An Improved Strategy for Data Layout in Convolution Operations on FPGA-Based Multi-Memory Accelerators

Convolutional Neural Networks (CNNs) are fundamental to modern AI applications but often suffer from significant memory bottlenecks due to non-contiguous access patterns during convolution operations. Although previous work has optimized data layouts at the software level, hardware-level solutions for multi-memory accelerators remain underexplored. In this paper, we propose a hardware-level approach to mitigate memory row conflicts in FPGA-based CNN accelerators. Specifically, we introduce a dynamic DDR controller generated using Vivado 2019.1, which optimizes feature map allocation across memory banks and operates in conjunction with a multi-memory architecture to enable parallel access. Our method reduces row conflicts by up to 21% and improves throughput by 17% on the KCU1500 FPGA, with validation across YOLOv2, VGG16, and AlexNet. The key innovation lies in the layer-specific address mapping strategy and hardware-software co-design, providing a scalable and efficient solution for CNN inference across both edge and cloud platforms.