Using NPUs to Increase Performance and Power Efficiency of ML Applications