The secret behind the autopilot chip

With the opening of the era of intelligent and connected vehicles, autonomous driving technology has become the focus of attention in the industry. In autonomous driving technology, the chip has the status of a "central hub", which controls the operation of autonomous vehicles through the automatic driving platform. Therefore, the number of suppliers of autonomous driving chips that want to enter the market is also gradually increasing.

At present, "large computing power", "computing efficiency" and "open platform" have become the focus of the development of autonomous driving technology. "China Business Daily" reporter found through market observation that NVIDIA, Horizon, Qualcomm, Mobileye and other chip manufacturers have continuously updated the relevant chips that support higher-level automatic driving, and the cooperation model with car companies has also changed from the original closed "all-inclusive model".

In the field of automatic driving chips, there are not many chip suppliers who can mass-produce and deliver "big computing power" chips to car companies, which also leads to the current models equipped with "large computing power" chips are not many, and the ambitions of car companies to develop self-developed automatic driving chips have gradually emerged, and the intelligent network battle around automatic driving chips has long begun.

"Big computing power" chips become darlings

With the formal implementation of the national standard of "Automobile Driving Automation Classification", the level of automatic driving is developing in the direction of a higher level, and the evolution of the high class also puts forward higher requirements for the computing power of the chip.

At NVIDIA's 2022 GTC (Global Commodity Exchange Center) conference, NVIDIA CEO Jen-Hsun Huang announced that NVIDIA's self-driving chip Orin was officially put into production this month. Official information shows that the computing power of NVIDIA Orin chip can reach 254TOPS. At the same time, Nvidia also announced the Atlan chip, which has a target hash rate of 1000TOPS and is scheduled to be delivered in 2025.

There is no doubt that the current market for autonomous driving chips has undergone obvious changes, and "big computing power" is a major trend. In addition to Nvidia, chips with a computing power of more than 100TOPS have been introduced, such as the Horizon Journey 5, which has released a single chip with a maximum hash rate of 128TOPS; and mobile EyeQ Ultra, which has a maximum hash rate of 176TOPS for a single chip.

Wang Ping, CEO of Cambrian Xingge, also said at the recent Chief Intellectual Officer Conference and The Annual Meeting of Machine Heart AI Technology that "big computing power" will become one of the two major trends of intelligent driving chips. It is reported that this year and next year, Cambrian Xingge will launch two blockbuster chips, one of which will have a high-end intelligent driving chip AI computing power of more than 400TOPS.

Looking at the chips currently mounted on real cars, most of the computing power of a single chip is not more than 100TOPS, such as the NGP auxiliary driving system of Xiaopeng P7, the NOP auxiliary driving system of NIO ES6, and the NOA auxiliary driving system of ideal ONE. This also leads to the current mass production of real vehicles of autonomous driving hovering in the real sense of the L4 level gate can not move forward.

At the same time that the "big computing power" chip has gained the favor of the market, the chip computing power efficiency has also been mentioned again. With a single chip computing power exceeding 200 TOPS, NVIDIA occupies a considerable share of the current autonomous driving market. In the eyes of the outside world, the current autonomous driving chip market is who has "big computing power" that is, who owns the market. But the industry has also raised concerns about computational efficiency.

At the 2022 Electric Vehicle 100 Forum held in Beijing a few days ago, Yu Kai, founder of Horizon, said that in the case of Moore's Law slowing down, it is no longer possible to rely on increasing the density of transistors and improving computing performance. Based on this, Horizon proposes to the chip computing power that "the computing power is not as fast as the calculation", hoping to improve the calculation efficiency of the chip.

In fact, Yu Kai has said many times that Horizon does not simply pursue physical computing power, but also pays more attention to the computational efficiency of deep neural network algorithms on chips. Compared with Horizon Journey 5 and NVIDIA Orin-X, Horizon uses only half of the chip area and half of the computing resources, but it can still get a fairly high computing performance. To some extent, the chips that Horizon pursues are more economical.

Under the premise of the same vehicle-grade chip, "large computing power" and "computing efficiency" are the focus of the industry willing to reach procurement intentions or strategic cooperation. But these two are not contradictory, and chips that have both are more favored by the industry.

The transformation of cooperation models with car companies

Before the current Nvidia market share peaked in half the sky, the industry's more willing to use autonomous driving chips came from Mobileye, which currently belongs to Intel. For a long time, Mobileye has adopted a black box solution, that is, the chip, operating system and intelligent driving system software and hardware are all integrated and packaged to car companies.

At the beginning of the 21st century, the market demand for intelligent driving has not yet been revealed, coupled with the small number of entrants in the field of automatic driving, technological development is in its infancy, and car companies have a blank space in the field of automatic driving. At this time, the all-inclusive model of chip suppliers has become the main choice of car companies. In 2007, BMW, GM and Volvo models were all equipped with chip services provided by Mobileye.

