Lookup Table-Based Computing-in-Memory Macro Approximating Dot Products Without Multiplications for Energy-Efficient CNN Inference

This paper presents a lookup table-based dot product approximation (LDA) for energy-efficient convolutional neural network (CNN) inference as another computing-in-memory (CIM) technique. Using the proposed LDA, computations of CNN can be performed without multiplications; however nearest neighbor search (NNS) is required, resulting in a large energy overhead. To overcome this issue, we propose a time-domain associative memory (TAM) that allows energy-efficient NNS. When the proposed LDA is applied to CNN, the inference accuracy decreases, because LDA involves an approximate computation. To resolve this accuracy degradation, training methods for the LDA-based CNN are proposed. The proposed LDA-based CIM macro using TAM is fabricated in a 65 nm CMOS process. Measurement results indicate that TAM consumes less energy than previously reported associative memory. Owing to the energy-efficient TAM, the LDA-based CIM macro can achieve an energy efficiency of 138 TOPS/W, which is 7.5 times higher than those of the conventional CIM macros fabricated in the same process node.