Optimizing Internal Communication of Compute-in-Memory-Based AI Accelerator

In CIM (compute-in-memory) based system, the weights of neural network models need to be mapped to the memory array, and the mapping policy has a huge impact on the performance of the system. In this paper, existing weight mapping policies are analysed and categorized into two types: positionwise and channel-wise mapping. Channel-wise policy causes non-continuous memory address at its output, while positionwise policy suffers from its large amount of input data and the problem of data concatenation. After that, a novel weight mapping policy, named as mixed dimension mapping, is proposed to overcome the limitations of the existing policies. Experimental results show that it reduces the communication load of the system by 14%-32% and avoids data concatenation completely.