TY - JOUR
T1 - A Charge-Domain 4T2C eDRAM Compute-in-Memory Macro With Enhanced Variation Tolerance and Low-Overhead Data Conversion Schemes
AU - Jung, In Jun
AU - Kim, Do Han
AU - Jo, Minyoung
AU - Ko, Dong Han
AU - Lee, Youngkyu
AU - Jung, Seong Ook
N1 - Publisher Copyright:
© 2023 IEEE.
PY - 2024/4
Y1 - 2024/4
N2 - In this brief, a novel charge-domain 4T2C eDRAM-CIM macro is proposed that has three key features: 1) a novel 4T2C eDRAM cell with an enhanced PVT variation tolerance that resolves the limitations of previous eDRAM-CIM macros such as current domain operation, large bit-cell size, cell leakage, and low energy- and area-efficiency, resulting in high linearity (R2 = 0.9998) and 76.8% reduction in 3σ PVT variation, 2) a quarter-ADC-reduction scheme with an offset-calibration comparator that reduces the number of ADC by 73% while improving the accuracy drop by 7.82%, and 3) an array-embedded DAC that reduces the area overhead by 64.2% compared to current-based DAC. The proposed 4T2C eDRAM-CIM macro is fabricated in 65nm LP technology and achieves 43.02- to 49.20-TOPS/W and 2.4-TOPS/mm2 when 4b×4b MAC operation is performed with 250MHz. In addition, an 90.03% accuracy at the CIFAR-10 dataset with the ResNet-20 network is achieved.
AB - In this brief, a novel charge-domain 4T2C eDRAM-CIM macro is proposed that has three key features: 1) a novel 4T2C eDRAM cell with an enhanced PVT variation tolerance that resolves the limitations of previous eDRAM-CIM macros such as current domain operation, large bit-cell size, cell leakage, and low energy- and area-efficiency, resulting in high linearity (R2 = 0.9998) and 76.8% reduction in 3σ PVT variation, 2) a quarter-ADC-reduction scheme with an offset-calibration comparator that reduces the number of ADC by 73% while improving the accuracy drop by 7.82%, and 3) an array-embedded DAC that reduces the area overhead by 64.2% compared to current-based DAC. The proposed 4T2C eDRAM-CIM macro is fabricated in 65nm LP technology and achieves 43.02- to 49.20-TOPS/W and 2.4-TOPS/mm2 when 4b×4b MAC operation is performed with 250MHz. In addition, an 90.03% accuracy at the CIFAR-10 dataset with the ResNet-20 network is achieved.
KW - Compute-in-memory (CIM)
KW - charge-domain
KW - embedded DRAM (eDRAM)
KW - low-overhead data conversion
KW - retention time
KW - variation tolerance
UR - http://www.scopus.com/inward/record.url?scp=85177033306&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85177033306&partnerID=8YFLogxK
U2 - 10.1109/TCSII.2023.3332749
DO - 10.1109/TCSII.2023.3332749
M3 - Article
AN - SCOPUS:85177033306
SN - 1549-7747
VL - 71
SP - 1824
EP - 1828
JO - IEEE Transactions on Circuits and Systems II: Express Briefs
JF - IEEE Transactions on Circuits and Systems II: Express Briefs
IS - 4
ER -