TY - GEN
T1 - Characterization and Mitigation of Relaxation Effects on Multi-level RRAM based In-Memory Computing
AU - He, Wangxin
AU - Shim, Wonbo
AU - Yin, Shihui
AU - Sun, Xiaoyu
AU - Fan, Deliang
AU - Yu, Shimeng
AU - Seo, Jae Sun
N1 - Funding Information:
ACKNOWLEDGMENT The authors thank Winbond Electronics for RRAM chip fabrication support. This work is partially supported by NSF grants 1652866/1715443/1740225, JUMP CBRIC/ASCENT (SRC program sponsored by DARPA), NSF/SRC E2CDA, and SRC AIHW program.
Publisher Copyright:
© 2021 IEEE.
PY - 2021/3
Y1 - 2021/3
N2 - In this paper, we investigate the relaxation effects on multi-level resistive random access memory (RRAM) based in-memory computing (IMC) for deep neural network (DNN) inference. We characterized 2-bit-per-cell RRAM IMC prototypes and measured the relaxation effects over 100 hours on multiple 8 kb test chips, where the relaxation is found to be most severe in the two intermediate states. We incorporated the experimental data into SPICE simulation and software DNN inference, showing DNN accuracy for CIFAR-10 dataset could degrade from 87.35% to 11.58% after 144 hours. To recover the largely degraded accuracy, mitigation schemes are proposed: 1) at the circuit level, the reference voltage for RRAM IMC could be calibrated after 80 hours when the relaxation is saturated. 2) At the algorithm level, the weights are trained with lower percentages to be quantized to the two intermediate states. With both schemes applied, the accuracy could be recovered to 87.32 % for long-term stability.
AB - In this paper, we investigate the relaxation effects on multi-level resistive random access memory (RRAM) based in-memory computing (IMC) for deep neural network (DNN) inference. We characterized 2-bit-per-cell RRAM IMC prototypes and measured the relaxation effects over 100 hours on multiple 8 kb test chips, where the relaxation is found to be most severe in the two intermediate states. We incorporated the experimental data into SPICE simulation and software DNN inference, showing DNN accuracy for CIFAR-10 dataset could degrade from 87.35% to 11.58% after 144 hours. To recover the largely degraded accuracy, mitigation schemes are proposed: 1) at the circuit level, the reference voltage for RRAM IMC could be calibrated after 80 hours when the relaxation is saturated. 2) At the algorithm level, the weights are trained with lower percentages to be quantized to the two intermediate states. With both schemes applied, the accuracy could be recovered to 87.32 % for long-term stability.
KW - RRAM
KW - deep neural network
KW - in-memory computing
KW - multi-level cell
KW - relaxation effect
UR - http://www.scopus.com/inward/record.url?scp=85105577765&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85105577765&partnerID=8YFLogxK
U2 - 10.1109/IRPS46558.2021.9405228
DO - 10.1109/IRPS46558.2021.9405228
M3 - Conference contribution
AN - SCOPUS:85105577765
T3 - IEEE International Reliability Physics Symposium Proceedings
BT - 2021 IEEE International Reliability Physics Symposium, IRPS 2021 - Proceedings
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 2021 IEEE International Reliability Physics Symposium, IRPS 2021
Y2 - 21 March 2021 through 24 March 2021
ER -