如果用一个词概括HBM这几年的进化，那就是：堆得越来越高。HBM本质上是一种"把 DRAM 垂直叠起来"的存储技术。层数越高，单颗 HBM 的容量越 大、带宽越高，对 AI GPU 来说就越香——因为 AI 真正稀缺的从来不是算力，而是喂数据的速度。 因此，HBM 的演进路线也非常清晰：从4层到8层、12层，再逼近16 层。8 层是 HBM 真正成熟、规模化出货的主力，它是过去一段时间 AI GPU 的"最常 见配置"，良率稳定、供应链也最成熟；12层则成为近两年的主力量产方向，在容量、性能与成本之间取得了更理想的平衡，也最适合大规模出货。而截 至目前，HBM 已经正式迈入16层堆叠的量产前夜：在刚刚结束的 CES 2026 上，SK 海力士已经展出了全球首款16层 HBM4 样品，单堆栈容量提升至 48GB。 但层数的提升，并不只是"多堆几层"这么简单。事实上，每增加 4 层，整个系统的制造难度都会显著上一个台阶：贴装精度、焊点间距、Z 方向高度控 制、翘曲、底填（MUF）可靠性……所有原本还能被工艺余量掩盖的问题，都会被 16 层这种高度放大到"生死线"级别。 面对困局，行业分化出了两种声音：一种是追求终 ...

Core Viewpoint - The evolution of HBM technology is characterized by increasing stack heights, enhancing capacity and bandwidth, which is crucial for AI GPUs due to their need for high data feeding speeds [1]. Group 1: HBM Technology Development - HBM has progressed from 4 layers to 8 layers, 12 layers, and is approaching 16 layers, with 8 layers being the most common configuration for AI GPUs [1]. - The introduction of 16-layer HBM4 has been showcased by SK Hynix, with a single stack capacity of 48GB [1]. - Increasing the number of layers significantly raises manufacturing challenges, including precision in mounting, solder joint spacing, and reliability issues [1]. Group 2: Hybrid Bonding and Fluxless Technology - Hybrid bonding is a cutting-edge interconnection technology that eliminates solder and flux, aiming for direct connections with higher I/O density [4]. - The recent JEDEC revision allows for a height increase in HBM modules, providing more space for traditional micro-bump technology [6]. - Fluxless technology is emerging as a transitional solution to address the limitations of traditional interconnection methods, particularly in high-density applications [8][12]. Group 3: TCB and Its Variants - Thermal Compression Bonding (TCB) is a key method for HBM stacking, allowing for higher interconnect density and precision [9][10]. - TCB has various types, including TC-CUF, TC-MUF, TC-NCP, and TC-NCF, each addressing specific challenges in high-density applications [12]. - The industry is moving towards Fluxless TCB to mitigate issues related to solder residues and improve yield and reliability [12][13]. Group 4: Industry Perspectives and Equipment Suppliers - SK Hynix remains cautious about adopting Fluxless technology for HBM4, preferring to continue with its Advanced MR-MUF process [19][21]. - BESI is seen as a proponent of hybrid bonding, focusing on preparing for future demands while facing short-term challenges due to slower-than-expected adoption rates [24]. - ASMPT emphasizes TCB as the core platform for HBM stacking, particularly during the transition from 12 to 16 layers, while also pushing for Fluxless advancements [25][26]. Group 5: Competitive Landscape - Hanmi Semiconductor is positioned as a key player in the "improvement route," optimizing TCB equipment for SK Hynix's processes [27]. - Hanwha Precision Machinery is emerging as a competitor, developing TCB equipment and exploring Fluxless technology to disrupt the existing supply chain [28]. - Kulicke & Soffa (K&S) is recognized for its stability and large-scale manufacturing experience, serving as a foundational player in the industry [29]. Conclusion - The delay in Fluxless technology adoption highlights the complexities of advanced packaging, emphasizing the need for a balance between innovation and production stability [31].

Core Viewpoint - The article discusses the challenges faced by Rapidus in achieving its ambitious goals in the semiconductor industry, particularly in the context of AI chip production and the transition to 3D IC technology. Group 1: Rapidus and AI Chip Production - Rapidus is focusing on advanced packaging technologies to secure orders from major clients like GAFAM in the growing AI market [4][8] - The company aims to mass-produce 2nm chips by 2027, but there are doubts about its capability to achieve this in the front-end process [7][8] - The article argues that Rapidus's goal of ultra-short turnaround time (TAT) for AI chip packaging is unrealistic due to various technological and supply chain challenges [71] Group 2: Transition to 3D IC Technology - The semiconductor industry is experiencing a paradigm shift from front-end processing to back-end 3D IC technology, which integrates multiple chips into a single package [29][31] - This shift is driven by the limitations of traditional scaling methods and the need for higher performance in AI applications [26][29] - Rapidus's entry into the 3D IC field aligns with industry trends, but achieving its goals will require overcoming significant hurdles [31][71] Group 3: Challenges in HBM Production - The production of High Bandwidth Memory (HBM) is a bottleneck for AI chip manufacturing, with a lead time of approximately six months [67] - HBM production is complex and costly, with a significantly lower yield compared to standard DRAM, making it a critical factor for companies like Rapidus [66][67] - The current market for advanced HBM is dominated by suppliers like SK Hynix, which has sold out its 2025 production capacity, further complicating Rapidus's plans [68][71]