AutocarMax – Media Under the Hundred People Association Focusing on the Evolution of the Automotive Industry

Author: Zheng Wen, Zhou Changxian

“Know thyself; nothing in excess.” These are the first two sentences of the three inscriptions on the Apollo Temple at the Delphi Oracle in ancient Greece.

These ancient sayings have a sobering effect on people’s secular lives, and the same goes for the rapidly changing automotive industry in this unprecedented century.

However, in the rapidly changing automotive industry, with the emergence of new products, new technologies, and new business models, more and more contradictions and problems are exposed. Some people go with the flow, some take the unconventional path, and the industry moves forward in chaos and controversy.

Recently, at the Geely Auto first Intelligent Automotive Domain Controller Summit, major OEMs and suppliers expounded their understanding and response to hot issues such as domain controller, computing power, and software-defined cars. The exchange and collision of ideas gradually revealed some industry consensus.

AutocarMax attempts to summarize several trend changes and industry thinking on certain common topics from the high-density information in the forum.

The Dilemma and Remodeling of the Electrical and Electronic Architecture

In the past 30 years, the number of Electronic Control Units (ECUs) in cars has grown rapidly, from a few to dozens or even hundreds. Take Volvo as an example. Since 1995, the number of ECUs has grown at a rate of about 4-5 per year.

The phenomenon that arises from this is that the code for automotive software has grown from tens of thousands of lines to billions of lines, making its complexity already surpass that of an ordinary computer. Bosch once predicted that the code for an L2-level autonomous vehicle is approximately 100 million lines, L3 will be between 200 million and 300 million lines, and L5 autonomous driving will require a direct code of 1 billion lines.

As a software engineer, writing a few hundred lines of code may not be difficult, but as the number of lines expands to billions, the maintenance of the entire system becomes extremely difficult and even risky.

From another perspective, the dilemma of the existing electrical and electronic architecture is even clearer.Starting from the perspective of wire harness, the wiring harness of the entire vehicle is becoming more and more complex. “We roughly estimated that currently in a car, if all the wiring harnesses are connected together, they will be several thousand meters long,” said Han Ying, the head of the Electronics and Electrical Architecture at Infineon’s Automotive Electronics Business Unit. “These wiring harnesses have many complicated layouts, and the space is getting smaller and smaller. They need to be manually laid out, and automation production cannot improve efficiency for various reasons, making it necessary for the electronic and electrical architecture to change.”

If we list the system complexity of the current architecture more concretely, it is as follows:

• The network structure is complex, which easily forms isolated islands of information and data, and the gateway and domain controllers become bottlenecks.
• In order to achieve function iteration, more and more ECUs are added, resulting in a large number of ECUs and computational waste, unable to form cooperation.
• In most cases, they are developed by different suppliers, and the framework and software cannot be reused cooperatively, making it difficult to unify OTA.
• In most cases, a user’s functions are distributed in multiple different ECUs, and changing/upgrading a function often requires co-working between multiple ECUs/suppliers, and sometimes there are situations where the bus information is not supported, causing slowness and high cost.
• ECUs are basically black boxes, and external developers cannot program them or define new features through software.
• Hardware upgrades cannot be performed.

The only solution to these problems is to shuffle the deck and re-establish a system with a more reasonable underlying logic. That is to say, abandoning the traditional way of implementing a large amount of repetitive engineering and evolving to a focus on innovation and iteration. At this point, the traditional distributed electronic and electrical architecture has reached its system growth limit, and thus the domain centralized electronic and electrical architecture has emerged.

Taking the automatic driving domain controller as an example, all the sensors related to automatic driving are centralized in this controller, which has relatively complete heterogeneous computing power (it can be simply and roughly understood as enabling different architecture processors to work together), and the software architecture can be continuously improved under a relatively stable hardware system. Other general purpose domain controllers can provide more powerful processing capabilities than traditional MCUs (microcontroller units) and support large-scale software deployment.

Ultimately, whether it is the automatic driving domain controller or the general-purpose domain controller, these application software can drive hardware to continuously upgrade within a relatively stable framework through continuous iterative development.According to data from GAS Research Institute, the current shipment volume of domain controllers in China is about 600,000, and it is estimated that the shipment volume of autonomous driving domain controllers will exceed 4 million sets by 2025. Meanwhile, the shipment volume of intelligent cockpit domain controllers is expected to exceed 5 million, with a compound growth rate of over 50%.

Taking autonomous driving domain controllers as an example, the competition can be divided into four categories: the first category is traditional foreign Tier 1, the second is local Tier 1, the third is internet technology and software companies, and the fourth is car manufacturers.

The advantage of the first category lies in the abundant accumulation of vehicle architecture and chassis technology and leading core technology. The disadvantage is the complexity of business lines, slow development of localized scenarios, and lack of flexibility.

