleware. At the hardware architecture level, Bosch is actively promoting the design and mass production of plug-and-play blade type hardware to ensure the scalability of the cabin-driving integrated platform. At the chip level, Bosch is promoting the development of on-board chips and conducting technical exchanges with chip manufacturers. At the electronic and electrical architecture level, Bosch is committed to fully communicating among major machine manufacturers, jointly participating in electronic and electrical architecture planning, and helping major machine manufacturers to achieve a regional central electronic and electrical architecture as soon as possible.
China Information and Communication Research Institute: The automotive industry is in the digital transformation
Whether it is digitalization or informationization, it talks about how to integrate information and communication technology with various industries, thus transforming production methods, business models, and industrial organizational methods to improve quality and efficiency. This is a systematic transformation process. Sometimes, digitalization, networking and intelligentization are also mentioned. At this time, digitalization is a narrow discussion. The overall digital transformation should include all these three contents. It can be said that digitalization and informationization are closely related, but they are also a new stage.
Customization and C2M (Customer-to-Manufacturer) have a long history in the automotive industry. C2M in the digital era means that users are more involved and can participate in the design of some core functions. When Germany proposed “Industry 4.0”, the first consideration was how to achieve large-scale customization. The adaptation of production lines and enterprise systems to large-scale customization is a very important exploration globally, and the automotive industry may be the best in this respect.
Intelligent manufacturing includes existing production lines, research and development, supply chains, and so on. The information automation level of the automotive industry is basically the best in the manufacturing industry, and the introduction of technologies such as artificial intelligence and 5G communication will take it to a new level.
The digital transformation of the automotive industry has two characteristics:- The scope of product connection can determine the extent to which the automotive industry value chain will be extended;
- The depth to which data can be applied depends on the degree to which value can be added through data.
Digital transformation is a systematic change, and the core lies in the transformation of automotive industry values, operations, and business. The organization of automotive enterprises will also face changes, so the strategic design of enterprise digital transformation is very important in this process. Transformation of business under the guidance of strategic design is crucial because it lies in the value of the enterprise. In addition, how to achieve the transformation of operations around the transformation of business is very critical for enterprises in the midst of change.
Jidu Auto: Intelligent Auto 3.0 Will Focus More on Software Security
In the era of fuel and electric cars, innovation in the automotive industry was driven by energy. Now, electric cars are undergoing comprehensive digital upgrades, and intelligence has become a new trend. Artificial intelligence brings technological innovation, efficiency improvement, and experience subversion. Jidu Auto believes that 2023 will be the first year of competition in automotive intelligence, and the era of intelligent auto 3.0 is already here.
In the era of intelligent auto 3.0, the basic concepts of automotive products are: free movement, natural communication, and self-growth. Jidu Auto self-developed the electronic and electrical architecture and domain controller, and fully ensured the overall security of the architecture from the perspective of software and hardware integration. Software research and development, verification of software capabilities, and the safe and stable operation of software are the issues that Jidu Auto is most concerned about.
Cambricon Technologies: Localization of Vehicle Chips and Local Manufacturing and Packaging Issues Still Exist
Cambricon Technologies is a subsidiary of Cambricon Holdings, a leading artificial intelligence chip company incubated by the Institute of Computing, Chinese Academy of Sciences. Cambricon Technologies is committed to building the world’s leading automatic driving chip company.
Cambricon Technologies believes that the trends of automatic driving in the next five years are as follows:
- The installation rate of L2+ automatic driving systems will rapidly popularize and will exist in the long term. The overall penetration rate of L2+ and above in the next five years may exceed 50%;
- L4-level automatic driving solutions in restricted scenarios will gradually land, but there is still a long way to go before large-scale production;
- The closed-loop cooperation of car-road-cloud will further promote the continuous upgrading of driving and riding experience. The scale landing of the intelligent driving system will face multiple challenges.
Cambricon Technologies positioning itself as a vehicle chip company and will work closely with Tier 1 companies, sensor companies, algorithm companies, and OEMs to form a network of cooperative relationships.# 2022 Cold Warbler will have two heavyweight chips officially released. One is the SD5226 series product for the L4 market, which can support on-board training, and the other is the SD5223 chip for the L2+ market. Subsequently, products aimed at other segmented markets will be launched, first of which is the high-end autonomous driving chip SD5226 series product for the L4 market. Currently, L4-level AD domain controllers adopt solutions with 2 or even 4 SOCs, which brings risks and challenges such as complex systems, limited board-level bandwidth, excessive power consumption, and long mass production cycle.
