TY - JOUR
T1 - Advanced 2T0C DRAM Technologies for Processing-in-Memory - Part II
T2 - Adaptive Layer-Wise Refresh Technique
AU - Yook, Chan Gi
AU - Shim, Wonbo
N1 - Publisher Copyright:
© 1963-2012 IEEE.
PY - 2024
Y1 - 2024
N2 - Two-transistor-zero-capacitor (2T0C) DRAM cell has been proposed and extensively investigated as a memory device for processing-in-memory (PIM) applications. In this two-part article, we propose a novel vertical-transistor on the gate (VTG) 2T0C DRAM cell structure and the refresh technique for PIM applications and demonstrate their effectiveness. We described the improved retention characteristics of VTG DRAM in Part I. In Part II, we introduce the adaptive layer-wise refresh technique to minimize refresh energy consumption while maintaining the inference accuracy. Additionally, we developed a customized simulation framework to evaluate the inference accuracy and hardware performance of the 2T0C DRAM-based PIM macro. Through the simulations reflecting the device characteristics extracted in Part I, the layer-wise refresh technique can achieve the same inference accuracy of 92% and 91%, with refresh energy consumption reduced by 22.9% and 16% respectively, compared to the conventional refresh method.
AB - Two-transistor-zero-capacitor (2T0C) DRAM cell has been proposed and extensively investigated as a memory device for processing-in-memory (PIM) applications. In this two-part article, we propose a novel vertical-transistor on the gate (VTG) 2T0C DRAM cell structure and the refresh technique for PIM applications and demonstrate their effectiveness. We described the improved retention characteristics of VTG DRAM in Part I. In Part II, we introduce the adaptive layer-wise refresh technique to minimize refresh energy consumption while maintaining the inference accuracy. Additionally, we developed a customized simulation framework to evaluate the inference accuracy and hardware performance of the 2T0C DRAM-based PIM macro. Through the simulations reflecting the device characteristics extracted in Part I, the layer-wise refresh technique can achieve the same inference accuracy of 92% and 91%, with refresh energy consumption reduced by 22.9% and 16% respectively, compared to the conventional refresh method.
KW - Deep neural network (DNN)
KW - processing-in-memory (PIM)
KW - refresh technique
KW - two-transistor-zero-capacitor (2T0C) DRAM
UR - https://www.scopus.com/pages/publications/85207420108
U2 - 10.1109/TED.2024.3469183
DO - 10.1109/TED.2024.3469183
M3 - Article
AN - SCOPUS:85207420108
SN - 0018-9383
VL - 71
SP - 6639
EP - 6646
JO - IEEE Transactions on Electron Devices
JF - IEEE Transactions on Electron Devices
IS - 11
ER -