CUDA yes. I'm using an AMD Ryzen for data pre-processing and then an NVIDIA GPU for the actual AI training. The TensorFlow pre-processing pipeline is heavily optimized for parallelism, so it works perfectly on AMD CPUs with lots of cores.

As for MKL, yes it is impressive for some applications like real-time audio processing, but AFAIK with TensorFlow MKL is only used if you don't use a GPU, so pretty much never.