Dopant network processing units: towards efficient neural network emulators with high-capacity nanoelectronic nodes

Abstract The rapidly growing computational demands of deep neural networks require novel hardware designs. Recently, tuneable nanoelectronic devices were developed based on hopping electrons through a network dopant atoms in silicon. These ‘dopant processing units’ (DNPUs) are highly energy-efficient and have potentially very high throughput. By adapting the control voltages applied to its electrodes, single DNPU can solve variety linearly non-separable classification problems. However, using device has limitations due implicit single-node architecture. This paper presents promising approach information by introducing DNPUs as high-capacity neurons moving from multi-neuron framework. implementing testing small multi-DNPU classifier hardware, we show that feed-forward improve performance 77% 94% test accuracy binary task with concentric classes plane. Furthermore, motivated integration memristor crossbar arrays, study potential combination linear layers. We simulation an MNIST only 10 nodes achieves over 96% accuracy. Our results pave road towards emulators offer atomic-scale low latency energy consumption.