Novel Parallelization Strategies for High-Performance DNN Training on HPC Systems