2.86% accuracy drop on MNIST is massive. This accuracy could be achieved with dramatically simpler algorithms. We use neural networks because they are more accurate, so if you take that away there is no point.
Also, they didn't actually build any hardware, they only ran some simulations.
I think this type of solutions will be more and more relevant as we progress toward more on-chip processing as system on chip (SoC) become more and more ubiquitous.