Frankly, this is starting to become less of a hurdle in that pure CUDA code is now rarely written in the DL community (instead opting to call optimized libraries like CuDNN). The biggest issue is replicating all the good "core computing libraries" that NVIDIA offers: CUBLAS, NPP, CuDNN, Tensor-RT, etc.
Frankly, if AMD can start putting this stuff together over the next 2-3 years it's usually not too bad to have deep learning frameworks leverage this stuff.
Frankly, if AMD can start putting this stuff together over the next 2-3 years it's usually not too bad to have deep learning frameworks leverage this stuff.