GenAI 存储解决方案第 6 部分：边缘计算

Core Insights - Semiconductor innovation is central to improving technology in daily life, with companies adopting flexible strategies to achieve higher returns on investment capital (ROIC) [1] - Collaboration within the supply chain has become increasingly essential in the semiconductor sector [1] - Hardware is evolving to accommodate GenAI applications and changes in user interfaces [1] Group 1: Market Trends and Innovations - From 2025 to 2027, various players are driving advancements in the semiconductor industry [2] - Dynamic Random Access Memory (DRAM) solutions have their own advantages and disadvantages, improving bandwidth, latency, speed, capacity, and power/heat characteristics, but also presenting cost and timeliness challenges [3] - Customers need to commit to reduce risks associated with innovation, while manufacturers must alleviate cost burdens [3] Group 2: Specific Applications and Technologies - In the smartphone sector, Processing In Memory (PIM) is currently the most innovative solution, albeit limited in quantity and primarily supporting Neural Processing Units (NPU) [3] - High Bandwidth Memory (HBM) is expected to enhance performance, although application scenarios remain unclear [3] - By 2026, Apple is anticipated to shift from Package on Package (PoP) to discrete packaging to improve bandwidth in iPhone Pro Max and foldable phones [3] Group 3: Automotive and Other Sectors - In the automotive sector, particularly with Tesla, the use of Application Processors (AP) and Low Power Double Data Rate (LPDDR) memory is increasing alongside the expansion of autonomous driving [4] - HBM4 is expected to be applied in autonomous vehicles post-2027, while Wide I/O usage is projected to expand in augmented reality (XR), drones, and gaming to improve latency performance [4] - NVIDIA's DIGITS technology aims to combine GPUs with HBM to expand memory bandwidth and enhance CPU bandwidth through SOCAMM, expected to be realized by mid-2025 [4]