Revisiting convolutional design for efficient CNN architectures in edge-aware applications

While Vision Transformers (ViTs) have recently demonstrated impressive performance in computer vision tasks, their high computational demands and memory usage limit their applicability in real-time and edge AI scenarios. In contrast, Convolutional Neural Networks (CNNs) remain the preferred choice for such environments due to their lower latency, inductive bias, and efficiency. This study examines the impact of five widely used convolutional operations (standard 2D spatial, grouped, shuffle, depthwise separable and shift) when integrated into the ResNet-50 architecture. All model variants are trained on the Tiny-ImageNet and CIFAR-10/100 dataset under standardized GPU-based settings and evaluated across three edge AI platforms: Raspberry Pi 5, Coral Dev Board and Jetson Nano. The analysis includes parameter count, FLOPs, accuracy, inference time, power consumption and detailed runtime decomposition. Results show that while depthwise separable convolutions offer theoretical efficiency, they suffer from increased memory access on memory-bound platforms. In contrast, shuffle and shift convolutions yield better trade-offs between accuracy, computational load, and inference speed. These findings provide actionable insights for designing hardware-aware, deployment-optimized CNN architectures suitable for resource-constrained applications.