Novel Parallelization Strategies For High-Performance Dnn Training On Hpc Systems