With the widespread adoption of generative AI, semiconductors are required to provide high computational performance to handle massive computational models like Transformer, as well as sufficient data transfer bandwidth. Current AI chips face two primary bottlenecks: computational processing capability and data transfer speeds between processing units and memory, as well as between GPUs.
The core of semiconductor evolution accelerated by Generative AI: An introductory book explaining everything from basics to next-generation structures through diagrams
This article is a translation. Read the Japanese original
To resolve these issues, technologies such as HBM (High Bandwidth Memory), which stacks DRAM vertically to prevent data congestion; advanced packaging technologies that place chips and memory in close proximity; and ultra-high-speed interconnect technologies (such as NVIDIA's NVLink) that allow multiple GPUs to operate virtually as a single massive GPU have become essential.
Furthermore, structural evolution is advancing to address the limits of miniaturization. Moving from traditional planar structures to FinFET structures, where channels are shaped like fins and controlled from three sides, the industry is currently in a transition period toward GAA (Gate-All-Around) structures. In GAA, the gate surrounds the channel on all sides, dramatically increasing control capability. Samsung Electronics began mass production of its 3nm generation using GAA in 2022, TSMC plans to introduce it in its 2nm generation starting from the end of 2025, and Intel aims to transition to mass production of its GAA structure, referred to as "RibbonFET," starting in 2026.
"Semiconductors Explained Through Diagrams: Basics, Mechanisms, and Industry Map" (written by Zubu), published by SB Creative, provides a systematic explanation using diagrams, covering everything from semiconductor fundamentals to manufacturing processes and the complex industry structure, including fabless and foundry models.
Sources
- AIがもたらした「半導体」の進化 光電融合、HBM、新トランジスタ――果てなき性能向上、どうかなえる？ （ITmedia AI+、2026-09-30）
- 図解で深掘る きちんとわかる 半導体