电子行业：AI应用侧深度渗透，驱动国产先进封装技术寻求突破

Investment Rating - The report rates the electronic industry as "Overweight" (first-time rating) [1] Core Insights - DeepSeek has achieved three major breakthroughs in algorithm efficiency, significantly reducing memory usage and inference latency while maintaining model performance [1][2] - The optimization of algorithms does not diminish the value of the computing power industry but rather restructures demand, leading to a multi-tier growth engine in the trillion-level computing power market [1][29] - The demand for high-performance computing chips remains strong due to the continuous expansion of model parameters and training data [1][35] Summary by Sections Breakthroughs in DeepSeek Architecture - DeepSeek optimizes from input processing to computation and output stages, enhancing computing efficiency while reducing redundant calculations [2] - The introduction of a KV cache mechanism reduces inference complexity significantly, leading to lower resource consumption [2][5] - The Multi-Head Latent Attention (MLA) technology reduces KV cache space by over 90% while maintaining performance [5] Efficiency Improvements - The DeepSeek-V3 model activates only about 5.5% of parameters per token, significantly lowering computational load [8][10] - The GRPO algorithm allows for high-level reasoning capabilities without the need for supervised fine-tuning, achieving better performance with reduced training costs [21][25] Market Demand and Growth - The report predicts that the global advanced packaging revenue will grow from approximately $37.8 billion in 2023 to $69.5 billion by 2029, with a compound annual growth rate of 11% [1] - The demand for advanced packaging technologies, such as 2.5D and 3D packaging, is expected to rise as traditional methods face limitations [1][51] Investment Recommendations - The report suggests focusing on key players in the 2.5D/3D packaging technology, including equipment manufacturers and OSAT companies [1]