The technology described herein is directed towards optimizing power consumption of devices, e.g., in a datacenter. A modified (two-tier) genetic algorithm performs a carbon footprint-based optimization in a first tier to determine a candidate range of coefficients for each device type, e.g., servers, switches and storage devices/systems that likely reduce carbon footprint of each device type. In a second tier of the genetic algorithm, those ranges of coefficients are used in conjunction with actual power usage-based carbon footprint scores of individual devices to find respective sets of coefficients that minimize respective objective functions for the servers, the switches and the storage devices. The sets of coefficients can be used for power capping the devices. Device performance constraint-based intelligent selection can be used in one or both tiers to speed up convergence.