The Quantitative Comparisons of Analog and Digital SRAM Compute-In-Memories for Deep Neural Network Applications

Abstract

Recently, static random-access memory (SRAM) compute-in-memory (CIM) research has been actively studied for energy efficient acceleration of deep neural network (DNN). The SRAM CIM research can be divided into analog CIM (ACIM) and digital CIM (DCIM) depending on the computing mechanism of MAC operation. Although both ACIM and DCIM are claimed energy efficient to accelerate DNN, detailed analysis and comparisons between two CIMs are hard to find. For the CIM designers who need to decide which type of CIMs can be selected for their DNN application, quantitative analysis in terms of energy and accuracy would be greatly helpful. In this paper, we compare the ACIM and DCIM in terms of the energy efficiency <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\text{TOPS}/\mathrm{W})$</tex> and accuracies. For ACIM design, BL chargesharing scheme and 3-bit flash ADC are selected for MAC operations, while adder-tree based MAC operations is used for DCIM design. Both approaches are designed with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$256\mathrm{x}64$</tex> macro using <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$28\text{nm}$</tex> CMOS process. The simulation results show that DCIM improves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\times 1.96$</tex> of energy efficiency <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\text{TOPS}/\mathrm{W})$</tex> and better CIFAR-1O accuracy compared to ACIM approach.