However, such an "all-inclusive model" has become a major factor restricting the personalized customization and differentiated competition of car companies in the current era of intelligent networking. This has also led many car companies to turn to more open chip manufacturers such as NVIDIA and Horizon.

At present, the cooperation model adopted by NVIDIA is based on self-developed chips and operating systems, and autonomous driving companies such as Xiaoma Zhixing and car companies and other relevant parties can design automatic driving software and hardware systems suitable for their own needs on this basis. Compared with the black box solution, this undoubtedly gives car companies greater independent research and development rights.

The cooperation model given by Horizon is more open, including building an open source in-vehicle operating system; authorizing chip IP to car companies to help car companies build chips. The former is called Together OS, horizon in the SOC (System Level Chip) and other development is completed, the middle of the underlying software through the open source OS open mode with the vehicle system development. The latter is called the BPU licensing model, that is, the development of the whole vehicle can penetrate from all levels such as chips, operating systems, and software and hardware systems for autonomous driving.

At the beginning of 2022, Ideal Auto founder Li Xiang said on his personal public social platform that due to the inability to meet the demand for self-research of the full boardwalk of intelligent driving of ideal cars, Ideal Cars stopped cooperation with Mobileye at the end of 2020 and began to use Horizon's J3 chip to carry out full-board road self-research of intelligent driving.

Obviously, with the horn of the era of intelligent networking, one of the competitiveness of car companies in intelligent driving comes from mastering more core chip technology, constantly penetrating to the bottom of the industrial chain, and seeking a high degree of collaboration between software and hardware.

In the specific implementation, the car companies implement the full boardwalk self-research demand for funds is greater. Although in this mode, car companies can control chip design, product functions and research and development efficiency to a greater extent, but due to capital and other investment, only a few car companies such as Xiaopeng Automobile currently use the full boardwalk for self-research.

More car companies outsource the software and hardware development, hardware production and chip solution integration of the automatic driving foundation to the upper-level suppliers, while the automatic driving software part is completed by the car companies to achieve the competitiveness of the intelligent driving personalized customization and differentiated experience.

Domestic chip manufacturers began to break through

At a time when the industry is paying increasing attention to "high computing power" and "computing efficiency", domestic autonomous driving chip manufacturers have also begun to break through.

It cannot be ignored that in the current autopilot chip market, most of the models facing the L4 level use foreign chip vendors such as NVIDIA and Qualcomm, and it is true that domestic chip vendors, such as Huawei, Horizon, etc., should not be underestimated, but how to let more car companies or autonomous driving companies choose domestic chips is a problem in front of us.

Wang Ping said that from the perspective of car companies, it is hoped that car companies can give domestic chip companies more opportunities, through joint development projects, traction of domestic SOCs to become SOCs that are more in line with the needs of car companies, and more use of localization chips to improve supply chain security. In addition, it supports and guides the construction of the ecosystem, and encourages strong cooperation between domestic chip companies, algorithm companies and other enterprises.

At the same time, Wang Ping hopes that enterprises in the semiconductor industry can realize the local landing of advanced process vehicle-grade manufacturing and packaging at the manufacturing end as soon as possible.

Under the premise of continuously improving computing efficiency, Horizon opens up the cooperation circle with more car companies through a more open platform, and understands that car companies want to participate in chip design, software and hardware system development, etc., in an attempt to achieve curve overtaking in more customized and personalized aspects.

It is worth mentioning that at present, most of the manufacturing of autonomous driving chips implements the FOUNDRY system, and chip design companies do not have manufacturing lines.

Jen-Hsun Huang revealed on a conference call in March that Nvidia was interested in considering having Intel foundry chips. "Intel intends to let Nvidia use Intel's manufacturing facilities, and Nvidia is very interested in exploring this possibility," it said. However, the discussion about foundry contracts takes a long time because it involves integrating the supply chain. At present, the above two sides have not yet had a specific timetable for cooperation.

Previously, most of NVIDIA's chip foundry was carried out at TSMC. Therefore, in the closed loop of the whole industry chain, domestic chip manufacturers can also gradually transform to front-end design through the advantages of the original manufacturing.

As Miao Wei, deputy director of the Economic Committee of the National Committee of the Chinese People's Political Consultative Conference, said, if new energy vehicles are compared to the first half and intelligent networked vehicles are compared to the second half, the Chinese automotive industry has achieved great results in the first half, but the decision to win or lose is still in the second half. The development of autonomous driving chips will affect the development of autonomous vehicle technology, how to obtain a greater market share of domestic chips, but still need technological breakthroughs, strengthen cooperation and other multi-dimensional development.