TY - GEN
T1 - Refresh Methods and Accuracy Evaluation for 2T0C DRAM based Processing-in-memory
AU - Yook, Chan Gi
AU - Shim, Wonbo
N1 - Publisher Copyright:
© 2023 IEEE.
PY - 2023
Y1 - 2023
N2 - Processing-in-memory (PIM) technology has been expected to be an effective memory technology for artificial intelligence (AI) computation with high data throughput and energy efficiency. Among various PIM component devices, 2-transistor (2TOC) DRAM has attracted attention due to its superior sensing speed, on/off ratio, and chip density. However, its retention characteristic requires frequent refresh, which results in considerable power consumption. Therefore, finding the effective refresh method that minimizes energy consumption while maintaining accuracy is crucial. In this paper, we introduce burst refresh and periodic refresh methods and compare their performance. We extract the accuracy-time graph when the refresh techniques described above are applied through our developed 2TOC DRAM-based PIM simulation framework, which can evaluate inference accuracy. Under the conditions of this paper, when a 32 ms refresh is performed, burst refresh causes the accuracy drop to 83 % just before the refresh, while periodic refresh showed an improved performance to maintain accuracy at around 88-92 %.
AB - Processing-in-memory (PIM) technology has been expected to be an effective memory technology for artificial intelligence (AI) computation with high data throughput and energy efficiency. Among various PIM component devices, 2-transistor (2TOC) DRAM has attracted attention due to its superior sensing speed, on/off ratio, and chip density. However, its retention characteristic requires frequent refresh, which results in considerable power consumption. Therefore, finding the effective refresh method that minimizes energy consumption while maintaining accuracy is crucial. In this paper, we introduce burst refresh and periodic refresh methods and compare their performance. We extract the accuracy-time graph when the refresh techniques described above are applied through our developed 2TOC DRAM-based PIM simulation framework, which can evaluate inference accuracy. Under the conditions of this paper, when a 32 ms refresh is performed, burst refresh causes the accuracy drop to 83 % just before the refresh, while periodic refresh showed an improved performance to maintain accuracy at around 88-92 %.
KW - 2TOC DRAM
KW - benchmark simulator
KW - DRAM
KW - inference accuracy
KW - Processing-in-Memory (PIM)
KW - refresh
UR - http://www.scopus.com/inward/record.url?scp=85169836839&partnerID=8YFLogxK
U2 - 10.1109/ITC-CSCC58803.2023.10212837
DO - 10.1109/ITC-CSCC58803.2023.10212837
M3 - Conference contribution
AN - SCOPUS:85169836839
T3 - 2023 International Technical Conference on Circuits/Systems, Computers, and Communications, ITC-CSCC 2023
BT - 2023 International Technical Conference on Circuits/Systems, Computers, and Communications, ITC-CSCC 2023
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 2023 International Technical Conference on Circuits/Systems, Computers, and Communications, ITC-CSCC 2023
Y2 - 25 June 2023 through 28 June 2023
ER -
