interesting, FPGA was used to speed up bitcoin mining, outdating the use of GPU - now it is coming to neural networks.
I think main advantage here is not the speed but power consumption and ease of deployment (small form factor etc.). Also as mentioned in the article, they can do a lot of customization there. Such as custom design floating point units, lookup tables for sigmoid calculations etc.
Imagine 3-D printable hardware that is able to adapt to functional needs like a lisp machine.