This research explores enhancing Deep Neural Network (DNN) training efficiency and robustness using model parallelism on multi-GPU platforms. The approach focuses on optimizing the distribution of model components across hardware to improve throughput and stability during the training process.