The way machine learning is implemented in hardware (and digital described in an HDL) is evolving. New architectures and new ways of processing the data is inevitable.. Machine learning will require better and custom hardware. At some point, machine learning would be so fast and be able to handle complex tasks. So, in the end, machines will design the machines themselves. We are now in the slow part of an exponential curve. Nobody can accurately predict when this happens. It is unavoidable that machines become smart enough to develop machines themselves, which would be the fast exponential part of the curve.
My Quora space : HW accelerators eating AI