Work

A 6T--11T In-Memory Computing Architecture for Near-Sensor Pipelined Binary Neural Network

Year: 2026
Type: article
Abstract: This paper presents the fabrication and experimental validation of an interlaced 6T-11T CMOS in-memory computing circuit designed for low-silicon-area pipelined binary neural networks. The circuit implements a weight-stationary architecture in which input activation values are sequentially stored in 6T SRAM columns, while 11T cells retain the convolution weights and perform binary XNOR-based multiply-and-accumulate operations, naturally aligning with the column-parallel readout topology of CMOS image sensors. The circuit has been fabricated in a 180 nm standard CMOS process and experimentally characterized. Measured results on a binary convolutional layer demonstrate operation at 1 MHz with a simulation-assisted energy efficiency of 18.53 TOPS/W. The interlaced organization embeds the line-buffer functionality required for streamed convolution within the IMC array itself, so that image rows are processed as they are read out from the sensor.
Language: English
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7204246013
FWCI:
Cites: 15
Cited by:
Related to: 10