A Task Allocation Method for Chiplet-Based CNN Accelerators

Chiplet-based Convolutional Neural Network (CNN) accelerators have emerged as a promising solution for efficient computational acceleration. However, communication between chiplets incurs higher costs compared to intra-chiplet communication. To address this challenge, this paper proposes a task allocation method designed to minimize inter-chiplet communication overhead. This method leverages the advantages of Time Division Multiplexing (TDM) and Space Division Multiplexing (SDM), consisting of two stages: task grouping and task partitioning. The CNN is divided into multiple task groups, which are allocated to different chiplets used SDM strategy. Subsequently, within each task group, the partitioning strategy of CNN layers is determined, and TDM strategy is used to allocate each CNN layer to processing units (PEs) within the chiplets. Experimental results demonstrate that the proposed method reduces inter-chiplet communication overhead by 59 % compared to uniform task allocation method.