CUDA is superior to OpenCL. Also NVIDIA provides CuDNN, a proprietary library with very efficient implementation of deep learning primitives. If you want to train models faster, you have to use them.