2024年中国端侧大模型行业研究：算力优化与效率革命 如何重塑行业生态

Industry Overview - The end-side large model is defined as a large-scale AI model running on the device side, typically deployed on local devices such as smartphones, IoT, PCs, and robots, with smaller parameter sizes compared to traditional cloud-based large models, enabling direct computation on the device without relying on cloud resources [4][11] - The end-side large model market in China reached 800 million yuan in 2023 and is expected to grow to 2.1 billion yuan in 2024, driven by strong downstream demand, particularly in the smartphone and autonomous driving industries [4][17] - End-side large models offer significant advantages in cost, energy efficiency, reliability, privacy, and personalization compared to cloud-based inference, making them suitable for personalized AI applications [4][15] Industry Drivers - The shift of AI processing to the edge is driven by cost advantages, energy efficiency, reliability, performance, latency, privacy, security, and personalization [16] - The cost of AI inference is significantly higher than training, and edge terminals can provide leading energy efficiency, especially compared to cloud-based solutions [16] - End-side AI processing can offer comparable or even better performance than cloud-based solutions, especially during peak demand periods, reducing latency and improving reliability [16] - End-side large models inherently protect user privacy by keeping queries and personal information on the device, which is crucial for enterprise and workplace applications [16] - Personalization is enhanced as digital assistants can tailor services based on user preferences and behaviors without compromising privacy [16] Market Size and Growth - The end-side large model market in China is projected to grow from 800 million yuan in 2023 to 2.1 billion yuan in 2024, with a CAGR of 58% [17] - The growth is driven by the rapid development of the smartphone and autonomous driving industries, which are integrating advanced AI functionalities to enhance user experience and safety [17][18] - The global AI chip market, which supports end-side large models, reached 20 billion USD in 2021 and is expected to exceed 70 billion USD by 2025, with end-side AI chips becoming a significant growth driver [18] Industry Chain Analysis - The upstream of the end-side large model industry includes AI chip suppliers, cloud computing service providers, and data service providers, while the midstream consists of end-side large model technology companies and end-side technology enterprises [20][21] - The downstream applications span across various industries such as finance, automotive, healthcare, education, and entertainment, with end-side large models being deployed in devices like smartphones, IoT devices, and robots [21] - Model compression technologies, such as knowledge distillation, pruning, and quantization, are crucial for reducing the parameter size and computational complexity of end-side large models, enabling them to run efficiently on resource-constrained devices [22][24] Cost Structure - The cost structure of end-side large models includes hardware costs, such as AI chips, which are essential for accelerating deep learning tasks and reducing energy consumption and latency [27] - R&D costs, including personnel and GPU expenses, are significant, with the average annual salary of a deep learning engineer in the US being approximately 140,000 USD, and high-end GPUs like NVIDIA GeForce RTX 3090 costing around 1,500 USD [28] - Other costs include management, operational, and marketing expenses, which are necessary for the sustainable development of end-side large model projects [25][26] Industry Scenarios - The adoption of end-side large models is influenced by industry-specific demands for data security, privacy protection, the prevalence of smart devices, and the maturity of AI large model technologies [29][31] - Industries with high data security requirements, such as finance, healthcare, and government, are expected to see significant growth in end-side large model applications [30] - The increasing prevalence of smart devices, such as home health monitoring devices, is driving the demand for end-side AI applications, particularly in sectors like education and healthcare [32] Business Scenarios - End-side large models are particularly suited for applications requiring low latency, real-time computation, and high levels of personalization, such as AI smartphones, autonomous driving, and robotics [33] - In AI smartphones, end-side large models enhance privacy and reduce latency by processing data locally, improving user experience in applications like voice assistants and image recognition [34] - Autonomous driving benefits from end-side large models by enabling real-time decision-making and improving safety, as the models can process data locally without relying on cloud connectivity [34] - Robotics applications, especially in home service and healthcare, leverage end-side large models to provide personalized services and improve efficiency by processing data locally and adapting to user behavior [35] Competitive Landscape - Leading large model companies, such as SenseTime, Alibaba, and FaceWall Intelligence, are leveraging their technical expertise and ecosystem advantages to dominate the end-side large model market [42][43] - These companies are using advanced technologies like algorithm optimization and model compression to overcome the computational limitations of end-side devices, enabling complex AI functionalities to run efficiently on mobile and IoT platforms [43] - The competitive landscape is expected to intensify with the integration of cross-domain technologies, such as natural language processing, computer vision, and edge computing, driving innovation in end-side large model solutions [44] - Ecosystem building and innovative cooperation models, such as joint R&D and data sharing agreements, are becoming key factors in shaping the competitive landscape, with end-side large models driving the growth of the AI chip market, which is expected to reach 2.286 billion units globally in 2023 [45] Policy Environment - The Chinese government has positioned the AI industry as a core national strategy, with supportive policies for AI infrastructure and generative AI, creating a favorable environment for the development of end-side large models [40][41] - Policies such as the "Management Measures for Generative AI Services" and the "Overall Layout Plan for Digital China Construction" provide regulatory guidance and support for the development of end-side large models, emphasizing data security, privacy protection, and technological innovation [41]