Dynamic bit-width reconfiguration for energy-efficient deep learning hardware