Abstract: The Memristive Crossbar Array (MCA) serves as the fundamental hardware component of the Computing-in-Memory (CIM) architecture, enabling matrix operations to be performed with O(1) time complexity. However, due to the limited bit-width of device, existing methods often require configuring a large number of memory cells to represent numerical values, leading to increased hardware resource consumption and making it difficult to achieve both high precision and high energy efficiency. To address this issue, paper proposes a mixed-precision quantization method based on crossbar-aware. This method first employs K-means clustering to optimize output channel rearrangement, enhancing weight distribution consistency within sublayers to reduce quantization error and improve post-quantization model accuracy. Building upon this, sublayers are partitioned according to the physical constraints of the MCA, ensuring the output channel count aligns with parallel processing capacity of the MCA. This reduces the number of dequantization operations and lowers computational complexity. Simultaneously, an array-aware regularization term is introduced, combining the number of MCA required per sublayer with group Lasso regularization. This dynamically induces bit-level sparsity in weights, reducing hardware resource overhead while compressing bit width. Experiments show that the method is able to quantize the network model to an average of 1.3-bit with no more than 0.2% loss in accuracy and a reduction in hardware area overhead of about 74% compared to traditional quantization methods on different neural networks (ResNet/VGG). Compared with existing quantization schemes, the method proposed in this paper achieves a synergistic optimization of accuracy and hardware resources at very low bit-width.

摘要： 忆阻交叉阵列作为存内计算架构的核心硬件载体，可在O(1)时间复杂度内实现矩阵运算。然而，受器件有限位宽的限制，现有方法往往需要配置大量存储单元来表示数值，导致硬件资源消耗增加，高精度与高能效难以兼得。针对这一关键问题，提出一种基于阵列感知的混合精度量化方法。该方法首先结合K-means聚类对输出通道进行重排优化，以提升子层内权重分布的一致性从而降低量化误差，提高量化后模型精度；在此基础上，依据忆阻阵列的物理约束划分子层，使子层的输出通道数与阵列并行处理能力相匹配，减少反量化操作数，降低计算复杂度。同时，引入阵列感知正则化项，将子层所需阵列数量与组Lasso正则化相结合，动态诱导权重的位级稀疏性，在压缩位宽的同时降低硬件资源开销。在不同网络(ResNet/VGG)上的实验结果表明，该方法将网络模型量化至1.3位时精度损失控制在0.2%的同时，降低约74%的硬件面积开销。与现有量化方案相比，所提出的方法在极低位宽下实现了精度与硬件资源的协同优化。