I present an abstraction of the Hopfield-model for neural networks which is suitable for physical chip design using commerically available two-dimensional gate arrays. It can be shown that ±1-bonds combined with a dilution of about 80–90% of the original Hopfield-connections still lead to a comparable performance of the network. Furthermore the learning capability of the chips is discussed. Future extensions concerning programmable designs are outlined. The impact on aspects of brain research is discussed.