E-MAC: Enhanced In-SRAM MAC Accuracy via Digital-to-Time Modulation

Abstract

In this article, we introduce a novel technique called E-multiplication and accumulation (MAC) (EMAC), aimed at enhancing energy efficiency, reducing latency, and improving the accuracy of analog-based in-static random access memory (SRAM) MAC accelerators. Our approach involves a digital-to-time word-line (WL) modulation technique that encodes the WL voltage while preserving the necessary linear voltage drop for precise computations. This eliminates the need for an additional digital-to-analog converter (DAC) in the design. Furthermore, the SRAM-based logical weight encoding scheme we present reduces the reliance on capacitance-based techniques, which typically introduce area overhead in the circuit. This approach ensures consistent voltage drops for all equivalent cases [i.e., $(a { \times} b) = (b \times a)$ ], addressing a persistent issue in existing state-of-the-art methods. Compared with state-of-the-art analog-based in-SRAM techniques, our E-MAC approach demonstrates significant energy savings ( $1.89\times $ ) and improved accuracy (73.25%) per MAC computation from a 1-V power supply, while achieving a $11.84\times $ energy efficiency improvement over baseline digital approaches. Our application analysis shows a marginal overall reduction in accuracy, i.e., a 0.1% and 0.17% reduction for LeNet5-based CNN and VGG16, respectively, when trained on the MNIST and ImageNet datasets.

Keywords

• 6T-static random access memory (SRAM)
• convolutional neural network (CNN)
• image classification
• processing in memory (PIM)