Arithmetic-unit and processor design for neural networks

The last decade saw a proliferation of research into the design of neurocomputers, many of which did not get beyond the prototype stage. We argue that, on the whole, neurocomputers are no longer viable; like, say, database computers before them, their time has passed before they became a common reality. We consider the implementation of hardware neural networks, from the level of arithmetic to complete individual processors and parallel processors and show that currents trends in computer architecture are not supportive of a case for custom neurocomputers. We argue that in the future, neural network processing ought to be mostly restricted to general-purpose processors or to processors that have been designed for other applications. There are just one or two two exceptions to this.