Scholarlink

<p>Compute-in-memory technology offers promising solutions for neural network acceleration but its potential is severely limited by inflexible and resource-intensive analog-to-digital converters. Here, we present a memristor-based analog-to-digital converter featuring adaptive quantization for diverse output distributions. Our design employs analog content-addressable memory cells with programmable overlapped boundaries to establish optimized quantization thresholds, demonstrating excellent integral and differential non-linearities. Extensive experiments validate the robustness of our approach by achieving 89.55% accuracy on CIFAR-10 (VGG8) at 5-bit adaptive quantized precision and maintaining competitive performance on ImageNet (ResNet18) through a proposed super-resolution strategy under experimental memristor variations. Compared to state-of-the-art designs, our converter achieves a 15.1× improvement in energy efficiency and a 12.9× reduction in area. Furthermore, integrating our converter into CIM systems reduces the energy and area overhead by up to 57.2% and 30.7%, respectively. This work establishes a paradigm for efficient and accurate signal quantization in practical compute-in-memory systems.</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

Memristor-based adaptive analog-to-digital conversion for efficient and accurate compute-in-memory

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月6日

查看全文
  • Haiqiao Hong,
  • Zhiyuan Du,
  • Mingrui Jiang,
  • Ruibin Mao,
  • Yuan Ren,
  • Fuyi Li,
  • Wei Mao,
  • Muyuan Peng,
  • Wei Zhang,
  • Zhengwu Liu,
  • Can Li,
  • Ngai Wong

摘要

Compute-in-memory technology offers promising solutions for neural network acceleration but its potential is severely limited by inflexible and resource-intensive analog-to-digital converters. Here, we present a memristor-based analog-to-digital converter featuring adaptive quantization for diverse output distributions. Our design employs analog content-addressable memory cells with programmable overlapped boundaries to establish optimized quantization thresholds, demonstrating excellent integral and differential non-linearities. Extensive experiments validate the robustness of our approach by achieving 89.55% accuracy on CIFAR-10 (VGG8) at 5-bit adaptive quantized precision and maintaining competitive performance on ImageNet (ResNet18) through a proposed super-resolution strategy under experimental memristor variations. Compared to state-of-the-art designs, our converter achieves a 15.1× improvement in energy efficiency and a 12.9× reduction in area. Furthermore, integrating our converter into CIM systems reduces the energy and area overhead by up to 57.2% and 30.7%, respectively. This work establishes a paradigm for efficient and accurate signal quantization in practical compute-in-memory systems.