The centralization of automotive grade computing platforms and the unlimited expansion of computing power is not a future trend

Reprinted: Gaishi Automobile Author: Jie Quanmin

With the evolution of automotive electronic and electrical architecture from distributed to centralized, domain controllers with wider control range and stronger computing power are now experiencing hot development. Driven by the trend of software defined cars, the direction of multi domain integration is becoming increasingly clear, and automotive grade computing platforms are moving towards centralization, becoming the focus of attention for related enterprises.

The computing platform is moving towards centralization

To meet the rapidly iterating user demands, software defined cars have become an industry consensus and are considered a key transformation direction for related enterprises to win the future. To achieve this, computing platforms, especially centralized computing platforms, are crucial.

Recently, at the 2021 Intelligent Vehicle Domain Controller Innovation Summit hosted by GAC Motor, Zeng Jienan, an expert in the basic software platform of SAIC Zero Beam Software Branch, stated that intelligent vehicles now face three needs: meeting the needs of thousands of users, continuously evolving user experience, and active perception and decision-making of vehicles.

To meet these three needs, Zeng Jienan believes that one of the key factors is to achieve the transformation from distributed to centralized at the hardware level, which requires a centralized computing platform to provide computing power foundation. Hardware provides us with a computing power foundation, allowing us to develop more algorithms and functions on it

Bi Shuzhan, the head of the BU MDC Solution Department at Huawei Intelligent Automotive Solutions, also pointed out in his speech at the summit that the development of intelligence poses many technical challenges. Among them, the application software on the car will be based on SOA architecture (service-oriented architecture), which is characterized by decoupling, modularization, and flexible deployment. The upper layer software will be updated quickly to meet the needs of end users, while the experience and functionality need to be iterated quickly.

Of course, the key issue is to have a solid intelligent driving computing platform base, so that the upper level application software can provide better and faster iterations. Bi Shuzhan stated that in order to achieve the development of upper level software SOA architecture, a centralized computing platform is needed at the bottom level, with domain centralized ZOA architecture (platform hardware centralized architecture) being the main approach at this stage.

ZOA and SOA technology trends; Image source: Huawei keynote speech PPT

Liu Huikai, Director of Autonomous Driving Algorithm Research and Development at Lantu Automotive Technology Co., Ltd., holds the same view. In his speech, he mentioned that the implementation of software defined cars improves the user experience of vehicles and extends their lifecycle. Additionally, as the car is used for longer periods of time, its personalization level also increases. "The most important aspect of implementing software defined cars is to have a centralized controller or create a highly integrated computing platform to support these concepts

From this perspective, in the era of software defined cars, computing platforms need to move towards centralization, and its importance has been recognized by many enterprises. As Yang Ke, Vice President of Guoqi Zhikong (Beijing) Technology Co., Ltd., said in his speech at the summit, computing platforms have become the focus of a new round of technological competition.

Need strong computing power support

No one can deny the importance of computing power for centralized computing platforms. Bi Zhangchang clearly pointed out in his speech that one of the characteristics of centralized computing platforms is their high computing power.

At this domain controller innovation summit, according to Jiang Hanping, General Manager of Product Planning and Management Department of Hubei Xinqing Technology Co., Ltd., in the distributed architecture, most modular functions are stacked through ECU units, which is a 1+1 process. The requirements for chip capability are relatively low, focusing on reliability and security. MCU requires 50 DMIPS (executing millions of integer operation instructions per second). Now we are starting to use domain controllers, which integrate chips that are close to functional and information security to form a domain where decisions can be made. In this case, the MCU has achieved 2k DMIPS.

After the domain controller, there is domain fusion. For example, Tesla has three domains, and in this case, the computing power has already reached the SoC level, with a computing power of 5-20K DMIPS. At present, central computing concentrates a large amount of computing power and places distributed control on the ECU side, resulting in computing power approaching 50-300K DMIPS.

Trends in electronic and electrical architecture and chip computing power; Image source: Theme speech PPT of Xinqing Technology

For the computing platform's demand for computing power, Liu Huikai believes that this largely depends on the huge upgrade of the perception system, including the installation of laser radar and 8 million cameras. In addition, the implementation of complex systems or SOA architecture also puts forward higher computing power requirements for the computing platform. According to its disclosure, the intelligent driving perception systems of domestic automakers basically adopt a strong perception route of multiple redundant sensors including "laser radar+millimeter wave+camera".

