Dot Product Engine Using Gated Schottky Diode with Quantized Weight

Hardware-based neural networks are expected to be a new computing breakthrough beyond conventional von Neumann architecture because of their low power operations. In this work, we investigate effect of quantized weight level on inference accuracy. Inference accuracy degrades when the number of conductance level decreases from 64 to 2. However, inference engine can be demonstrated easily as the number of quantized level decreases. Furthermore, in ternary weight, neural network becomes resilient to device variation with tuned weight threshold.