Title:Comparative Analysis of CNN Performances Using CIFAR-100 and MNIST Databases: GPU vs. CPU Efficiency
Volume: 18
Author(s): Rania Boukhenoun, Hakim Doghmane, Kamel Messaoudi*El-Bay Bourennane
Affiliation:
- Electrical Engineering and Renewable Energy Laboratory (LEER), Mohamed Cherif Messaadia University, Souk-
Ahras, Algeria
Keywords:
CNN architecture, deep learning, image classification, training time, gpu-nvidia, CPU, GPU.
Abstract:
Introduction: Convolutional Neural Networks (CNNs) and deep learning algorithms
have significantly advanced image processing and classification. This study compares three CNN
architectures VGG-16, ResNet-50, and ResNet-18, and evaluates their performance on the CIFAR-
100 and MNIST datasets. Training time is prioritized as a critical metric, along with test accuracy
and training loss, under varying hardware configurations (NVIDIA-GPU and Intel-CPU).
Methods: Experiments were conducted on an Ubuntu 22.04 system using the PyTorch framework.
The hardware configurations included an NVIDIA GeForce GTX 1660 Super GPU and an Intel
Core i5-10400 CPU. The CIFAR-100 dataset, containing 60,000 color images across 100 classes,
and the MNIST dataset, comprising 70,000 grayscale images, were used for benchmarking.
Results: The results highlight the superior efficiency of GPUs, with training times reduced by up
to 10x compared to CPUs. For CIFAR-100, VGG-16 required 13,000 seconds on the GPU versus
130,000 seconds on the CPU, while ResNet-18, the most time-efficient model, completed training
in 150 seconds on the GPU and 1,740 seconds on the CPU. ResNet-50 achieved the highest test
accuracy (~80%) on CIFAR-100. On MNIST, ResNet-18 was the most efficient, with training
times of 185 seconds on the GPU and 22,000 seconds on the CPU.
Discussion: This study highlights the clear advantage of GPUs over CPUs in reducing training
times, particularly for complex models such as VGG-16 and ResNet-50. ResNet-50 achieved the
highest accuracy, while ResNet-18 was the most time efficient. However, the use of simpler datasets
(MNIST and CIFAR-100) may not fully capture the complexity of the real world.
Conclusion: This study emphasizes the importance of hardware selection for deep learning workflows.
Using the CIFAR-100 and MNIST datasets, we demonstrated that GPUs significantly outperform
CPUs, achieving up to a 10x reduction in training time while maintaining competitive
accuracy. Among the architectures tested, ResNet-50 delivered the highest test accuracy (~80%)
on CIFAR-100, demonstrating superior feature extraction capabilities compared to VGG-16 and
ResNet-18. Meanwhile, ResNet-18 proved to be the most time-efficient architecture, completing
training in 150 seconds on the GPU, a significant improvement over VGG-16's 13,000 seconds.
These results highlight the advantage of residual connections in reducing training complexity and
achieving higher performance. The results underscore the critical role of both architecture selection
and hardware optimization in advancing deep learning workflows.
