For machine learning Tensor Cores in the upcoming Volta GPUs seem much better idea, delivering 120 TFLOPS. Although the price would be a multiple of 400 USD, for sure.
> For machine learning Tensor Cores in the upcoming Volta GPUs seem much better idea, delivering 120 TFLOPS.
Yes, but the real-world benchmark numbers nvidia published show a much smaller speedup than one would expect if one looked only at the 120TFLOPS number.
This is because (a) "tensor core" is just a fancy name for "fast 4x4 matmul instruction" -- it's not applicable to everything you might do on the GPU, and (b) memory bandwidth did not increase commensurate with compute speed.
AFAIK AMD is busy pushing out support for MIOpen into Tensorflow etc.