TY - GEN
T1 - Design Exploration of Dynamic Multi-Level Ternary Content-Addressable Memory Using Nanoelectromechanical Relays
AU - Li, Taixin
AU - Zhong, Hongtao
AU - George, Sumitha
AU - Narayanan, Vijaykrishnan
AU - Shi, Liang
AU - Yang, Huazhong
AU - Li, Xueqing
N1 - Publisher Copyright:
© 2023 IEEE.
PY - 2023
Y1 - 2023
N2 - Multi-Level Ternary Content Addressable Memories (ML-TCAMs) are a type of TCAM that can calculate the hamming distance between the stored data and the input vector, which can be used to accelerate several specific applications. There have been several existing current-domain and charge-domain ML-TCAMs based on SRAMs and nonvolatile memories (NVMs). However, they fail to meet a good balance between area and computational accuracy tradeoffs.In this paper, for the first time, we explore the design of dynamic ML-TCAMs that achieve both high cell density and high accuracy, and propose DyLAN, the current-domain dynamic ML-TCAM using the 4-terminal nanoelectromechanical (NEM) relays. Specifically, combined with the nearly zero OFF-state leakage and stable ON-state current of the 4-terminal NEM relays, this paper proposes DyLAN-W with ultra-long retention time and DyLAN-S with ultra-low single refresh overhead and high density, respectively. Results show that DyLAN achieves up to 2.7 x and 4.9x area reduction compared with the 16T SRAM ML-TCAM and the charge-domain ML-TCAMs, respectively, and increases the few-shot learning accuracy by 13.7% (from 79.9% to 93.6%) on average compared with the state-of-the-art nonvolatile ML-TCAM, i.e., the 2FeFET ML-TCAM.
AB - Multi-Level Ternary Content Addressable Memories (ML-TCAMs) are a type of TCAM that can calculate the hamming distance between the stored data and the input vector, which can be used to accelerate several specific applications. There have been several existing current-domain and charge-domain ML-TCAMs based on SRAMs and nonvolatile memories (NVMs). However, they fail to meet a good balance between area and computational accuracy tradeoffs.In this paper, for the first time, we explore the design of dynamic ML-TCAMs that achieve both high cell density and high accuracy, and propose DyLAN, the current-domain dynamic ML-TCAM using the 4-terminal nanoelectromechanical (NEM) relays. Specifically, combined with the nearly zero OFF-state leakage and stable ON-state current of the 4-terminal NEM relays, this paper proposes DyLAN-W with ultra-long retention time and DyLAN-S with ultra-low single refresh overhead and high density, respectively. Results show that DyLAN achieves up to 2.7 x and 4.9x area reduction compared with the 16T SRAM ML-TCAM and the charge-domain ML-TCAMs, respectively, and increases the few-shot learning accuracy by 13.7% (from 79.9% to 93.6%) on average compared with the state-of-the-art nonvolatile ML-TCAM, i.e., the 2FeFET ML-TCAM.
UR - http://www.scopus.com/inward/record.url?scp=85172127971&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85172127971&partnerID=8YFLogxK
U2 - 10.1109/ISVLSI59464.2023.10238633
DO - 10.1109/ISVLSI59464.2023.10238633
M3 - Conference contribution
AN - SCOPUS:85172127971
T3 - Proceedings of IEEE Computer Society Annual Symposium on VLSI, ISVLSI
BT - 2023 IEEE Computer Society Annual Symposium on VLSI, ISVLSI 2023 - Proceedings
A2 - Kastensmidt, Fernanda
A2 - Reis, Ricardo
A2 - Todri-Sanial, Aida
A2 - Li, Hai
A2 - Metzler, Carolina
PB - IEEE Computer Society
T2 - 26th IEEE Computer Society Annual Symposium on VLSI, ISVLSI 2023
Y2 - 20 June 2023 through 23 June 2023
ER -