CNN model acceleration and optimization: CPU and GPU performance analysis and evaluation for multiclass classification

The field of agriculture has undergone a revolutionary evolution towards "smart agriculture," integrating AI technology through deep learning techniques, especially CNN. This profoundly renews computer vision in various tasks, including image detection and classification. However, most existing work focuses primarily on improving accuracy indicators. The associated energy cost is often overlooked. This article presents an empirical performance evaluation that emphasizes not only model accuracy but also energy efficiency, incorporating statistical analysis. We focused on hardware acceleration with the NVIDIA GPU, with the help of the CUDA API, through libraries optimized for matrix computations, such as cuDNN. Based on the Fruits-360 dataset, we carried out an in-depth evaluation through a comparative analysis of CNN performance on the two hardware platforms (CPU vs. GPU). Our results show that the GPU delivered a significant speedup, being 4x faster than the CPU in terms of frames processed per second, with a latency reduction of 75%, which increases throughput compared with the CPU. We also confirmed that the GPU demonstrated greater energy efficiency, with an improvement factor of 2.78, underlining the interest in our approach for applications such as embedded vision, where hardware resources are limited. Although the model performed well on a single dataset, validation on other datasets would be necessary to confirm its robustness and improve its generalization. • Our approach offers an optimal, innovative, and rarely explored balance between high accuracy and hardware optimization in multi-class classification. • CPU–GPU benchmarking for smart agricultural sorting. • CUDA-cuDnn-accelerated CNN inference: 4× GPU speedup over CPU under identical workloads, 75% lower latency, and Higher Throughput. • Up to 3.5× higher energy efficiency on the GPU Platform and higher GFLOPS/J than CPU. • This optimal balance between accuracy, energy efficiency, and cost makes our approach particularly well-suited to embedded systems.