For efficient HPC, graphics cards do one thing very well. Vector Math.
For large regression sets the better implemented the vector sets are and the better managed the memory is the better the result will be.
Cuda is an chip centric implementation for NVidias graphics cards that gets the most out of their cards. Cuda -> PTX -> unpublished ML -> Card
Cuda is a not terribly low level API for Graphics processing..
The tools that NVidia have invested in to sit on top of if FFT libraries, BLAS, dnn are just that tools.
OpenCL is an open standard that is designed by a committee and is not AMD's answer to Cuda. I wish OpenCL was ubiquitous but it isn't.
Mantle has the ability to organize around structures that are natural to the architecture of the AMD processor environment to make optimization easier for it's compilers.
AMD is a bit of a dark horse in the race... Intel plus Intel Mic // IBM Open plus Nvidia Tesla // AMD Opteron plus ?? //
The cards in AMD's suite all have display ports attached to them and the FirePros are clearly targeted at Rendering.
Some development cards with silicon that balanced fp32 and fp64 and a strongly supported API in combination with tool ecosystem investment could make AMD fun again.
The megawatts installations aren't going to jump on 500watt cards but in the workstation compute or small cluster arena it could get a toehold.