This paper proposes a Floating-Point (FP) Digital Compute-in-Memory (CIM) architecture with a simplified FP alignment unit designed for Multiply-And-Accumulate (MAC) operations in machine learning (ML) models. By leveraging the distribution characteri...
This paper proposes a Floating-Point (FP) Digital Compute-in-Memory (CIM) architecture with a simplified FP alignment unit designed for Multiply-And-Accumulate (MAC) operations in machine learning (ML) models. By leveraging the distribution characteristics of batch normalized data, where FP data is typically bounded within a narrow range, the proposed design eliminates the need for maximum exponent searching to align FP data during MAC operation. This simplification significantly reduces the complexity of the hardware, minimizes area overhead, and improves efficiency. The proposed FP CIM architecture is evaluated using the ResNet18 model with both the CIFAR-10 and CIFAR-100 datasets. For CIFAR-10, the inference accuracy showed only a 0.39% drop compared to the baseline. When evaluated with the more complex CIFAR-100 dataset, the proposed architecture maintained stable inference accuracy with an appropriate maximum exponent, under dataset-dependent exponent constraints. These results demonstrate that the proposed simplified FP alignment unit can effectively preserve inference performance while achieving substantial improvements in configurability, area efficiency, and energy consumption, achieving up to 25.33% energy improvement.