Current assisted driving technology is undergoing a profound transformation from modularity to end-to-end paradigm. The traditional solution relies on high-precision maps and FP16-level computing power, mainly supporting high-speed scenarios; Segmented end-to-end solution turnoverLightweightMaps usually require E-level or greater computing power and more than 800G bandwidth to achieve urban navigation; The end-to-end solution of the large model requires 10E or more computing power and more than 3.2T bandwidth, and uses FP8/FP16 mixed-precision computing to promote the implementation of global piloting and active interaction. In the technology iteration, the geographic modality realizes the transition from all-element HD Map to lightweight SD Map, and the unit data volume is reduced by more than 90%. However, the three core challenges of insufficient algorithm intelligence, cloud computing power cost constraints, and lagging map updates continue to restrict the process of large-scale inclusiveness of technology.
On July 22, 2025, Wang Zezhong, director of Tencent Automotive Solutions, said at the 8th Assisted Driving Conference in 2025 that in the context of the rapid iteration of assisted driving technology, the current industry is facing many challenges and pain points, such as insufficient algorithm intelligence, cloud computing power constraints, AI algorithm limitations, and lagging map data. As a technology company, Tencent is deeply empowering the self-driving ecosystem through cloud services, map data and AI capabilities, helping car companies reduce costs and improve efficiency. Relying on massive data ecology and cloud-graph integration services, Tencent is promoting the evolution of assisted driving from modularity to end-to-end, realizing the efficient combination of data closed-loop and compliance, and creating a sustainable innovation engine for the industry.
Wang Zezhong | Director of Automotive Solutions at Tencent
The following is a summary of the speech:
The evolution route of assisted driving business and industry challenges
Assisted driving technology is undergoing a transformation from traditional perception modularity to end-to-end large models, and this process is accompanied by significant challenges. The traditional stage relies on small network computing power, which is mainly used for assisted driving on high-speed and urban expressways, and requires high-definition map data support. At present, segmented end-to-end requires E-level computing power or greater network bandwidth, which is suitable for urban NOA scenarios and uses light map data. In the future, the one-stage end-to-end of large models will require 10E-level computing power, support global navigation and active interaction, and rely on light maps defined by algorithms.
Source: Speaker material
At present, there is still room for improvement in algorithm intelligence, effective perception, and control algorithms, and there is still room for improvement in the computing power pressure brought by NOA across the country, the limitations of AI algorithms, and the contradiction between rapid function expansion and lag in map data updates. In terms of algorithmic intelligence, because algorithmic intelligence has not yet fully matched the human level, the cockpit manned navigation map information cannot provide a good experience for self-driving, so self-driving requires a special car map instead of a human driving map. In terms of computing power and AI algorithms in the cloud, with the advancement of intelligent driving equality, vehicles of about 100,000 yuan also need to have national NOA functions, and in the case of insufficient effective perception, the computing power of domain control chips should be used as little as possible to establish real-time complete vector maps. In terms of map data, the rapid expansion of functions makes it difficult for map merchants to keep up with the data, and how to make full use of the data returned by car companies has become an urgent issue to be solved in the industry.
Source: Speaker material
The core role of map data and Tencent’s solution
Map data is the cornerstone of assisted driving, and Tencent provides full life cycle support through the intelligent driving cloud map product matrix. In the initial stage of R&D, Tencent provides compliance cloud and data compliance solutions, and combines data collection and annotation services based on training data such as HD, point cloud, HD Air/SD, etc.; The mass production expansion stage covers HD AIR/SD/HD data products, lane-level guidance services and vehicle-side data engines; The operation stage focuses on data backhaul compliance, map construction and integration, and data mining, such as lane-level events and behavior analysis.
Source: Speaker material
Tencent Maps’ advantage lies in massive data coverage: 1 billion+ user positioning calls every day, 11 million+ road kilometers coverage (95% in China), and 26 million+ daily vehicle activity to ensure that road network elements are updated at a daily level. The self-driving online data service provides 5D+HA road network, supports vehicle-side composition and over-the-horizon maps, covers SD attributes of the whole road network, light map HD Air lane-level information, small data volume but rich information, and supports day-level updates and flexible expansion.
Source: Speaker material
Cloud computing power and compliance cloud services reduce costs and improve efficiency
The lack of cloud computing power is the bottleneck of the scale of assisted driving, and it has created the concept of “dedicated cloud” intelligent driving cloud zone advocated. At the same time, Tencent has built four intelligent driving cloud zones to support efficient cross-regional transmission, remote disaster preparedness and elastic expansion to ensure compliance and efficiency in the whole data process. The deployment model includes public cloud, private cloud and distributed cloud integration to adapt to the different needs of car companies. For example, the AI infrastructure model built for car companies integrates cloud services, mass production data backhaul and annotation algorithms to accelerate R&D.
