Introduction

On October 12, 2021, at the “Promoting the Sustainable Development of the Vehicle Semiconductor Industry” summit of the China Electric Vehicle Hundred People Association held in Nanjing, Wang Ping, CEO of Horizon Robotics, gave a speech on the topic of “Smart Breakthrough of Single-car and Cloud-edge Cooperative Driving”.

During the speech, it was mentioned that “Horizon Robotics is developing a 7nm-based SoC product with computing power greater than 200TOPS, and will achieve SOP on vehicles through various automotive certifications in the future.”

Despite being a newly-established company, Horizon Robotics’ first product is an intelligent driving AI chip with 7nm process and computing power greater than 200 TOPS. With many questions in mind, Jiuzhang Autonomous Driving recently interviewed Mr. Wang Ping, CEO of Horizon Robotics.

Main Body

Jiuzhang Autonomous Driving: Horizon Robotics is involved in cloud-edge-end chip-related business. What was the consideration behind setting up subsidiary, Horizon Robotics-Xingge, to enter the field of in-car driving and develop intelligent driving AI chips?

Horizon Robotics-Xingge: First of all, we believe that the in-car field is a very promising market. The automotive industry has a large scale, and autonomous driving is a great application scenario for AI. From the perspective of industrial application of artificial intelligence, autonomous driving is an indispensable market.

The significance of this field is different for the country and society, as intelligent vehicles are not just consumer goods, but also involve factors such as data security and social security. Therefore, it is necessary to localize high computing power and advanced process chips for intelligent driving.

Finally, speaking mainly from the perspective of Horizon Robotics itself, before being listed, we mainly focused on developing cloud-edge-end AI chips. Setting up Xingge Technology to enter this field was because we had accumulated enough knowledge, talents and financial resources to support our entry into the field at the right time, and we felt that we made the right choice.

Jiuzhang Autonomous Driving: In terms of volume, the AI chips in the in-car field are much smaller than those in the smartphone industry. There are already many powerful manufacturers in the market, and some OEMs are planning to develop their own AI chips. Does Horizon Robotics-Xingge’s direct entry into the in-car field pose a risk, and how are you evaluating this risk?

Horizon Robotics-Xingge: First, I believe that the market for chips in the in-car field might not match the scale of the smartphone market in terms of the number of chips. However, the value of a single chip installed in a car is increasing every year, making it a large incremental market. For example, the high-performance chip, OrinX, developed by NVIDIA, costs several hundred dollars per chip, and NIO ET7 requires the installation of 4 of these intelligent driving chips.The intelligent driving control chip market is a very promising market. Global new car sales are approximately 90 million units per year. In China, sales are close to 30 million units. In the future, intelligent driving AI chips are expected to become standard equipment, leading to an increasing market penetration rate.

Secondly, I believe that it is not easy for automakers to develop their own AI chips. Although Tesla has already developed its own AI chip, it is only for its own use and accounts for a small proportion compared to the size of the entire automotive industry.

In the future, there will be few OEMs like Tesla that can independently develop and successfully apply chips, because chip development requires talent and related technology accumulation, which takes a long time. I think professionals should handle specialized matters to optimize the automotive industry division of labor and development.

“Cambricon”: Cambricon’s positioning is to develop universal and high-performance chips; that is, what is the extent and what are the characteristics of its universality?

We call our car-mounted intelligent driving chip a heterogeneous SoC chip, which includes not only AI modules but also CPU, GPU, DSP, and ISP processing modules.

When we talk about universality, we specifically refer to the AI module’s universality on SoC; it has a universal AI software stack or basic software platform that allows OEM car manufacturers or algorithm companies to flexibly and conveniently move or develop their algorithms on it. In other words, it has good flexibility and compatibility.

On the contrary, the so-called specialized AI modules do not have universality in software platforms. Their software stacks have specialized optimization for certain chips or algorithms. Although they can reduce power consumption and improve performance to some extent, they also have an adverse effect.

Because it is a specialized module, the algorithms that can be adapted are limited, which means that when OEM or algorithm companies need to move or develop other algorithms, they need to spend a lot of time and effort on adaptation work, and the workload of software development may be so large that it could even be impossible. For example, even if Mobileye opens up the AI section to OEM manufacturers, it is still difficult for anyone to run the perception algorithm on its chips.

“Cambricon” started relatively late in the field of car installation. Some domestic chip companies already have mass production and vehicle installation experience, and they are also building their ecosystems. What are the areas that Cambricon needs to catch up on, or what challenges will it face?The Song of Cambrian Journey: Firstly, although we started making chips for automobile applications relatively late, Cambrian’s previous experience in cloud and edge computing chips has provided us with transferable capabilities that can be applied to the field of automobiles. Therefore, we do possess a certain technological foundation. Speaking of whether we started early or late, it is actually relative, or in other words, it depends on different modules.

Secondly, I believe that there is still a long way to go in the field of AI chips for intelligent driving in automobiles, from driver assistance to true autonomous driving, and we predict that it will take at least 10-20 years or even longer.

Earlier, we asked what the gap was for Cambrian Journey. First of all, regarding the capabilities of designing AI architectures with high computing power and universality, we can directly rely on our parent company’s cloud and edge computing chip architecture design capabilities. Secondly, we have accumulated enough in the area of basic software. The only area of inadequacy is, I think, the level of requirements concerning automotive-grade functional safety and reliability. This is a new challenge that requires us to improve in a timely manner. Although the requirements for cloud and edge computing and automotive applications differ in terms of capabilities, it is still difficult, but we are confident in overcoming this challenge.

