This is all more or less true with the current public code. We're building our own standalone stack that will interop with TF, Keras, PyTorch, etc. Basically we are fixing the problem for AMD and Intel on desktop but also embedded chip vendors as well. Currently our stack runs Google's Xception net in Keras slightly faster than TF 1.2 + cuDNN 5.1 on Tesla K80 so there's good evidence OpenCL itself is not that slow.