Abstract Modern-day deep neural community (DNN) devices are really configurable with big many choices that substantially have an impact on their non-functional conduct, as an example inference time and energy use. General performance products let to comprehend and predict the consequences of these configuration choices on procedure behavior, but ar