Source: Speaker material
Tencent Cloud’s case shows that the comprehensive cost reduction solution significantly reduces the operating costs of car companies, such as increasing the utilization rate of computing power resources by more than 30%. The cloud-map integrated service accelerates through one-stop training, helps the data closed loop, solves the problem of fast function expansion and data cannot keep up, and ensures that assisted driving takes into account cost and performance when the popularity of 100,000 yuan models.
Source: Speaker material
AI algorithm improvement and data closed-loop system
The lack of AI algorithm intelligence is the bottleneck of the assisted driving experience, and Tencent combines cloud graph capabilities to promote algorithm upgrades. To achieve national NOA when there is still room for improvement in effective perception insufficient perception and regulation algorithms, Tencent provides a perception computing power improvement solution, integrating traditional small models, large device models and event models, and optimizing algorithms using map data. Cloud and map integration to build a full closed-loop data service system: through vehicle-side data backhaul (such as road procurement data), cloud processing (automatic element extraction and quality inspection) and algorithm matching (laser fusion positioning), a R&D integration link is formed.
Source: Speaker material
For example, AI large models drive the co-construction of environmental experience layers, such as bumpy road suggestion speed and dangerous road section information, to improve the efficiency of assisted driving decision-making. Tencent is hereNew energyNOA maps accounted for 49.01%, leading the industry, confirming the effectiveness of the scheme.
Source: Speaker material
In terms of deployment mode, Tencent adopts “one cloud and multiple forms” to provide automotive companies with the most suitable AI infrastructure, including distributed cloud, private cloud, private cloud, etc., taking into account security compliance and open compatibility. Its intelligent driving cloud zone also builds a global automotive storage and computing network, supports advanced technologies such as serverless GPU resource elastic scheduling, and can provide efficient data services and computing power support for assisted driving research and development.
Source: Speaker material
Vehicle-cloud integrated data closed-loop link
Tencent has built a closed-loop data link for vehicle-cloud integration to comprehensively improve the efficiency of intelligent driving research and development. In the data collection and return process, the road acquisition data backhaul includes sensor data and vehicle module data to ensure the real-time and completeness of the data. With the help of AI annotation engine, automatic annotation of 4D time and space, combined with manual quality inspection, greatly improves annotation efficiency, annotation efficiency by 16 times, and reduces cost by 22.6 times. Through the AI retrieval engine, it can achieve millisecond-level retrieval of millions of samples, quickly build a long-tail scene library, and provide rich sample data for model training.
Source: Speaker material
In terms of model training and deployment, after simulation evaluation and model training and tuning, mass production model deployment and OTA upgrade are realized to ensure that the assisted driving model can be continuously iteratively optimized and optimized. This closed-loop link covers the entire life cycle from data generation to model application, providing one-stop training acceleration through the combination of cloud and graph, helping intelligent driving R&D improve efficiency, and at the same time accumulating rich experience in million-level mass-produced vehicle services, providing compliant privacy desensitization SDK and other services for the car side.
Source: Speaker material
Future outlook and industry synergy
As the large-scale coverage of intelligent driving has become a rigid need, algorithm upgrades and map reforms are being promoted, and Tencent will continue to focus on cloud, graph, and AI capabilities and work hand in hand with industry colleagues. In the future, assisted driving technology will develop in the direction of higher precision and intelligence, with higher requirements for the real-time and richness of map data, and Tencent’s intelligent driving cloud map will continue to optimize the multi-layer data form, realize the flexible release and rapid update of map elements, and meet the needs of assisted driving in different scenarios. In terms of computing power and algorithms, with the continuous advancement of AI technology, Tencent will further improve the performance of AI annotation, retrieval and other engines, reduce R&D costs, improve model training efficiency, and help the industry break through technical bottlenecks.
In the face of industry challenges, Tencent will deepen cooperation with car companies, algorithm companies and other partners to jointly explore key issues such as data compliance utilization and efficient scheduling of computing power, promote the popularization and development of assisted driving technology, so that a safer and more convenient assisted driving experience can benefit more users, and jointly move towards a bright future in the era of self-driving.
(The above content comes from the keynote speech “Combining Tencent’s Capabilities and the Autonomous Driving Industry” delivered by Wang Zezhong, Director of Tencent Automotive Solutions, at the 8th Intelligent Assisted Driving Conference in 2025 on July 22, 2025.) ）