Carbon-Aware Scheduling of Generative AI Workloads in Renewable-Powered Data Centers

The booming development of the Generative Artificial Intelligence (GenAI) workloads has markedly risen the rate of energy consumption and carbon footprint in data centers. Even though the facilities powered by renewable energy would provide a sustainable alternative to this issue, they become unstable with the situation of variable energy supply, which makes scheduling of GenAI training and inference processes with high compute intensity difficult. In this chapter, a carbon-conscious framework of scheduling is presented, which strives to match GenAI workloads to the real-time and predicted renewable energy supply and grid carbon intensity. The model unites the forecasting of renewable-oriented energy, the scheduling of activities through carbon sensibilities and the migration of cloud workloads among the distributed data facilities. To minimize emissions whilst preserving the quality of services, the system is based on temporal shifting, migration, and hybrid scheduling. With the real renewable-generation and carbonintensity traces in experiments it is demonstrated that approximately 28.5% reduction in carbon emissions is reached by matching the workloads to the periods of high renewable-energy availability. The cross-region placement also gives up to 40% additional emission reduction when the renewable supply is not the same in all areas. It is also found that the study focused major operational restrictions, such as training checkpoint overhead, inference latency sensitivity, and accelerator energy elasticity. Generally, the research proves that carbon-aware scheduling can significantly reduce the carbon footprint of GenAI workloads and that carbon-aware scheduling provides a promising direction to implementation of sustainable and active AI systems in data centers that utilize renewable energy.