Cold Warbler plans to release the SD5226 series chip solution next year. To meet the market demand, SD5226 further improves its AI computing power to over 400 TOPS, and its CPU’s maximum computing power surpasses 300 KDMIPS. It adopts the 7nm process, independent security island design, and provides the first autonomous driving solution based on a single SOC for the L4 level. In addition, the highlight of this chip is that it can support on-board training. SD5226 supports an on-board self-learning architecture, truly entering the era of intelligent cars 2.0.
Huawei: Software-Defined Cars
Intelligence will be the focus of electric vehicle competition, fundamentally changing the way people interact with cars and the experience, and changing the future pattern of the automotive industry. Through years of practice and summarization, Huawei believes that intelligent cars have four major characteristics: managing watts with bits, driving horsepower with computing power, software-defined functions, and continuous experience optimization.
In the digital architecture of intelligent cars, Huawei provides the basic elements of the intelligent car digital platform iDVP, including the computing and communication architecture CCA, the in-vehicle operating system, and a complete toolchain to reduce the difficulty of software system development. In this architecture, there is a core operating system, Huawei’s self-developed VOS, which is fully compatible with multiple heterogeneous chips and AUTOSAR, and has passed both the car regulation and information security certification. Huawei provides a model-based toolchain, making the migration of the original ECU’s code and the development of a new system very efficient.
Huawei hopes to build diversified cabin experiences based on the Hongmeng car OS. This requires the joint participation of hardware and software partners. We have established partnerships with more than 150 hardware and software partners, and more car applications will be available this year. In the latest cabin demo car, partner’s car skies, electronic rearview mirrors, holographic projection, steering systems, and intelligent health seats have been deployed, and the experience is excellent. We are promoting the commercialization of these scenarios.# Smart Car Industry Development Requires Massive Computing and Software Talents
The development of intelligent cars requires a large number of computing and software talents. We gather talents in the industry and provide them with systematic knowledge, training, and practical opportunities to help them grow. At the same time, talents joining the industry also promote its development, forming a talent training system that combines production, education, and research. Huawei also actively invests in talent training by organizing training camps for more than 10 well-known universities and sponsoring multiple smart driving contests. We have engineers go to campuses to guide students in developing smart driving systems. We have also established a talent certification system for course study, experiential learning, and examinations, promoting the learning and development of talents in universities.
Guo-Qi Zhikong: Launching an Intelligent Platform Architecture that Adapts to Software Definition and Digitally Driven Platforms
The intelligent network-connected automotive operating system we discuss is an overall product architecture. The main purpose today is not to introduce the product, but to highlight several key features that need attention. This includes the hardware and operating system, which are similar to mobile phones, which allows original equipment manufacturers and their tier-one suppliers to independently and efficiently develop intelligent driving applications on top, achieving the so-called software-defined car.
In short, the intelligent driving operating system needs to have several efficient advantages that support high-cost, low-cost application development. It also supports different hardware platforms and uniform operating systems. In addition, China is also promoting network connectivity, requiring that the internal operating system, like a computer kernel, has external network connectivity functions, is safer, not just ASIL D functional safety, but also includes data security and information security, further realizing the car rule level.
Regardless of the time, tens of thousands of new cars always exist with insufficient hardware intelligence. This includes the several billion cars already in service. During the software-defined life cycle, the software of the vehicle must be upgraded and will definitely exceed its own hardware.
Therefore, it is necessary to consider how to break through the boundary of car limits and the extension of sensing on computing borders. Based on the overall framework of the car-cloud integration, the focus is on the car control system, emphasizing the extension of the car control system to be tested on the road. From a calculation perspective, the main calculation unit is found on the road test, vehicle end and cloud side. Therefore, the car control system needs to use edge computing to achieve the extension of the car control system. Here, we need to understand that the intelligent car operating system also includes information and data security, and can logic run in the car and cloud, achieving collaborative computing.
