If AMD can write an abstraction layer similar to CuDNN, and write the initial integration to TensorFlow and Caffee I think they'd gain enough community support to finish out the rest.

80% of deep learning users use the TensorFlow, Caffe, or Torch frameworks.