The advantage of the second category lies in providing customized customer development that is more suitable for domestic traffic scenarios. The disadvantage is that there is a bottleneck in the core software and algorithm advantages and a serious dependence on overseas companies.

The advantage of the third category is the outstanding technology capabilities of AI, big data algorithms, etc. The disadvantage is that the platform requires cross-model integration, and the system platform needs to break through interoperability.

In summary, traditional foreign Tier 1 and local Tier 1 have two matching strategies. The former is more inclined to integrate with chip, middleware, and component companies, to provide a complete package of domain controller solutions. The latter is more inclined to integrated and collaborative division of labor. Of course, many companies in China have begun to continuously expand their boundaries.

It’s worth noting that mainstream car manufacturers represented by BMW and Toyota have also begun to develop a new electronic and electrical architecture based on domain controller technology.

Will computing power expand without limit?

As mentioned earlier, domain controllers will be a very important part in driving the integration of smart car software and hardware for a long time to come. It is divided into application algorithms, middleware, system software, chips, and hardware platforms.

Compared with traditional ECU, the number of devices, PIN/wiring/welding points, total power consumption, and EMC (electromagnetic interference) intensity of the calculation platform on smart cars are more than 10 times that of the former. The single-board area is more than 6 times, and the PCB board layers are more than 4 times.

## LanTu Automotive Technology Co., Ltd.’s Director of Autonomous Driving Algorithm Research and Development, Liu Huikai believes that the computing platform’s demand for computing power largely depends on the massive upgrade of the perception system, including the incorporation of Lidar and an 8 million pixel camera, as well as the implementation of complex systems or SOA architecture which puts forward a higher computing power requirement.

“L2 requires less than 10 tops of computing power; L3 requires 30-40 tops; L4 requires over 100 tops, and the industry has yet to define the required computing power for L5,” Liu Huikai pointed out. Currently, the computing power of the computing platform can only support the development needs of some L3 and L4 requirements.

Bishiuzhan, the head of the BU MDC solution department of Huawei’s Intelligent Automotive Solution, believes that, compared to traditional ECU, a computing platform for L2-L5 realizes a challenge of 10-100 times in the fields of hardware and software engineering.

“For example, if the chip computing power ranges from 200 tops to 400 tops, the power is around 100 to 300 watts, and such a significant heat dissipation problem has never been encountered before, posing enormous challenges in engineering such as liquid cooling anti-freezing, EMC complex environment, etc.”

Liang Shuang, Co-founder and CTO of Chaoxing Future, stated that the competition for computing power had already begun, and “for example, this year NVIDIA released the first SoC (system-on-chip) with 1000 tops in the industry, which is more than one order of magnitude improvement compared to Tesla’s FSD single chip with a computing power of 72TOPS.”

In addition, he also noted that there is a trend where domestic players are leading foreign ones, such as Horizon Robotics J5 with a maximum of 128 tops, and Black Sesame Technologies’ A1000Pro with 106 tops, which was announced completed the tape-out just last month.

No one can deny the importance of computing power for centralized computing platforms, but is more computing power always better? Is blindly increasing chip computing power really in line with user thinking? Can the support for intelligent driving system computing platforms only be achieved through chip stacking?

Liu Huikai believes that intelligent automobiles must also be user-oriented. However, throughout the industry’s development, there are some deviations. For example, the weapons race for sensors and chips is becoming increasingly intense; hardware is embedded, software income business model is implemented, but excessive promotion of OTA is ahead of time. The government has begun to strengthen supervision on this.

“Considering the need for occupying technological high ground, capital market financing, and attractive marketing, certain companies are more aggressive in technology deployment and mass production, which is irresponsible for users,” Liu Huikai stated.

Industry insiders seem to have reached a consensus on the series of issues he presented.

“Computing power cannot grow infinitely, and chip PPA (power, performance, area) is crucial. When we designed chips before, we were highly concerned about power consumption. Some car manufacturers say they now use alternative energy, and you don’t have to worry about electricity issues. But we immediately tell them that you don’t have to worry about electricity issues, but you have to worry about heat dissipation,” said Jiang Hanping, General Manager of Product Planning and Management Department at Hubei Xinqing Technology Co., Ltd.

“So, we don’t believe that the unlimited expansion and pre-embedding of computing power are future trends, especially in SoC. We need precise and efficient computing power to adapt to the changes of electronic architecture,” he continued.

Looking at the intelligent cockpit, the overall trend of car CPU chips started with 60K DMIPS (computing power), and in the next five years, the mainstream intelligent cockpit in the market would have more than 60K DMIPS. The curve, which is slowly evolving based on previous cockpits, may still be within the range of 30K DMIPS. In this case, it is challenging to adapt to centralization of computing power. The car’s electronic body requires 10 DMIPS, the chassis 15K DMIPS, semi-automatic driving 350K DMIPS, and intelligent network connection 20K DMIPS.