Deng Zhidong, Tsinghua University: Venturing into the Road of Innovation, Development, and Commercialization of Smart Cars
Today, there are mainly three aspects to share:
- Can L3+ pure visual automatic driving be achieved through artificial intelligence supercomputing and pre-trained massive models?
- How to achieve collaborative innovation, development, and commercialization of single smart vehicles and network-connected new infrastructure?
- How to move from L2 assisted driving to L3+ automatic driving, and its technological implementation path.## The first aspect
We are currently seeing Tesla’s L2 pure vision FSD assistance driving, which has achieved a good user experience. With just a few targeted pushes in specific regions, it has gained more than 50,000 ordinary end-users and is currently the world’s largest L2 fleet. Its use of pure vision has advantages in avoiding the technical challenges brought by using multimodal fusion perception.
Since the launch of the FSD beta test version in October 2020, no accidents have occurred. However, it is ultimately an L2 automatic driving assistance, not L3 or L4 pure vision automatic driving. We believe that to improve to L3 or L4 automatic driving, we need to make maximum use of existing data-driven artificial intelligence, including automatic driving application scenarios for limited areas, maximum use of labelled big data and artificial intelligence supercomputing, and pre-trained massive models.
The second aspect
How do we explore a path to collaborative innovation development and commercialization landing? We believe that the key is to coordinate single bike intelligence with the new infrastructure of internet of vehicles.
Single bike intelligence is actually the foundation, and developing it can be said to be a top priority. We need to encourage all kinds of automatic driving solutions and explore and innovate. To pursue single bike intelligence at L4 or even L5 level, encourage L3+ sensor solutions dominated by pure vision, vision or lidar, and explore and innovate all kinds of solutions.
How can we leverage our unique advantages in China’s relatively backward single-bike intelligence to accelerate the promotion of 5G-V2X new infrastructure for L3 and L4 automatic driving, especially the construction of intelligent road networks in limited areas, and explore the path of innovation, development, and commercialization of L3+ automatic driving in China through the collaboration of single bike intelligence and intelligent internet of vehicles.
The last aspect
Let’s explore the technical implementation path for moving from L2 assisted driving to L3+ automatic driving.
According to the new standard, the key difference between L2 and L3+ or assisted driving and automatic driving is based on determining whether there is a safe driver in the driver’s seat. If there is no safe driver in the driver’s seat, but instead in the passenger seat, the back seat, or even outside the car, this is true L3+ automatic driving.
For L3 automatic driving, because there is no driver in the driver’s seat, it may be impossible to introduce human perceptual capabilities into this closed loop. For L4 automatic driving, because it does not require human takeover, there is not only no human perceptual loop but also no human-introduced decision-making loop. Therefore, cancelling the onboard safety driver will undoubtedly force us to face the truly key core technology and promote the large-scale commercial landing practice of automatic driving.The primary basis for revoking the car-mounted safety driver is that whether it is for RoboTaxi, unmanned trunk line and feeder line freight transport, or unmanned public transportation, the main goal is to ensure absolute safety. Why is this a recent major goal? In addition to the aforementioned key core technology breakthroughs such as long-tail problems, 5G applications, AI edges, digital twins, edge-cloud integration, and next-generation AI, this is also a real need for commercial model iteration. Because the cost of using an algorithm engineer as a safety driver in an L2 vehicle is definitely higher than that of a regular ride-hailing driver, this business model cannot be iterated. At the same time, true breakthroughs in key technologies and continuous innovation in business models can effectively promote the landing of China’s autonomous driving industry.
EHang: Setting the Score in the Flying Car Field with its Safe, Smart Air Mobility Solution
There are currently many chips in the car, from gateways, body controls, autonomous driving, cabins, power, and various other aspects of smart cars that require a large number of chips, with hundreds or even thousands of chips.
From a technical perspective, the main form of autonomous driving in the future will still be human-machine driving. Breaking through from L2 to L3 will be a relatively long process, involving the continuous upgrading of software, hardware, and data technology coordination with the autonomous driving system to achieve more functions of autonomous driving and driving experience. EHang’s Atlas A200, which is currently in mass production, perfectly covers levels L2-L3 of autonomous driving.
