How to install domestic autonomous driving chips in cars?

The pace of popularization of autonomous driving is accelerating.

On January 12, the Ministry of Industry and Information Technology issued a document stating that 3.521 million new energy vehicles will be sold in 2021, of which passenger cars equipped with combined assisted driving systems accounted for 20% of the new car market. Two years ago, the penetration rate of L2 assisted driving was only 3.3%.

The development of smart cars is inseparable from the brain of the car - the autonomous driving chip. Among domestic autonomous driving chips, Black Sesame Intelligent, CoreDrive Technology, Horizon Robotics, and Huawei are the main manufacturers. In addition, Leapmotor and NIO also have plans to develop their own chips.

This Wednesday (January 12), Heizhima Intelligence and Xinchi Technology held their own media communication meetings in Beijing to introduce the latest business progress.

Heizhima Intelligence was founded in 2016 and focuses on the autonomous driving chip track, while Xinchi Technology, which was founded two years later, has a wider range of chip business, covering autonomous driving, smart cockpits, central gateways, high-reliability MCUs, etc.

Autonomous driving chips have attracted much attention in the capital market. Heizhima Intelligent has completed six rounds of financing, with investors including SAIC Group, Weilai Capital, Xiaomi Changjiang Industry Fund, etc. On January 12, it also announced that it had received strategic investment from Bosch's Boyuan Capital; CoreDrive Technology has completed four rounds of financing, with investors including Hechuang Capital, Sequoia Capital, Lenovo Capital, Matrix Partners China, and Guokai Ronghua.

For a long time, the autonomous driving chip market has been dominated by Mobileye and Xilinx. In 2019, Mobileye's market share reached 70%. In recent years, with the intelligentization and electrification of automobiles, giants in the consumer electronics era such as Nvidia and Qualcomm have also expanded into the field of autonomous driving chips.

In 2022, Nvidia's autonomous driving chip Orin will be launched with the delivery of NIO ET7 and Zhiji Auto, and Qualcomm's Snapdragon Ride autonomous driving platform will also be launched in mass production of Great Wall Motors.

This year is also the year when domestic high-computing autonomous driving chips are put into cars. This time, domestic autonomous driving chip manufacturers have the opportunity to stand on the same starting line with international giants.

1. How much computing power is required for autonomous driving?

In the past few years, computing power has been regarded as one of the simplest and crudest standards for measuring an autonomous driving chip, presenting a state of "arms race".

In 2014, Tesla Model S was the first to use Mobileye's first-generation EyeQ chip, with a computing power of only 0.256TOPS (a unit of processor computing power, 1TOPS represents one trillion operations per second), which is almost the same as the computing power of today's smart access control systems; the Model 3 released by Tesla in 2019 is equipped with its self-developed autonomous driving chip, with a computing power of 144TOPS; in 2020, NIO ET7 and Zhiji Auto were launched with NVIDIA Orin chips, with a computing power of up to 1000TOPS.

How much computing power does an autonomous vehicle need? Yang Yuxin, CMO of Black Sesame Intelligence, said that this question can be viewed from two aspects. First, chip computing power will become a key point to attract users to buy. Now some car manufacturers have said that consumers have asked about computing power in 4S stores.

Second, with the trend of software-defined cars, pre-embedded computing power will become a common choice for car manufacturers. Even if the computing power is not used now, it will be used to upgrade the car's functions through OTA in the future. In fact, the current mainstream car models are all L2 or L2+ level, and a computing power of up to 100TOPS is sufficient.

It is worth mentioning that the computing power of autonomous driving that everyone often talks about is actually the computing power of the neural network acceleration unit - NPU. However, the performance of an autonomous driving chip, in addition to NPU, also includes CPU, GPU, ISP, DSP, etc., which is a contest of comprehensive capabilities.

Like the CPU function of a computer, the CPU mainly handles logical operations in autonomous driving scenarios; the NPU belongs to AI operations, realizing functions such as neural network model detection; the GPU is a graphics processing unit, which renders images based on camera data; the ISP is responsible for image signal processing, and the DSP is responsible for digital signal processing, handling large-scale data fusion and data operations.

