A Quantization Model Based on a Floating-point Computing-in-Memory Architecture

Abstract

Computing-in-memory (CIM) has been proved to perform high energy efficiency and significant acceleration effect for high computational parallelism neural networks. Floating-point numbers and floating-point CIMs (FP-CIM) are required to execute high performance training and high accuracy inference for neural networks. However, none of former works discuss the relationship between circuit design based on the FP-CIM architecture and neural networks. In this paper, we propose a quantization model based on a FP-CIM architecture to figure out this relationship in PYTORCH. According to experimental results we summarize some principles on FP-CIM macro design. Using our quantization model can reduce data storage overhead by more than 70.0%, and control floating-point networks inference accuracy loss within 0.5%, which is 1.7% better than integer networks.