记忆体内运算(PIM)技术
Search documents
三星与SK 海力士,首次超越台积电
半导体芯闻· 2025-12-24 10:19
Core Insights - The semiconductor industry's profit structure is changing due to the rapid increase in demand for artificial intelligence (AI) and a significant rise in memory prices [2] - Samsung Electronics and SK Hynix are expected to surpass TSMC in gross margin by Q4 2025, marking the first time since Q4 2018 that the memory industry outperforms foundries [2] - Micron, the third-largest memory manufacturer, anticipates its gross margin to reach 67% in Q2 of FY2026, indicating potential for profitability to exceed TSMC as well [2] Group 1: Memory Market Dynamics - The three major memory manufacturers have allocated approximately 18% to 28% of their DRAM capacity to high bandwidth memory (HBM), which has led to a quarterly price increase of over 30% for standard DRAM [2] - The shift in memory gross margins surpassing foundries is closely linked to changes in AI industry demand structures [2] Group 2: AI and Memory Technology - As AI applications transition from training to inference, the demand for high-speed data storage and instant access has significantly increased, relying heavily on HBM to supply data to GPUs for computation [3] - Despite general memory not matching HBM's performance, the demand for high-performance general memory is rapidly growing, with GDDR7 and LPDDR5X handling initial inference workloads [3] - Memory manufacturers are developing AI-oriented high-performance products, such as Processing-In-Memory (PIM) technology, which allows memory to perform some computations traditionally handled by GPUs [3]