Scholarlink

<p>The rise of large-scale models has significantly increased the demand for high computational power and throughput in artificial intelligence (AI) chips, presenting challenges for existing architectures. Compute-in-memory (CIM) has emerged as a promising solution to address these bottlenecks. This study redefines the traditional computing architecture pyramid by introducing the CIM pyramid, structured around different storage media. CIM architectures are classified into two categories: compute-intensive CIM, which leverages static random-access memory (SRAM) and embedded dynamic random access memory (eDRAM), and memory-intensive CIM, which utilizes DRAM and non-volatile memory (NVM). The research reviews and analyzes recent advancements in both compute-intensive and memory-intensive CIM, highlighting their development trends and challenges. Furthermore, the study suggests a hybrid, heterogeneous architecture that integrates both CIM types with traditional computing systems, aiming to address the diverse computational needs of large models in the future.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Expansion of the memory pyramid in the era of large models: compute-intensive compute-in-memory and memory-intensive compute-in-memory

  • 期刊论文

  • 发表日期：

    2025年9月2日

查看全文
  • Zhichao Liu,
  • Yanqi Zhang,
  • Zhican Zhang,
  • Yi Yang,
  • Zhaoyang Zhang,
  • Xing Wang,
  • Jinwu Chen,
  • Xiaomin Li,
  • Xin Si,
  • Jun Yang

摘要

The rise of large-scale models has significantly increased the demand for high computational power and throughput in artificial intelligence (AI) chips, presenting challenges for existing architectures. Compute-in-memory (CIM) has emerged as a promising solution to address these bottlenecks. This study redefines the traditional computing architecture pyramid by introducing the CIM pyramid, structured around different storage media. CIM architectures are classified into two categories: compute-intensive CIM, which leverages static random-access memory (SRAM) and embedded dynamic random access memory (eDRAM), and memory-intensive CIM, which utilizes DRAM and non-volatile memory (NVM). The research reviews and analyzes recent advancements in both compute-intensive and memory-intensive CIM, highlighting their development trends and challenges. Furthermore, the study suggests a hybrid, heterogeneous architecture that integrates both CIM types with traditional computing systems, aiming to address the diverse computational needs of large models in the future.