Cloud-based autonomous driving also uses large models for inference, but focuses on emphasizing V2X technology. The vehicle-side model will combine cloud instructions such as roadbed sensors and high-precision maps to control the vehicle.
According to Zhongtong Finance, GF Securities issued a research report stating that a single vehicle intelligent solution requires a large amount of cloud computing power to support model training. It is recommended to continue to pay attention to the progress of the intelligent solution related companies' computing construction. V2X has higher requirements for the construction of roadbed and cloud infrastructure, such as RSU and edge cloud MEC; at the same time, it is recommended to pay attention to innovative roadside perception solutions. Focusing on the inside of the car, no matter what scheme, the interconnection and intelligence of automobiles are the trend, so the T-Box carrying communication functions and the supporting vehicle-mounted antennas are important investment directions. The number of vehicle body sensors has increased and the response speed has accelerated, should pay attention to the upgrading progress of in-vehicle Ethernet.
The iteration of AI technology is driving the development of autonomous driving. In 2021, Tesla demonstrated a BEV+Transformer architecture autonomous driving model at AI Day, part of the underlying principles of which are similar to large language models such as GF's ChatGPT. BEV and Transformer can fully utilize the rich spatial environmental information provided by BEV and the multi-source heterogeneous data modeling ability of Transformer to achieve more accurate environment perception, longer-term motion planning and more global decision-making. It is worth noting that the autonomous driving model depends on large-scale training computing power support. Tesla continues to invest in computing power. According to its official Twitter, Tesla is expected to have a total computing power of 100 exaFLOPS by the end of 2024 to train and fine-tune FSD. Computing power is the main theme of the development of large-scale models, and the optical module is the base of computing power. The significant increase in GPU communication flow in AI training networks has promoted the construction of computing power network infrastructure and driven the continuous increase in demand for high-speed optical modules.
With excellent infrastructure and overall planning capabilities, China is expected to be the first country where cloud-based autonomous driving can be commercially successful, and it also uses large models for inference, but focuses on emphasizing V2X technology. China, represented by Baidu's RoboTaxi, is accelerating the commercialization landing.
The progress of communication technology is an important catalyst for accelerating the landing of autonomous driving. (1) Inter-vehicle communication: after the technology competition between C-V2X and DSRC, C-V2X currently has a prominent advantage in China. Compared with DSRC, C-V2X has advantages such as high bandwidth, wide range, fast speed, and low cost, which can reduce the probability of resource collision in high vehicle density scenes. In the field of autonomous driving, C-V2X can provide more accurate information perception and stronger networked intelligence, realizing the real-time sharing and interaction of information and coordinated perception and control needed for autonomous driving. (2) In-vehicle communication: the trend of upgrading the CAN bus to Ethernet is obvious. Vehicle-mounted Ethernet has the advantages of large bandwidth, low latency, low electromagnetic interference, and uniformity of link connection forms, which reduces the types of whole vehicle links and reduces costs. It can couple heterogeneous intelligent cabin, autonomous driving and vehicle networking, and GF Securities judges that vehicle-mounted Ethernet is expected to partially replace the CAN bus technology market.
Investment advice: Pay attention to investment opportunities in the communication sector of the three ends of car, road and cloud in the V2X era. A single vehicle intelligent solution requires a large amount of cloud computing power to support model training. Therefore, GF Securities recommends paying attention to the progress of the intelligent solution related companies' computing construction. V2X has higher requirements for the construction of roadbed and cloud infrastructure, such as RSU and edge cloud MEC; at the same time, GF Securities recommends paying attention to innovative roadside perception solutions, such as Wuhan Ligong Guangke's grating array technology. Focusing on the inside of the car, no matter what scheme, the interconnection and intelligence of automobiles are the trend, so the T-Box carrying communication functions and the supporting vehicle-mounted antennas are important investment directions; the number of vehicle body sensors has increased and the response speed has accelerated, should pay attention to the upgrading progress of in-vehicle Ethernet. Finally, high-precision positioning is a rigid demand of autonomous driving cars, and it is recommended to continue to pay attention to the Beidou satellite navigation industry chain.
Risk warning. The construction of computing infrastructure is not as expected; the intensification of trade frictions; the risk of reduced international cooperation due to geopolitical risks; the risk of macroeconomic performance falling short of expectations.