For every level increase in the level of autonomous driving, the computing power increases by an order of magnitude. It is generally believed that the computing power required for L2 is less than 10 TOPS, L3 is 30-40 TOPS, and L4 is over 100 TOPS. Currently, there is no clear definition of the computing power required for L5 in the industry Liu Huikai pointed out that the current computing platform's computing power can only support the needs of some L3 and L4 development.

He also mentioned that there are still some challenges in the rapid development and functional iteration of intelligent driving systems: firstly, they need to process massive amounts of data, which is costly, and the monthly amount of data generated by bicycles is very large; 2、 For training and simulation, strong computing power support is required.

Infinite expansion of computing power is not a future trend

In terms of centralized computing platforms, the importance of computing power is beyond doubt, and in this context, many enterprises have strengthened their layout in this area.

At the Domain Controller Innovation Summit, Liang Shuang, co-founder and Chief Technology Officer of Chaoxing Future, admitted that the arms race of computing power has already begun. "For example, this year Nvidia released the industry's first SoC with a computing power of 1000 TOPS, which is more than an order of magnitude higher than Tesla's FSD single-chip computing power of 72 TOPS. In addition, we see that there is also a trend in China to surpass foreign players, such as the Horizon J5 with a maximum of 128 TOPS, and the Black Sesame A1000 Pro with a 106TOPS chip that just announced completion of chip production last month

From the perspective of computing platforms, the launch of these chips provides a richer and more reliable selection. However, Liang Shuang reminded that the industry needs to consider whether the support for intelligent driving system computing platforms can only be achieved through chip stacking?

The answer is clearly negative.

Although the intelligence of automobiles requires stronger computing power, Jiang Hanping said, "Computing power cannot be said to grow infinitely. Chip PPA (power consumption, cost, and area) are all very critical. When we used to make chips, we paid great attention to power consumption. Some car manufacturers said that now it is new energy, and you don't have to worry about electrical problems, but you should know that even if you don't have to worry about electrical problems, you still have to worry about other issues such as heat dissipation. Therefore, we don't think that the infinite expansion and embedding of computing power is the future trend, especially on SoC. We need precise and efficient computing power to adapt to the changes in electronic and electrical architecture

Liang Shuang also pointed out that the computing power of chips is essentially a necessary or insufficient condition for intelligent driving systems. Although everyone knows that implementing a better system requires more computing power, the computing power that people are talking about more now is peak computing power. We often see a chip with poor optimization claiming to have 10TOPS computing power, but the actual application that runs out is equivalent to only 3-4TOPS computing power. Therefore, we believe that computing platform design is not just a problem of computing power, but a very complex problem that requires system optimization design

There are still other challenges in various aspects

As mentioned earlier, computing platforms not only need to solve the problem of computing power, but also face other challenges. In the keynote speech of the Domain Controller Innovation Summit, Dong Zuomin, Director of Intelligent Driving Architecture at Great Wall Motors, mentioned that the large computing platform mainly faces four dimensions of challenges, namely power consumption, heat dissipation, electromagnetic compatibility, and quality challenges.

In terms of power consumption, according to GAC Motor's understanding, usually more computing power is required to pay for more power consumption. Liang Shuang pointed out that the scenario of the car is limited. Taking L2+domain control as an example, its power consumption needs to be controlled within the range of 30-40 watts. Even if the upper water cooling system reduces the power consumption to several hundred watts or even kilowatts, for vehicles with batteries as the power source, the impact on the range is also very significant.

Based on this, Liang Shuang said, "Building a computing platform is essentially a problem of optimizing under resource constraints such as hardware and power consumption, which is a bit like dancing with shackles. We are facing increasingly complex systems and need to deal with increasingly complex and rapidly iterating sensors. The resources at hand are actually limited, and it is very important to make the computing platform easy to deploy, energy-efficient, and safe and reliable

Challenges faced by in vehicle computing platforms; Image source: PPT of Chaoxing Future Theme Speech

Bi Xianchang also believes that computing platforms face many challenges. He admitted that computing platforms, especially those targeting L2 to L5, have achieved challenges in hardware engineering and software engineering that are 10-100 times greater than traditional ECUs.

