CSUM: A 28-nm Vertical-Feature-Shift and Horizontal-Weight-Shift-Based Separate-WL 6T-SRAM Computing-in-Memory Unit Macro for Edge Depthwise Neural-Networks

The pursuit of lightweight design is a critical research focus and imperative for advancing edge-AI networks. Support for depthwise (DW) convolutions is essential for many lightweight convolutional neural network (CNN) models. This article introduces a 6T-SRAM-based computing-in-memory (SRAM-CIM) unit macro for edge DW neural networks, capable of: 1) implementing a current-domain CIM macro with low-area overhead for standard and DW convolution; 2) adopting a computing and shift unit (CSU) based on a separate-WL 6T structure to achieve vertical-feature-shift (VFS) and horizontal-weight-shift, thereby enabling shifting-in-memory (SIM) and optimizing memory data access for DW convolution; and 3) employing a computation-current-tracked current-to-digital conversion scheme to free the signal voltage margin limitation and reduce area overhead. A fabricated 28-nm 9-kb SRAM-CIM unit-macro achieves an energy efficiency of 14.24 TOPS/W and area efficiency of 500 GOPS/mm2for 8-bit inputs, 8-bit weights, and full-precision DW convolution with nine-channel accumulation outputs. Compared to recent SRAM-CIM macros, this work achieves the highest energy and area efficiency for DW convolution. When considering both standard and DW convolution, it also improves comprehensive energy and area efficiency by1.34×and1.47×, respectively.