Scholarlink

<p>Compute-in-memory (CIM) based on various devices such as static random access memory (SRAM) and resistive random access memory (RRAM) and other emerging devices such as indium-gallium-zinc-oxide (IGZO) transistor, magnetoresistive RAM (MRAM), and ferroelectric RAM (FeRAM) has been explored for better performance and energy efficiency on neural networks applications. However, CIM based on a single-type device suffers a variety of non-ideal metrics to reach the simultaneous optimal accuracy, density, and energy efficiency. This work presents equivalent-ideal CIM (Eq-CIM), a monolithic 3D (M3D) IGZO-RRAM-SRAM integrated architecture for robust and efficient compute-in-memory enabling equivalent-ideal device metrics. To overcome the non-ideality (variation, endurance, temperature, etc.) of the single-type devices, system-technology co-optimization (STCO) is performed. This work highlights the non-ideality-aware functionality breakdown for robust high accuracy and simultaneous high density/efficiency, by utilizing 2T0C IGZO for temporal activation storage, RRAM for high-density weight storage, and SRAM for accurate CIM. Device-to-algorithm variation transfer is applied to analyze the system-level accuracy. We benchmark Eq-CIM architecture on CIFAR-10/ImageNet, with 5.06× storage density and 5.05×/2.45× area/energy efficiency compared with single-type-device-based CIM, and high robustness (&lt;0.27% accuracy loss) from −40°C to 120°C.</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

A monolithic 3D IGZO-RRAM-SRAM-integrated architecture for robust and efficient compute-in-memory enabling equivalent-ideal device metrics

  • 期刊论文

  • 发表日期：

    2025年1月15日

查看全文
  • Shengzhe Yan,
  • Zhaori Cong,
  • Zi Wang,
  • Zhuoyu Dai,
  • Zeyu Guo,
  • Zhihang Qian,
  • Xufan Li,
  • Xu Zheng,
  • Chuanke Chen,
  • Nianduan Lu,
  • Chunmeng Dou,
  • Guanhua Yang,
  • Xiaoxin Xu,
  • Di Geng,
  • Jinshan Yue,
  • Lingfei Wang,
  • Ling Li,
  • Ming Liu

摘要

Compute-in-memory (CIM) based on various devices such as static random access memory (SRAM) and resistive random access memory (RRAM) and other emerging devices such as indium-gallium-zinc-oxide (IGZO) transistor, magnetoresistive RAM (MRAM), and ferroelectric RAM (FeRAM) has been explored for better performance and energy efficiency on neural networks applications. However, CIM based on a single-type device suffers a variety of non-ideal metrics to reach the simultaneous optimal accuracy, density, and energy efficiency. This work presents equivalent-ideal CIM (Eq-CIM), a monolithic 3D (M3D) IGZO-RRAM-SRAM integrated architecture for robust and efficient compute-in-memory enabling equivalent-ideal device metrics. To overcome the non-ideality (variation, endurance, temperature, etc.) of the single-type devices, system-technology co-optimization (STCO) is performed. This work highlights the non-ideality-aware functionality breakdown for robust high accuracy and simultaneous high density/efficiency, by utilizing 2T0C IGZO for temporal activation storage, RRAM for high-density weight storage, and SRAM for accurate CIM. Device-to-algorithm variation transfer is applied to analyze the system-level accuracy. We benchmark Eq-CIM architecture on CIFAR-10/ImageNet, with 5.06× storage density and 5.05×/2.45× area/energy efficiency compared with single-type-device-based CIM, and high robustness (<0.27% accuracy loss) from −40°C to 120°C.