Software-Defined Cars: Computing Architecture and Technology by Dr. Shu Jie from ARM

On June 29-30, the “2021 China Automotive Semiconductor Industry Conference” hosted by Gaishi Auto was grandly held. This conference mainly discussed the current situation of chip shortages in Chinese car companies, the construction of supply chain localization security, the design of in-vehicle chip platforms, the demand and application cases of chips in the fields of autonomous driving and intelligent cockpits, the application of power semiconductors in three electric systems, and chip testing and functional safety, to jointly explore the future development path of the industry. Below is the speech by Dr. Shu Jie, Senior Manager of ARM’s Automotive Market in China at this conference.

Good morning everyone, my name is Shu Jie, and I am responsible for the automotive ecosystem at ARM China. I am very happy to share with you today as a representative of an IP company. My topic today is the computing architecture and technology of software-defined cars. First, let me introduce ARM; some of you may ask how ARM chips are. In fact, we do not directly make chips, but sell IP to customers, who then make the chips.

Dr. Shu Jie, Senior Manager of ARM’s Automotive Market in China

In the semiconductor industry, buying IP to make chips is a very common business model, so you can think of ARM as taking on part of the R&D work for chip companies. For example, customers will hand over CPU, GPU, ISP, etc. to us to help them design, and then they will take this design and incorporate it into their chips to produce the entire chip. The chip IP that ARM provides can meet the different applications of customers, and from the perspective of IP, we bring great flexibility and design freedom to our customers. In addition to making IP products, ARM also maintains a large ecosystem, which is also part of my work at ARM China, where I participate in the maintenance of the automotive ecosystem.

For the ecosystem, we aim to facilitate customers to benefit from it, to better develop their products or applications. Currently, we have nearly 2000 licenses globally, with over 530 customers, and the cumulative shipment of chips based on ARM technology has exceeded 180 billion units globally, with 23.7 billion units shipped just last year, all of which are our customers’ chips.

Looking at the automotive sector, ARM started working in this area relatively early, as early as 1994. As shown in the diagram, many IPs suitable for different applications in today’s intelligent connected cars are listed, such as IVI instruments, infotainment systems, as well as ADAS, sensors, and chassis, etc. For IVI and ADAS, currently over 60% of the SoCs applied in this area are based on ARM technology. As for ADAS, there is no need to say much. If you are involved in chip applications or come from car manufacturers or Tier 1 suppliers, you are familiar with TI, Renesas, NXP, Ambarella, etc., all of which use ARM technology, and their chips all utilize ARM computing cores. In terms of sensors and vehicle body applications, as everyone knows, there has been a significant shortage of chips in this area in recent months, and ARM’s M-core series can also be used here.

For ARM’s customers in the automotive industry, currently, 15 top automotive chip manufacturers are our clients, and ARM has been cultivating new customers, and we also support local chip players to make automotive chips. From 2021 onwards, on average, each new car will have 18 ARM processors, covering applications from IVI to ADAS to sensors, etc. ARM has also been focusing on functional safety, scalable computing power, and interoperability, which can help the automotive industry undergo digital transformation. For example, in autonomous driving, SoCs with functional safety all have ARM computing cores inside. It is also emphasized that information security is critical, and ARM will have end-to-end validated information security solutions.

The previous part is a general introduction to the company. Now back to today’s topic, which mainly covers three parts: software-defined systems; software-defined cars and electronic electrical architecture; and the computing platform for software-defined cars.

When it comes to software-defined systems, what exactly is software-defined? When people first think of the term software-defined, they often think of hardware abstraction, requiring software to be independent of a specific chip, operating system, or vehicle model, so that it is fundamentally unaffected by hardware. Of course, software-defined is developed based on a service-oriented architecture (SOA), allowing services to be delivered independently, published and subscribed separately, and managed as services.

How to enable the development model of software-defined systems? The first approach is to extract software services from a single hardware platform through virtualization and container technology. The second approach is to abstract software services from distributed hardware through a real-time bus.

Currently, the main applications of software-defined are still in data centers, where there will be software-defined computing, software-defined networking, and software-defined storage. The technological support for data centers includes: container technology, virtualization technology, enterprise-level operating systems, etc., all of which enable the application of software-defined data centers.

In the future, software-defined will expand to edge applications, such as software-defined cars and software-defined industrial automation systems, which will also have corresponding technological support, such as heterogeneous computing platforms required for software-defined cars, as well as secure operating systems, real-time operating systems, and real-time program managers, all of which are technologies that meet the applications of future software-defined cars and software-defined industrial systems.

Looking at software-defined cars and electronic electrical architecture, autonomous driving is actually software-defined or software-driven. A Boeing 787 Dreamliner contains 14 million lines of code, while the autonomous driving function of a car may require hundreds of millions of lines of code. On the left is the traditional distributed architecture, and in the middle is the technology architecture that has emerged so far. Yesterday, I listened to Qualcomm’s speech, and their chips indeed have many projects in the cockpit domain controllers. We see that for the domain architecture application, the cockpit will come first, followed by body controllers, while autonomous driving and other domains will be relatively slower, but gradually new products will emerge.

In the future, the development of cars will move towards centralized architecture, which will include high-performance computing units for vehicles. In addition, it will connect to edge modules via Ethernet, and these edge modules will have computing power, which you can consider as edge gateways that can connect sensors and actuators to implement vehicle control. At this year’s Shanghai Auto Show, I saw that domestic Tier 1 suppliers also showcased similar centralized architectures. From the trend, everyone will mention what kind of computing platform to meet the computing needs of centralized architecture, which is the most concerning issue, as there is currently no truly mass-produced computing platform particularly suitable for centralized architecture. Of course, with the development of E/E architecture, cars will increasingly resemble data centers on wheels.

