院士邬贺铨：车路云协同的关键在于数据 未来更应关注“算力压缩”

Core Viewpoint - The development of intelligent transportation systems relies heavily on vehicle intelligence, but it has limitations in complex environments, necessitating a shift towards vehicle-road-cloud collaboration for comprehensive situational awareness and improved traffic management [1][2]. Data and Computational Challenges - Training Level 5 (L5) models requires 17 billion kilometers of data, with at least 100 million kilometers of real roadside data, which is challenging to collect [2] - Each vehicle generates approximately 1GB of data per second, leading to a total of around 12GB of data per vehicle during travel after compression [2] - Currently, only 1% of traffic data comes from real roads, with 90% sourced from closed roads and simulations, highlighting a significant data scarcity issue [2] Data Annotation and AI Solutions - The high cost of data annotation necessitates the development of AI-based methods to replace manual processes, although a portion of original data (10%-20%) should be retained to prevent data obsolescence [3] - The demand for computational power in intelligent driving is proportional to model parameters and training data, while inversely related to training duration and GPU utilization [3] Vehicle Communication and Computational Requirements - Different vehicles have varying application needs, requiring capabilities for direction indication, predictive actions, and communication with other vehicles and infrastructure [4] - Minimum computational requirements for vehicle levels L2, L3, L4, and L5 are 4-10 Tops, with L5 needing up to 1000 Tops, which current vehicles cannot support [3][4] Focus on Computational Compression - Strategies for reducing computational demands include utilizing generative AI techniques and attention mechanisms to streamline calculations [5][6] - The deployment of large models in the cloud can lower usage barriers, allowing for user-specific data adjustments [6] Network Infrastructure Development - Upgrading existing 5G networks and establishing V2X networks is essential for supporting intelligent transportation systems, requiring collaboration among various stakeholders [7] - A national unified V2X operator is proposed to standardize and scale network construction, with an estimated investment of 400 billion yuan to achieve comprehensive coverage and enhance urban traffic efficiency [7]