本地AI处理

Search documents
手机芯片,大变局
半导体行业观察· 2025-06-07 02:08
Core Viewpoint - Leading smartphone manufacturers are facing challenges related to local generative AI, standard smartphone functionalities, and increasing data interactions between mobile devices and the cloud, which put pressure on computing and power consumption [1][3]. Group 1: Mobile SoC Design Challenges - High-end smartphones utilize heterogeneous architectures in their System on Chip (SoC) designs, where multiple modules perform different tasks collaboratively [3]. - The rapid evolution of AI networks and diverse AI model requirements complicate mobile SoC design, necessitating support for both large-scale cloud models and efficient local models [3][4]. - The integration of AI capabilities into chips is becoming less challenging due to advancements in tools and processes over the past five to ten years [6]. Group 2: AI Processing and Architecture - The design focus is shifting towards optimizing power consumption in parallel processing of graphics, general computing, and AI operations [5]. - AI accelerators in mobile SoCs may include GPUs, NPUs, or high-end ASICs, with NPU becoming central for low-power tasks [7][8]. - The rise of multimodal models and generative AI tools adds complexity to design, requiring flexible and efficient computing structures [10]. Group 3: Local vs. Cloud Processing - Local processing of AI applications, such as facial recognition and photo editing, is preferred to reduce latency and enhance data privacy [13]. - Despite the increase in local AI processing, some tasks still need to be executed in the cloud due to battery and power limitations [13]. - The balance between local and cloud processing will be an ongoing challenge as AI models become more efficient [13]. Group 4: Key Trends in Mobile SoC Design - Three key trends driving changes in mobile SoC design include rising analog demands, the proliferation of visual and AI applications, and the high-performance computing requirements of modern applications [15]. - Designers must consider both hardware and software perspectives to remain competitive, emphasizing the need for collaborative efforts across disciplines [15].