CICC: Optimistic about the development of the medium- and long-term autonomous driving industry Manufacturers with advantages on the data side are expected to take the lead in achieving breakthroughs

Zhitong Finance APP learned that CICC released a research report saying that automatic driving is one of the important scenarios for AI landing, after 2012, the rapid development of deep learning technology has led to the rapid progress of automatic driving technology, in recent years, Transformer big model and other technologies have further improved the ability of automatic driving algorithms, and the trend of large computing chips and self-built AIDC in car factories that began to land in 2022 also provides underlying support for large models, CICC is optimistic about the development of the medium and long-term automatic driving industry. Vendors with advantages on the data side are expected to be the first to achieve breakthroughs.

The main points are as follows:

Deep learning is a watershed in the development of autonomous driving technology, in recent years, the Transformer big model is an important industrial trend, and algorithms are the core capabilities of major manufacturers. As "software 2.0", deep learning is the core driving force for the development of autonomous driving, which greatly improves the accuracy of image recognition at the perceptual level and effectively improves the algorithm capabilities at the decision-making level. CICC believes that in the future, deep reinforcement learning and traditional expert rules will coexist extensively. The large model represented by Transformer has become the mainstream route due to its obvious advantages over CNN in massive data pre-training, robustness, and generalization capabilities, and its emergence has also enabled enterprises with data advantages to form an efficient data closed loop, which is expected to bring subversive breakthroughs to autonomous driving algorithms.

Data is the winner or loser who determines the capacity of mass production. Deep learning is data-driven AI that requires massive amounts of data to be "fed." Data acquisition relies on on-board sensors: Tesla (TSLA. US) insist on choosing a visual camera-led sensor solution, complete the data closed loop, build a deep barrier; other manufacturers choose lidar, camera and other multi-sensor fusion solutions, lidar price reduction speed is the key to the development of the industry, CICC believes that the future of lidar as a security redundancy has a realistic need. In addition, the amount of data is the core element of the competition between the two types of schemes, and the level of automation, labeling quality and efficiency of data labeling are the key competition points.

At this stage, the computing power determines the upper limit of autonomous driving technology, and the giants have laid out first. Computing power in advanced autonomous driving precedes algorithms, and the current automatic driving AI chip is presented with NVIDIA (NVDA. US) as the leading "one super and many strong" pattern, Tesla leads the transformation of the domain controller architecture, catalyzing the change of the pattern of chip manufacturers and the mass production process, 2022 is the first year of the opening of the automatic driving large computing chip. CICC expects that the ratio of large computing power and energy efficiency in the future chip competition will be the primary competitive factor, and the importance of an open ecosystem will be increasing. At the cloud level, AIDC provides the large computing power needed to process large amounts of backhaul data, which is essential for training large models, and giants such as Tesla have a natural advantage in technology and money. Similar self-owned e-commerce business needs gave birth to Amazon (AMZN. US), Alibaba (BABA. US) and other CPU era cloud computing giants, CICC believes that AI giants such as autonomous driving have a natural first-mover advantage in the GPU/ASIC cloud computing market.

This article is compiled from the WeChat public account "Zhongjin Dianqing", authors: Chen Xingyu, Deng Xue, etc.; Zhitong Finance Editor: Chu Yunwei.