[C59] ConvMapSim: Modeling and Simulating Convolutional Weight Mapping for PIM Arrays

Abstract

While the selection of convolutional weight mapping method significantly influences the latency and energy efficiency of PIM systems, existing PIM simulators lack the flexibility needed to assess and compare various mapping approaches. Additionally, a rigorous definition and formulation of the weight mapping problem remain outstanding challenges. To address this issue, this paper first establishes a formal definition of the convolutional weight mapping method accompanied by a rigorous mathematical model for evaluating its computing cycles. Then, we introduce ConvMapSim, a novel simulation platform designed to compute the energy and latency performance of state-of-the-art mapping methods. Furthermore, the platform offers visualizations of mapping methods, providing intuitive insights into their structures and behaviors.