<p>Floating-gate memories (FGMs) show great promise for neuromorphic computing in efficient data-centric applications. However, their limited single-device state capacity remains insufficient for highly integrated precision computing. Here, we demonstrate 11-bit two-dimensional (2D) MoS<sub>2</sub> FGMs by contacting the 2D channels with bismuth electrodes, enabling 100 μA on-state current with 10<sup>8</sup> on/off ratio and reducing the current noise by 3 times (approaching the equipment limits) due to the Schottky barrier-free interfaces. Moreover, we employed a dual-pulse state editing scheme enhancing the stability of our FGMs. The devices show as high as 2,249 distinct conductance levels (&gt;11-bit) while maintaining 230 ns operation speed, &gt;10<sup>4 </sup>s retention, and &gt;10<sup>5</sup> cycle endurance. Furthermore, the gate-injection operation prevents the influence from generated defects during cycling, maintaining low noise even after 10<sup>5</sup> cycles and at 85 °C. Theoretical analysis reveals interfacial defects as the primary state-number limitation, suggesting 17-bit capacity is achievable through further trap density reduction. This work establishes 2D FGMs as promising candidates for high-bit-density, low-power neuromorphic hardware.</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

11-bit two-dimensional floating-gate memories

  • Yanrong Wang,
  • Yuchen Cai,
  • Feng Wang,
  • Tao Yan,
  • Shuhui Li,
  • Mingyang Cao,
  • Ruohao Hong,
  • Baoxing Zhai,
  • Kai Xu,
  • Xueying Zhan,
  • Jun He,
  • Zhenxing Wang

摘要

Floating-gate memories (FGMs) show great promise for neuromorphic computing in efficient data-centric applications. However, their limited single-device state capacity remains insufficient for highly integrated precision computing. Here, we demonstrate 11-bit two-dimensional (2D) MoS2 FGMs by contacting the 2D channels with bismuth electrodes, enabling 100 μA on-state current with 108 on/off ratio and reducing the current noise by 3 times (approaching the equipment limits) due to the Schottky barrier-free interfaces. Moreover, we employed a dual-pulse state editing scheme enhancing the stability of our FGMs. The devices show as high as 2,249 distinct conductance levels (>11-bit) while maintaining 230 ns operation speed, >104 s retention, and >105 cycle endurance. Furthermore, the gate-injection operation prevents the influence from generated defects during cycling, maintaining low noise even after 105 cycles and at 85 °C. Theoretical analysis reveals interfacial defects as the primary state-number limitation, suggesting 17-bit capacity is achievable through further trap density reduction. This work establishes 2D FGMs as promising candidates for high-bit-density, low-power neuromorphic hardware.