Scholarlink

<p>While AI-generated content (AIGC) strives to replicate human imagination, current models like score-based diffusion remain slow and energy-intensive. This inefficiency stems from conventional digital computers, where physically separated storage and processing units cause data-transfer bottlenecks, and discrete operations disrupt naturally continuous generation dynamics. Here we show a brain-inspired, analog in-memory computing system that overcomes these limitations. By employing resistive memory, our system integrates storage and computation to act as a time-continuous neural differential equation solver. We experimentally validate our solution with 180 nm resistive memory in-memory computing macros. While maintaining generative quality equivalent to the software baseline, our system accelerated both unconditional and conditional generation tasks, by factors of 69.0 and 116.5, respectively, compared to advanced digital hardware. Furthermore, it reduced energy consumption by 31.5% and 52.0%, respectively. Our approach expands the horizon for hardware solutions in edge computing for generative AI applications.</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

Resistive memory-based neural differential equation solver for score-based diffusion model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月12日

查看全文
  • Jichang Yang,
  • Hegan Chen,
  • Jia Chen,
  • Songqi Wang,
  • Qifan Zhu,
  • Xinyuan Zhang,
  • Yan Zeng,
  • Shaocong Wang,
  • Mingrui Yang,
  • Yifei Yu,
  • Xi Chen,
  • Bo Wang,
  • Binbin Cui,
  • Yi Li,
  • Ning Lin,
  • Meng Xu,
  • Yi Li,
  • Xiaoxin Xu,
  • Xiaojuan Qi,
  • Xumeng Zhang,
  • Dashan Shang,
  • Zhongrui Wang,
  • Han Wang,
  • Qi Liu,
  • Kwang-Ting Cheng,
  • Ming Liu

摘要

While AI-generated content (AIGC) strives to replicate human imagination, current models like score-based diffusion remain slow and energy-intensive. This inefficiency stems from conventional digital computers, where physically separated storage and processing units cause data-transfer bottlenecks, and discrete operations disrupt naturally continuous generation dynamics. Here we show a brain-inspired, analog in-memory computing system that overcomes these limitations. By employing resistive memory, our system integrates storage and computation to act as a time-continuous neural differential equation solver. We experimentally validate our solution with 180 nm resistive memory in-memory computing macros. While maintaining generative quality equivalent to the software baseline, our system accelerated both unconditional and conditional generation tasks, by factors of 69.0 and 116.5, respectively, compared to advanced digital hardware. Furthermore, it reduced energy consumption by 31.5% and 52.0%, respectively. Our approach expands the horizon for hardware solutions in edge computing for generative AI applications.