Evaluating and Comparing the Potentials in Primary Response for GPU and CPU Data Centers

The rapid growth of Large Language Models (LLMs) and Artificial Intelligence (AI) has transformed traditional CPU-centric Data Centers (DaCe) into more power-demanding GPU DaCes. Previous work has explored methods to reduce energy costs and carbon emissions in GPU DaCes. However, there remains a gap in understanding the potential of GPU DaCes for providing primary response, a crucial ancillary service for stabilizing the power system. Drawing on real-world job traces from a GPU-intensive DaCe operated by SenseTime and a CPU-intensive DaCe at Oak Ridge National Laboratory, we developed a mixed-integer linear programming model to assess the DaCe flexibility potentials considering individual jobs’ characteristics. We show that the GPU DaCe possesses a larger flexibility for delivering primary responses compared to the CPU DaCe. Furthermore, the GPU DaCe exhibits lower variability in flexibility across different times of the day and over a 7-month evaluation horizon, making them more dependable and stable sources for offering primary response.