Scholarlink

<p>Modern computing powers applications from data analysis to artificial intelligence but now faces limitations. The slowdown of device scaling and the bottleneck between memory and processors motivate architectures that unify computation and data storage. Convolution is a core operation in learning, vision, and signal processing, yet its conventional implementation incurs high energy, high latency, and limited scalability. Magnetic systems that host spin textures, such as domain walls, offer dynamic behaviors that enable computation beyond traditional logic. Here we introduce a compute-in-memory platform that performs convolution by sequentially shifting magnetic domains and sensing the resulting signals. Information is written directly into domain patterns, processed through controlled motion, and read electrically, forming a nonvolatile structure suited for convolution tasks. This approach supports applications including Fourier analysis, neural networks, and image processing, achieving 10<sup>3</sup> to 10<sup>5</sup> improvements in area, energy, and throughput over existing technologies, marking a concrete advance in spintronic computing.</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

Domain wall motion-driven magnetic convolutional accelerator

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月14日

查看全文
  • Bingqian Dai,
  • Tianyi Wang,
  • Albert Lee,
  • Shijie Xu,
  • Chin-Chung Chen,
  • Kin Wong,
  • Dingyi Li,
  • Malcolm Jackson,
  • Yang Cheng,
  • Puyang Huang,
  • Yaochen Li,
  • Chao Yun,
  • Qingyuan Shu,
  • Haoran He,
  • Lixuan Tai,
  • Hanshen Huang,
  • Tien-Kan Chung,
  • Yanglong Hou,
  • Kang L. Wang

摘要

Modern computing powers applications from data analysis to artificial intelligence but now faces limitations. The slowdown of device scaling and the bottleneck between memory and processors motivate architectures that unify computation and data storage. Convolution is a core operation in learning, vision, and signal processing, yet its conventional implementation incurs high energy, high latency, and limited scalability. Magnetic systems that host spin textures, such as domain walls, offer dynamic behaviors that enable computation beyond traditional logic. Here we introduce a compute-in-memory platform that performs convolution by sequentially shifting magnetic domains and sensing the resulting signals. Information is written directly into domain patterns, processed through controlled motion, and read electrically, forming a nonvolatile structure suited for convolution tasks. This approach supports applications including Fourier analysis, neural networks, and image processing, achieving 103 to 105 improvements in area, energy, and throughput over existing technologies, marking a concrete advance in spintronic computing.