Recent advances in artificial intelligence and deep learning are leading to an unforeseen demand for flexible yet efficient hardware architectures. Quantization of Neural Networks (NNs) seeks for an integer-only arithmetic, even towards extremely low precision such as 2-bits. Reduced integer precision can fully take advantage of Precision Scalable (PS) multipliers, that can adapt the datapath to increase efficiently the number of operations performed in parallel when operating at a low number of bits. Combining reduced arithmetic precision with PS multipliers leads to an improvement of the performance per energy ratio. However, reducing precision comes at the cost of reduced task accuracy, which might be unacceptable depending on the scenarios.