ADC-Free Current-Domain Compute-In-Memory Processors for Intelligent Image Sensing

Abstract

Current-domain compute-in-memory (CIM) architectures offer a promising pathway to reduce data movement and eliminate costly data conversion overheads in edge AI systems. However, conventional resistive crossbar implementations rely heavily on peripheral circuits such as analog-to-digital converters (ADCs), which dominate system energy and latency. This work presents a fully current-domain CIM architecture that performs matrix-vector multiplication (MVM) through direct summation of weighted input currents using current mirrors, eliminating the need for intermediate voltage domain conversion and ADCbased readout. Synaptic weights are implemented using a two-ReRAM voltage-divider structure that improves robustness to device variability by ensuring rail-to-rail switching behavior. An integrated current-domain non-linear activation stage suppresses leakage and background light currents. Circuit-level simulations demonstrate 9.6× improvement in energy efficiency and $8.3 \times$ reduction in latency compared to an ADC-based baseline. The proposed architecture enables scalable and energy-efficient current-domain in-memory computing for edge AI applications.