The problem of mapping neural nets onto massively parallel architectures is considered. The solution examined, based upon regular array structures, can support the mapping of any neural graph. In particular, the case of feed-forward multilayered nets is analyzed, and is proven that in this case the mapping suggested is easily implemented and optimizes a number of relevant figures of merit. The structure of nodes, I/O ports, and switches is taken into account with reference to the neural net case. It is seen that the claims of inherent fault tolerance for neural nets are not actually kept for all classes of faults of a digital implementation; moreover, it is considered that end-of-production defects require restructuring to grant nominal initial operation. An efficient and straightforward solution to the defect-tolerance problem is presented, allowing the most limited redundancy versus good harvesting characteristics.>