Autonomous driving chips integrate various functions, including SOC chip systems with functions such as image processing, GPU, DSP, NPU, information security, and functional safety. They can support high-powered autonomous driving chips above L3 level, and are a very complete system.
The development of autonomous driving requires computing power, high-power SOC chips, AI computing platforms, and good image processing capabilities as the basis for autonomous driving evolution. At the same time, the decoupling of software and hardware and the popularization of SOA mean that it is necessary to embed sufficient computing power to provide computing support for subsequent software iteration and upgrading. L4 and L5 autonomous driving may even require over 1,000 TOPS of computing power.
EHang is committed to promoting the development of China’s autonomous driving industry through leading technology development, a complete product system, an open ecosystem, and a flexible business model. We see the opportunities that this era has given us. In the field of smart new energy vehicles, China’s automobile industry has truly taken the lead globally. This kind of leadership actually needs more localized industrial chain support to meet the needs of domestic automakers.
Horizon Robotics: Embracing Centralized, High-Performance Computing
The main basis for revoking the car-mounted safety driver is that whether it is for RoboTaxi, unmanned trunk line and feeder line freight transport, or unmanned public transportation, the main goal is to ensure absolute safety. Why is this a recent major goal? In addition to the aforementioned key core technology breakthroughs such as long-tail problems, 5G applications, AI edges, digital twins, edge-cloud integration, and next-generation AI, this is also a real need for commercial model iteration. Because the cost of using an algorithm engineer as a safety driver in an L2 vehicle is definitely higher than that of a regular ride-hailing driver, this business model cannot be iterated. At the same time, true breakthroughs in key technologies and continuous innovation in business models can effectively promote the landing of China’s autonomous driving industry.With the development of intelligent vehicles, especially the demand for advanced autonomous driving technology, higher requirements have been placed on the overall software computing power and data transmission. Domain controllers and domain architectures can no longer meet the development needs of advanced autonomous driving, hence the birth of a central computing architecture. The characteristics of this architecture include a centralized controller and a domain controller architecture, which separates computation from execution, and is also a more open software development platform.
By virtualizing and containerizing hardware resources, developers can more easily develop software on this architecture, better support 10G Ethernet and high-speed interfaces, greatly improve the computational efficiency and data transmission capacity of the entire algorithm, and better support the development of advanced autonomous driving.
Horizon Journey 5 chip, released by Horizon Robotics last year, is a central computing platform chip that is intelligent throughout the whole vehicle for all scenarios. It is a powerful, high-performance, low-latency, and low-power consumption chip, with 128 TOPS AI computing power. On the Microsoft CoCo (data) set, it can process 1,283 frames per second, which is currently the most powerful computing chip in the industry, and it has only 60 milliseconds of processing delay, which is also the industry’s lowest. At 125 degrees, its power consumption is only 30 watts.
I’d like to also explain the difference between AI chip computing power and real AI performance. The AI computing power we usually refer to is measured in TOPS/Watt (how many TOPS per watt) or TOPS/Dollar (how many TOPS per dollar), which is determined by the entire hardware design and represents the theoretical peak of the AI chip, but it does not represent all the computing power that we can use in actual use.
There is an issue of effective utilization, which depends on the hardware architecture design of the entire chip and the compiler’s ability to compile the AI algorithm to the best fit for the hardware architecture and hardware. Besides these two factors, we also need to consider the third factor, which is the efficiency of the AI algorithm itself. It is represented in FPS/TOPS (how many frames per second can be processed per TOPS), which represents the algorithm’s advancement and complexity, and also reflects in the process of engineering the AI algorithm.
We have achieved industry-wide recognition with our open and win-win business cooperation model. Currently, we are in deep cooperation with more than 17 OEMs, with over 45 front-loading projects being developed. Last year, we shipped more than 1 million Journey chips, making us the only company in China that has achieved mass production of automotive-grade chips.Our positioning as a Tier-2 player is to serve as an enabler in the industry. With the journey chip as the cornerstone, we aim to build a comprehensive ecosystem in the intelligent automotive industry. Currently, we have extensive collaboration with mainstream OEMs and new entrants. Additionally, we cooperate extensively with Tier-1 companies, hardware, sensor and software partners to jointly develop the intelligent automotive industry.
This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.