The performance of computing systems has been increasingly choked by power consumption and memory access time within and between system components. Meanwhile, the explosion of artificial intelligence requires massive data-heavy computation. Therefore, it is crucial to develop energy efficient computing from devices to architectures. This work is developed along three streams: a steep device with low operation voltage, a novel device enabling complex logic operation, and an efficient modeling algorithm to quickly incorporate emerging devices into circuit designs. On the first front, tunnel field effect transistors (TFETs), which switch by modulating quantum tunneling, promise sub-60 mV/dec subthreshold swing and operate at low power consumption. Based on the unique properties of atomically thin 2D layered materials, two-dimensional heterojunction interlayer tunneling field effect transistor (Thin-TFET) was proposed as a ultra-scaled steep transistor. On the second front, we converted the “undesirable” ambipolar behavior in TFETs into XNOR logic operation, and proposed a one-transistor XNOR design: TransiXNOR. On the third front, we structured artificial neural networks with awareness of device physics, and developed an accurate, efficient, and generic device compact modeling algorithm: physics-inspired neural network (Pi-NN).