Jiuzhang Autonomous Driving: As you mentioned earlier, Cambrian Journey makes AI chips for automobiles, and some of the capabilities of its parent company, Cambrian, making AI chips for cloud and edge computing can be directly transferred and reused. Can you specifically describe what these capabilities are that can be transferred and reused?

The Song of Cambrian Journey: In fact, we can divide the design of automotive chips into three major parts: the chip architecture design, the basic software platform design, and the design requirements for automotive-grade applications.

Our initial goal was to make a large-sized, high-performance intelligent driving chip. The capabilities of designing architecture for large-sized and high-performance chips, as well as the ability to develop basic software, were already possessed by Cambrian prior to the formation of Cambrian Journey, and for these two areas of capability, we can mostly transfer and reuse them directly.

Of course, Cambrian’s previous products were mainly designed for industrial products, or higher-end industrial products with industrial-grade specifications. There are higher requirements for automotive-grade chips, such as in terms of reliability and functional safety, and they need to be adjusted and optimized according to the requirements of automotive-grade applications. However, the core foundational modules are the same.

Jiuzhang Autonomous Driving: Earlier you mentioned that the universality of the AI module in Cambrian Journey’s intelligent driving AI chip for automobiles is mainly reflected in the AI module. So, what are the technological barriers and challenges in developing a universal AI module?

The Song of Cambrian Journey: Firstly, a large number of software personnel are required. Currently, Cambrian has around 1,500 employees, of whom roughly 60% are software personnel. Why is this? This is because creating a universal AI hardware architecture to adapt to different networks requires adaptation to different operators. Therefore, everyone who develops a universal AI processor faces this problem.# About the Operator of Visual Perception

There are approximately 1,000 operators related to visual perception. If we want to make it universal, we need to devote a lot of effort in adapting these 1,000 operators. However, if we make a specialized AI perception module, we don’t need to follow such a complicated process. It only needs to use a portion of the network, perhaps only dozens of operators, and adapt them accordingly. The other 900 or so operators are not needed. The technology of creating an AI module itself is straightforward, but the challenge is how to adapt these different algorithms to our hardware. To achieve both generality and network utilization while considering low power consumption is crucial and also the difficult part.

This universality puts two abilities to the test: processor architecture and how to use one to support different types of algorithms and operators. Additionally, the other capability is the ability of the algorithm toolchain – how to make an easy-to-use SDK to quickly adapt to the new algorithm operators.

Furthermore, the process of upgrading to enhance computing power and performance is essential. The most significant restrictions come from the process. The higher the process, the more difficult it is to design and produce. Take the 7nm process as an example; it requires cooperation between chip design companies and wafer manufacturers to validate all IP related to 7nm. Increasing one process step requires significant effort, such as a large amount of human and financial resources.

Jiuzhang IMa: Besides the use of advanced processes to enhance computing power, can improving the architecture also increase computing power? Most chip architectures today are still based on the von Neumann architecture. Will new architectures emerge in the future that will overthrow this traditional architecture, such as Stored-Program Computing Integrated Chips? What are the prospects for such chips?

Hanwuji Xingge: In essence, increasing computing power requires increasing the number of transistors, and increasing the number of multiplication and accumulation operations. No matter how much the software architecture is improved under the traditional architecture, it can only improve hardware utilization, not actual computing power. Hardware remains fixed once the chip is manufactured, determining the number of multipliers and transistors inside.

From a macro perspective, the only way to significantly increase computing power is to enhance the process. Unless we don’t use deep learning or CNN networks, which are different concepts, altering the AI processor architecture only changes the degree of adapting to different networks.

New architectures may appear, but at least for now, we have not seen their widespread applications or practical productization.# Traditional Architecture Limitation

The traditional architecture separates processors and memories, which need to communicate through buses, thereby limiting the performance. The so-called “STT-MRAM on Logic” or “Compute-InMemory” means that the computing unit and memory are integrated. In other words, the same processing unit can perform both computing and storage functions, which will bring a significant change to the entire architecture compared to the previous one. This is not a small feat – all upper-level software and systems need to be changed and adjusted accordingly.

HanGu HiSilicon’s Strategy

For passenger cars, Level 4 autonomous driving has a long way to go, and in the short to medium-term, L2-level intelligent driving configurations will still dominate. For this level, high-performance, cost-effective chips with 14nm or 16nm processes have already met the requirements. Why is HanGu HiSilicon choosing to enter the market with high computing power 7nm chips? Will HanGu HiSilicon’s 7nm process advantage become a cost disadvantage for the L2 level automotive chip market?

First of all, the reason we started with large computing power chips is that we saw a strong demand for these kinds of chips among OEMs. Especially for new energy vehicles, as their batteries and motors use liquid cooling, the heat dissipation problems of large computing domain controllers can be relatively easily solved. Secondly, the use of a “digital engine” – large computing power chips instead of a fuel engine in electric vehicles – requires OEM’s to put effort into developing differentiation to give consumers a better driving experience.

Third, starting with large computing power can help us quickly grab the market and shape our brand influence. It is clear that there are currently relatively more chip manufacturers who are focused on low computing power, making the market competition more intense. Developing high computing power and advanced process autonomous driving chips, which have a higher technological barrier, is more in line with our brand positioning and makes it easier to enter the automotive market field.

At the L2 level, using a 7nm process may result in a cost disadvantage, so we are also prepared to deal with slightly lower process technology. However, our product plans come step by step, starting with large computing power chips and then gradually exploring other areas.

Conclusion

In the future, computing power will be one of the key performance parameters for intelligent automobiles. Demand for high computing power AI chips will continue to rise, and to address the pain points of the vehicle semiconductor and supply chain industries, the entire industry must work together and cooperate to enable large computing power, advanced process AI chips to be localized as early as possible.

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.