Liang Shuang acknowledges that chip computing power is essentially a necessary but insufficient condition for an intelligent driving system.

Although everyone knows that better systems require more computing power, currently, everyone talks more about peak computing power. We often see a chip that is poorly optimized claiming to have 10TOPS of computing power, but the actual equivalent of running applications is only 3-4TOPS of computing power. Therefore, he believed that “the design of the computing platform is not just a computing power issue; it is a very complicated issue that requires system optimization design.”

Dong Zuomin, the Director of the Intelligent Driving Architecture Department of Great Wall Motors, summarized the four-dimensional challenges of large computing platforms. They are power consumption, heat dissipation, electromagnetic compatibility, and quality challenges.

Liang Shuang pointed out that for L2+ domain controllers, power consumption needs to be controlled within a range of 30-40 watts. Even if the power is over several hundred or thousands of watts using water cooling, it will have a significant impact on the battery as a power source for vehicles’ endurance.”Making a computational platform essentially involves optimizing under constrained resources such as hardware and power, much like dancing with shackles,” said Liang Shuang. “We are facing increasingly complex systems and a need to deal with sensors that are becoming more complex and rapidly iterating. We have limited resources at hand, so it is very important to make the computing platform easy to deploy, highly energy-efficient, and secure and reliable.”

Challenges brought by software-defined cars

Today, despite some controversy, there is almost a consensus on SDV (software-defined cars). So what value does it bring to car companies and users?

Liu Huikai, Director of Automatic Driving Algorithm Research and Development at Voyah Automobile, believes that the value of SDV to car companies is the ability to improve product value through software upgrades and feature openness. By combining and matching submodules of safety, comfort, power, and entertainment, different application scenarios can be formed, and originally locked-in functions can be flexibly linked with other domains to ultimately achieve a user experience that is tailored to each car.

Cao Bin, General Manager of Neusoft Reach Automotive Technology (Shanghai) Co., Ltd., pointed out several key factors accelerating the development of SDV:

• The appearance of relatively standardized computing hardware platforms: automatic driving and central computing are especially critical.
• Based on a standardized computing hardware, a relatively clear software architecture hierarchy is formed.
• The software development of car companies focuses on application software, and the division of labor of software service companies is gradually refined.

In the process of the above-mentioned technological developments, there is a critical point of “standardization.”

We all know that shared power banks generally provide three types of charging interfaces: Android, Apple, and Type C. If the interface does not match the phone model, it cannot be charged. In software and hardware engineering development, the interface also needs to be matched. At this point, if the interface is not standardized, just matching the interface will require a lot of effort.

In terms of software, with the maturity of middleware and development tools, this part will gradually become thinner and be standardized. It is precisely this standardization that ensures that the applications developed by car companies can continue to iterate and develop without being hampered by engineering adaptation issues.

Some suppliers have good solutions in this regard. For example, according to Li Zhezhe, the R&D Vice President of NuoBo Automobile Technology Co., Ltd., NuoBo’s product design maximizes the level of configurability of the software architecture platform to achieve the goal of adapting to the needs of more car companies with minimal code changes.

In addition, there is hardware-oriented engineering adaptation. Because chips are still evolving, heterogeneous chip forms will exist for a long time. Therefore, this part will become more and more complex, and the workload will become larger, posing challenges to stability and performance.

Apart from standardization, there are also other important factors to consider.

For example, a cabin domain controller needs to adapt to twenty or so different models of vehicles, each with different requirements. “So how to build a domain controller platform that can cover so many complex and different needs requires me to imagine this platform can cover different application scenarios of high-end, mid-end, and low-end during the development process,” said Li Zhezhe.

“At the same time, in the process of platform system upgrade and migration, we try to reuse the hardware and software resources and experience accumulated originally, which can help customers save a lot of development time and cost. Therefore, flexibility, scalability, development cycle, and cost are also very important factors to consider in the domain controller design process.”

However, even if the domain controller bottom layer standardization of the supplier is relatively more convenient than that of the automaker, they still face some practical problems.

Han Ying pointed out that the pressure on automakers is not only in the construction and capability accumulation of the software team but also in the organizational structure.

Now the traditional organization structure of automakers is still divided by functions, such as powertrain, chassis, body electronics, and so on. However, in the future, many developments require functional integration. “This may require automakers to break the boundaries of the existing organizational structure and make certain integration in the organizational structure, which is quite challenging at present.”

“In addition to the well-known functional safety requirements such as active safety and passive safety in the automotive industry, there are also network security issues. Bisio believes that intelligent networked cars are always connected to the network, which means that they face various challenges from the outside at any time, including hacker intrusion.” And this is where you will find that information security, in addition to functional safety, is also facing enormous challenges.”

The old world has gradually collapsed, and the new world is still being created. The future of intelligent automobiles is still a long way to go.

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.