This paper investigates the accuracy of convolutional neural networks (CNNs) under constrained time cost, focusing on the trade-offs among factors such as depth, width, filter sizes, and strides. The authors propose an architecture that achieves 11.8% top-5 error on the ImageNet dataset while being 20% faster than AlexNet. The model is designed by progressively modifying a baseline model while preserving its time complexity. Through controlled experiments, the authors find that increasing depth significantly improves accuracy, but overly increasing depth can lead to accuracy degradation. They also explore trade-offs between depth and width, width and filter sizes, and the impact of adding pooling layers. The model J' achieves a top-1 error of 31.8% and top-5 error of 11.8%, with a training time of 3-4 days on a single GPU. The model is more accurate and faster than several competitive models. The study highlights the importance of understanding the factors that influence CNN accuracy under time constraints, which is crucial for practical applications in industrial and commercial scenarios. The results show that while increasing depth improves accuracy, it is not always beneficial, and careful trade-offs are necessary to achieve optimal performance. The model also demonstrates that using smaller filters can maintain accuracy while reducing complexity. The study provides insights into the design of efficient CNNs for real-world applications with limited time budgets.This paper investigates the accuracy of convolutional neural networks (CNNs) under constrained time cost, focusing on the trade-offs among factors such as depth, width, filter sizes, and strides. The authors propose an architecture that achieves 11.8% top-5 error on the ImageNet dataset while being 20% faster than AlexNet. The model is designed by progressively modifying a baseline model while preserving its time complexity. Through controlled experiments, the authors find that increasing depth significantly improves accuracy, but overly increasing depth can lead to accuracy degradation. They also explore trade-offs between depth and width, width and filter sizes, and the impact of adding pooling layers. The model J' achieves a top-1 error of 31.8% and top-5 error of 11.8%, with a training time of 3-4 days on a single GPU. The model is more accurate and faster than several competitive models. The study highlights the importance of understanding the factors that influence CNN accuracy under time constraints, which is crucial for practical applications in industrial and commercial scenarios. The results show that while increasing depth improves accuracy, it is not always beneficial, and careful trade-offs are necessary to achieve optimal performance. The model also demonstrates that using smaller filters can maintain accuracy while reducing complexity. The study provides insights into the design of efficient CNNs for real-world applications with limited time budgets.