The chip with the highest computing power at present of Heizhima Intelligence is the Huashan II A1000Pro, which has a computing power of 106TOPS and is designed for L3 autonomous driving. The other two chips have slightly lower computing power, namely the Huashan II A1000L, which has a computing power of 16TOPS and is designed for L2 autonomous driving; and the Huashan II A1000, which has a computing power of 58TOPS and is designed for L2+ autonomous driving.

However, high-computing chips also mean higher costs and are often more suitable for mid-range and high-end or even flagship models. According to Black Sesame data, high-computing chips with a computing power of more than 100TOPS will be installed in vehicles in 2022-2023.

For example, many popular models such as NIO ET7 and Zhiji L7 equipped with NVIDIA Orin chips will be delivered this year.

Xu Chao, vice president of CoreDrive Technology, mentioned: "There are different models for a car model. For example, the lowest-end model has nothing, while the highest-end model has everything. These two models often have the lowest sales, and their combined sales do not exceed 20%. The most popular model sells 80%."

Compared with NVIDIA's high-computing chips, the computing power of Xinchi Technology's current V9 series autonomous driving chips is not high, only 1 TOPS, but according to Dr. Tao Sheng, head of Xinchi Technology's autonomous driving, it can already achieve L2.99 autonomous driving.

Tao Sheng said that L1 is a bygone era, L2+ is an emerging era, and L3, L4, and L5 will be an era in the future. Xinchi did not pursue high computing power right away, but provided the most cost-effective solution based on a feasible premise. At present, V9T has partners using it for forward-looking L2+ functions.

CoreChi believes that 2023 will be the year when L3 will enter mass production, and has made product layout accordingly. According to the plan, CoreChi will launch an autonomous driving chip with a performance of 10 to 200TOPS this year - V9P/U, which has higher computing power integration and can support L3 level autonomous driving.

Unlike Heizhima Intelligence, which focuses on autonomous driving chips, Xinchi has a product matrix approach. In addition to the V9 series of autonomous driving chips, there are also the X9 series of smart cockpit chips, the G9 series of central gateway chips, and MCU vehicle control chips.

2. Tier 2 needs a flexible and open business model

As the status of chips in the automotive supply chain increases, the division of roles in the traditional automotive supply chain is also quietly changing.

In the traditional automotive supply chain, the roles of car companies, Tier 1 (first-tier suppliers), and Tier 2 (second-tier suppliers) are in a vertical linear relationship. Car companies do not need to contact Tier 2 chip manufacturers. Someone joked that "people in car companies don't even know which way the door of the chip factory faces."

However, three variables gradually broke this situation. The first is the shortage of automotive chips. AutoForecast Solutions previously released data showing that as of December 9, the global reduction in automobile production due to chip shortages in 2021 has reached 10.272 million vehicles, of which the Chinese automobile market reduced production by 1.982 million vehicles in 2021, accounting for 19.3% of the world.

In order to ensure production and supply, the chairmen of many automobile companies personally lead teams to purchase chips.

The second change is that different car brands used to have their own separate supply chains, but now a car company has several sub-brands. The OEM will merge the supply chain from the perspective of supply chain integration and supply chain security. Therefore, the OEM needs to communicate deeply with chip manufacturers.

The underlying changes are in technological change. The automotive electronic and electrical architecture is changing from the traditional ECU to the current "domain controller" architecture, and will evolve to "central computing + regional control" in the future. Under this trend of change, on-board chips are gradually becoming the brain of smart cars.

In order to better define their needs, car companies have to directly contact chip manufacturers and even cooperate in R&D. Xu Chao, vice president of CoreDrive Technology, said that the traditional automobile development process does not require the participation of chip companies, and most of them use standard components provided by Tier 1, and the electronic systems of mainstream models are less differentiated; now car companies have strengthened in-depth exchanges with chip companies, and chip companies have entered the market more than 16 months in advance, and the demand for differentiation has increased.

These changes are also gradually reshaping the roles of the industry chain. The traditional OEMs, Tier 1, and Tier 2 are beginning to transform into a system of OEMs to Tier 1.5 suppliers. OEMs used to connect with Tier 1, but now they connect with chip suppliers, hardware suppliers, and integrated testing service providers.

Under this model, how to more efficiently meet the customized needs from car manufacturers has become a problem that chip manufacturers must solve, which requires a more flexible and open business model.