For example, when the computing power of the chip ranges from 200 TOPS to 400 TOPS and the power ranges from 100 to 300 watts, it means that it faces significant challenges in heat dissipation. In addition, it also faces huge challenges in engineering issues such as liquid cooling anti condensation and EMC complex environments.

Software engineering will also face many problems, such as how to decouple the surrounding service-oriented form based on different upper layer support for SOA scheduling systems? How to decouple and isolate the data and management aspects? How to achieve deterministic low latency in operating systems?

In addition, there are security challenges. In addition to the well-known functional safety requirements such as active and passive safety in the automotive industry, there is also network security. Intelligent connected vehicles are always connected to the internet, and at this time, they are constantly facing various external challenges, including hacker intrusions. At this time, you will find that in addition to functional safety, information security is also facing huge challenges, "Bi Shuzhan added.

Partial response plans and countermeasures

From the above, it can be seen that computing platforms are indeed facing many severe challenges nowadays. So, how can relevant companies respond to these challenges?

In response to this, Bi Shuzhan stated in his speech at the summit that Huawei can provide a specialized computing platform. According to its disclosure, there are currently two platforms in the industry. One platform may come from different manufacturers for chips, hardware, operating systems, and middleware, known as a combination platform. The specialized computing platform it refers to is provided by a company for chips, operating systems, and middleware.

Regarding the advantages of specialized computing platforms, Bi Shuzhan pointed out that if each component comes from different manufacturers, it is very difficult to communicate and coordinate problems encountered during mass production, which will generate a huge workload and lower efficiency. In addition, once a new demand comes from the host factory, it may require the implementation of the operating system, middleware, and even at the chip level. This cycle planning is very difficult to coordinate, and the response cycle also needs to be calculated on a semi annual basis. However, each layer of the specialized computing platform is provided by one company, so there is no such problem. Of course, this is just one of the characteristics of specialized platforms. According to its introduction, specialized platforms also have the characteristics of high performance, high security, and fast response.

According to GAC Motor's understanding, Huawei can provide a series of products for different scenarios, which it calls "unified platform architecture, serialized hardware, shared software, and continuous iteration of functions". This platform supports smooth upgrading of upper layer applications. The specific product models include: MDC 810, with 400+TOPS computing power, which can meet L2+, L3, L4, L5 application scenarios; MDC 610， Provide computing power of 200+TOPS, targeting L4 scenarios, mainly for passenger cars; MDC 210， Has 48TOPS computing power, targeting L2+scenarios; MDC 300F， Targeting commercial vehicle scenarios such as mining trucks, high-speed logistics, and industrial parks.

Huawei Intelligent Driving Computing Platform Solution; Image source: Huawei keynote speech PPT

As mentioned earlier, it is crucial to achieve a computing platform that is easy to deploy, highly efficient, and secure. Liang Shuang believes that the technical solution to such a difficult problem also needs to be achieved through software and hardware collaboration.

As for Chaoxing Future, its idea is to "focus on neural network computing acceleration processing as the core, optimize and even design hardware friendly models through model compression and structure search from the perspective of software collaboration, including customized acceleration for other calculations besides neural networks, significantly reducing the computing power overhead of general-purpose units such as CPUs and shortening the processing latency of computing platforms. After processing all energy efficiency optimizations, optimization will also be carried out on the basic software environment, and ultimately a heterogeneous computing platform will be provided to customers to support software and hardware technologies

According to Liang Shuang, the high-level autonomous driving computing platform released by Chaoxing Future in May this year can support multiple sensor access and processing, up to 20 4K cameras, 3 Gigabit Ethernet ports, and 12 CAN-FD. In the intermediate computing configuration scheme, Nvidia Xaiver SOC is adopted to meet the computing power requirements, while introducing heterogeneous schemes FPGA and MCU to meet user computing power requirements while addressing other computing needs besides neural networks. At the same time, a lot of work has been done on functional safety. We will also provide comprehensive drivers, basic software environments, and toolchains to truly enable users to achieve plug and play. We have now implemented the application of autonomous driving computing platforms for multiple vehicle models, levels, and scenarios with many customers

Chaoxing Future High level Autonomous Driving Computing Platform; Image source: PPT of Chaoxing Future Theme Speech

Of course, not only these two companies, but also numerous companies have laid out their layouts around computing platforms, and their solutions are diverse. However, they all lead to the same goal, ultimately aiming to better meet the needs of end users and gain profits as a result.