Earlier, I mentioned the need to create software-defined systems. So how does ARM do this? We implement software-defined edge nodes from three aspects. First, we have some work done in the ecosystem and global information security, such as the platform standard CASSNI and the certification project PSA. The second is chip IP in software-defined systems; we will have device-side IP and will also support heterogeneous system product IP combinations. The third is cooperation with ecosystem partners, such as open-source communities or commercial software providers, to work together on some solutions. For example, for software-defined automotive applications, we will develop digital cockpit solutions and also work on ADAS or autonomous driving-related solutions.

Specifically, what is the strategy driven by platform standards or ecosystem standards? You see this is the CASSINI project. Software-defined applications in CASSNI will have continuous applications, meaning that the cloud-native experience must be good, and we will use it to meet the applications of software-defined cars.

PSA, which is first applied in IoT. Information security will have additional requirements for vehicles, so we will provide best security practices and standard-based APIs in PSA, allowing the guidelines for information security to be conveniently used in vehicles, which is the purpose of PSA certification. We also have the System Ready project, which defines architectural standards and compliance plans before tape-out to ensure the portability of operating systems and hypervisors on automotive chips. This means doing the necessary certification work before chip tape-out to ensure that the chip can support the programs and operating systems that are already in use in the market after it is released. Of course, specific solutions will also be mentioned, including software standards, open-source software stacks, and solutions with commercial software companies to assist in the realization of software-defined cars.

The previous part covers some solutions based on ARM ecosystem standards. So what are ARM’s core products? On the one hand, there will be AE IP targeted at automotive applications, which we call automotive-enhanced IP. In addition, there will be a “safety-ready” plan for automotive applications. For our AE products, all IPs are specifically designed for automotive applications, and many features are also aimed at future applications that cars may need. We will implement lock-step technology to meet functional safety requirements, and the lock-step structure is transparent to software, so that everyone does not need to spend too much effort on software. Of course, we are constantly expanding more IP products to meet automotive applications.

Let me talk about the “safety-ready” plan, which is a multi-year project aimed at reducing customer design investment and accelerating their deployment, allowing customers to get their chips out as soon as possible. Some of our IP products have passed certification, which can help customers more easily or quickly pass functional safety certification. In China, if you want to create a product for automotive use, the time for functional safety certification is very long, but if there is IP support, it will facilitate everyone in passing this certification. Our IP products are also included in Safety Ready.

The Cortex-A78AE is the last automotive-grade CPU under our V8 architecture, offering a 30% performance improvement over the previous generation and meeting various workload requirements, supporting functional safety ISO26262.

If you want to create a software-defined solution, it is first inseparable from the ecosystem and partners, which include chip customers, open-source communities, commercial software providers, autonomous driving technology providers, and industry alliances. The second is the software platform and tools, which include standard platform architecture, open-source development platforms, commercial development platforms, middleware-based application support players; standard tools and development environments are also indispensable, as it is difficult to create chip solutions without them. Finally, it is related to the computing core, which I categorize as system architecture. In addition to CPUs that everyone is familiar with, many chip players will also mention XPUs, which refer to applications for data processing beyond CPUs, such as ISPs, VPUs, and DPUs, all of which belong to the XPU category and need to be considered in the system architecture. This system architecture is designed for automotive use, so it must also address functional safety and information security. Since it is a software-defined car, it is also inseparable from OTA.

The computing core here not only includes real-time MCUs but also single-threaded high-performance big-core CPUs and multi-threaded CPUs. For the accelerator part, machine learning is definitely needed, and customers will also need GPUs for computer graphics processing and auxiliary calculations, which also need to be considered in the system architecture.

The third part is the computing platform used for software-defined cars. As mentioned earlier, our ecosystem needs different players to participate, but to truly create a chip based on ARM solutions, here we have Mali-G78AE, NPU, Mali-C71AE. A very important aspect of ARM solutions is the security island. In addition, there is the information security part; when I introduced the previous content, I mentioned that information security will have significant performance improvements in ARM V9 architecture CPUs. Besides these IPs, the others are provided by third-party partners.

Here I have an example; NVIDIA chips have received a lot of attention in the past year, and this image comes from their official website. First, look at the computing platform on the left, which mentions that the chip itself supports centralized architecture and can also support software-defined cars, deeply integrated with autonomous driving. It also supports containerization, functional safety, and information security. The many logos below represent various companies involved; this Orin chip has been chosen by these companies as the next-generation computing platform chip.

Recently, NVIDIA released a conceptual design for the Atlan chip, which will be mass-produced in 2025, retaining CPU, GPU, accelerator, and security island, but adding Bluefield (smart network card). This is the first time I have seen a company include this technology in automotive SoCs from publicly available conceptual SoCs.

I have included news interviews with the founder of XPeng Motors, who has expressed many sentiments shared by automakers and Tier 1 suppliers. He stated that fewer chips mean lower costs, greater computing power, and simpler management. Many companies are developing chips, but whether these products will still exist five years from now or whether they will be integrated into other products remains uncertain.

Finally, to summarize, if we are to use software-defined automotive systems in the future, we actually need solutions similar to current data centers. ARM’s computing technology with functional safety can assist in the realization of software-defined cars. ARM’s software and hardware partners can help developers implement software-defined applications, as they have many tools and solutions that can assist everyone. That concludes my presentation. Thank you all.