https://www.quora.com/What-is-the-underlying-reason-for-AMD-GPUs-being-so-bad-at-deep-learning
Lots of work need to be done to reach Nvidia’s performance. Cudnn is a huge step for performance that AMD does not have yet.
There was OpenCL support for torch, but pragmatically everyone just turned to CUDA devices. If someone is willing to make a cudnn equivalent maybe it would change.
It’s also not only a software problem. But apparently AMD is planning on producing DL oriented hardware
http://www.amd.com/en-us/press-releases/Pages/radeon-instinct-2016dec12.aspx