T-2T SRAM: A CMOS 8-bit T-2T SRAM Ladder-Based Analog Current-Domain Compute-in-Memory (CIM) Architecture

  • Kim, Honggu; 
  • Kim, Jaeyoun; 
  • Mun, Jeongmin; 
  • An, Yerim; 
  • Shim, Yong
Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

This paper presents a CMOS 8-bit T-2T SRAM ladder-based analog current-domain Compute-In-Memory (CIM) architecture for multi-bit neural network applications. A 1-to-2 ratio CMOS 8-bit T-2T SRAM ladder structure was implemented across the entire Multiply-Accumulate (MAC) macro, where the area scales linearly (similar to N) with N-weight, while the power consumption scales sub-linearly (similar to 2(1 -1 / 2 (N-1)), enabling areaand power-efficient multi-bit CIM operation. The proposed T-2T SRAM ladder-based Successive Approximation Register Analogto-Digital Converter (SAR ADC) effectively reuses the memory read and write peripheral circuits for asynchronous SAR operation, removing the need for dedicated SAR logic and register, thereby improving silicon area efficiency. Pre-distortion technique was implemented within input Digital-to-Analog Converter (DAC) stage, to compensate for the non-linearities within the Synaptic T-2T bitcells. The proposed T-2T SRAM architecture exhibits PVT tolerance through an inherent differential tracking and compensation mechanism between the main and reference computing paths, effectively suppressing PVT-induced variations and preserving MAC linearity. Fabricated using TSMC 65nm GP CMOS process, the proposed T-2T SRAM chip achieved a peak energy efficiency of 4.99 TOPS/W, a peak throughput of 1.5 GOPS/Kb, and a peak compute density of 0.114 TOPS/mm2 . Using the hardware-aware neural network analysis framework, we achieved 92.11% classification accuracy on the CIFAR-10 dataset with a VGG-9 model, and 67.43%/90.28% Top-1/Top-5 accuracy on the CIFAR-100 dataset with a ResNet-20 model.

키워드

In-memory computing; SRAM chips; Common Information Model (computing); Common Information Model (electricity); Computer integrated manufacturing; Arrays; Voltage; Architecture; Computer architecture; Media Access Control; Compute-in-memory; CIM; CMOS; multiply-accumulate; MAC; current-domain computing; static random access memory; SRAM; neural network hardware; NN hardware; MACRO; ACCELERATOR
제목
T-2T SRAM: A CMOS 8-bit T-2T SRAM Ladder-Based Analog Current-Domain Compute-in-Memory (CIM) Architecture
저자
Kim, Honggu; Kim, Jaeyoun; Mun, Jeongmin; An, Yerim; Shim, Yong
DOI
10.1109/TCSI.2026.3714492
발행일
2026
유형
Article; Early Access
저널명
IEEE Transactions on Circuits and Systems I: Regular Papers