Workflow
集成80个HBM 4,台积电封装:疯狂炫技
TSMCTSMC(US:TSM) 半导体行业观察·2025-06-13 00:46

Core Viewpoint - TSMC's advanced packaging technology, SoW-X, aims to meet the demands of the next-generation AI semiconductor market by integrating high-performance computing chips and HBM4 modules, significantly enhancing performance and efficiency [1][3][4]. Group 1: SoW-X Technology Overview - SoW-X is a next-generation packaging technology set to be mass-produced by TSMC in 2027, designed for high-performance systems semiconductors like GPUs and CPUs, as well as AI semiconductors [3]. - The technology allows for direct connections between memory and system semiconductors without the need for traditional substrates, utilizing a fine copper redistribution layer (RDL) for inter-chip connections [3][4]. - SoW-X can integrate up to 16 high-performance computing chips and 80 HBM4 modules, achieving a total memory capacity of 3.75 TB and a bandwidth of 160 TB/s [3]. Group 2: Performance and Efficiency - Compared to existing AI semiconductor clusters with the same number of computing chips, SoW-X reduces power consumption by 17% and improves performance by 46% [3]. - The overall performance per watt of SoW-X is approximately 1.7 times higher than that of current AI semiconductor clusters, enhancing power efficiency and eliminating challenges associated with existing substrate connections [4]. Group 3: Market Implications and Challenges - TSMC positions SoW-X as an innovative technology platform aimed at the next generation of high-performance computing and AI industries, although short-term impacts on the AI memory market may be limited due to current low demand for ultra-large AI semiconductors [6]. - The predecessor of SoW-X, launched in 2020, has seen limited adoption among customers, including Tesla and Cerebras, indicating the niche market nature of this technology [6]. - The complexity and high technical difficulty of integrating numerous chips make it challenging for SoW-X to immediately replace popular AI semiconductor packaging technologies [6].