Fine grain power optimization in MPSoCs architectures is now available. It is possible to independently adjust the local frequency/voltage of each processor. The objective of this work was to investigate a new system-level approach to reduce the MPSoC global power consumption at run-time. Our proposal aims to dynamically adjust the local frequency/voltage settings of each processor to save energy while maintaining real-time deadline guarantees. The developed algorithm is fully decentralized and requires only local information from nearest nodes. This algorithm is based on a combination of subgradient methods with “consensus” concepts as a mean to accelerate convergence. A telecom test-case was used to illustrate our approach. Simulation results showed that an optimization as close as 4% compared to pareto solution can be achieved. Depending on applied constraints, energy may reach 45% regarding a worst-case static configuration. Moreover, when dealing with different application standards, our optimization process offers gains of up to 80%.