BARQ: Boundary-Aware Regularized Training for Accurate Inference on Computing-in-Memory Accelerators with Low-Precision A/D Conversion

Computing-in-Memory (CIM) with ReRAM crossbar arrays accelerates Deep Neural Network (DNN) computations by executing operations directly within memory, which reduces both data movement and energy consumption. Recent studies show that lowering the precision of Analog-to-Digital Converters (ADC) in crossbar peripheral circuits can significantly reduce system area and power overheads at the risk of increasing quantization-induced numerical errors that degrade inference accuracy. This paper proposes a Boundary-Aware Regularized Quantization (BARQ) technique to maintain accuracy in CIM accelerators with low-precision ADCs. BARQ introduces ADC bit-width boundary constraints to regulate model weights during quantization-aware training (QAT), mitigating overflow-induced clipping errors and enhancing unstructured weight sparsity. Additionally, we introduce a novel weight initialization strategy based on Euclidean projection, which minimizes initial quantization error and facilitates faster and more stable convergence. Experiments on the CIFAR-10, CIFAR-100, and ImageNet datasets show that BARQ achieves minimal accuracy loss of only 0.59% even with 3-bit ADCs. Furthermore, by enhancing weight sparsity, BARQ further reduces energy consumption by up to 75% and improves hardware efficiency by 3.79× compared to conventional quantization methods.