With the widespread adoption of deep learning across industries and academia, the energy consumption of AI computing clusters has surged, creating significant economic and environmental challenges. Improving AI server energy efficiency has therefore become a critical research focus. Traditional dynamic voltage and frequency scaling (DVFS) research has primarily focused on regulating either CPU or GPU/NPU frequency in isolation, which cannot fully unlock the energy-saving potential of AI servers. Naively applying joint DVFS requires exhaustively evaluating all frequency combinations to find the most energy-efficient one, which becomes prohibitively expensive with many frequency states. To address this limitation, this study proposes CoDVFS, a Coordinated Dynamic Voltage and Frequency Scaling framework that employs Bayesian optimization to rapidly converge to optimal frequency configurations for both CPUs and GPU/NPUs. Experimental results on AI servers running HPL-MxP workloads demonstrate that CoDVFS enhances server energy efficiency by 17%, achieving significant energy savings. We also analyze the relationship between whole-server and accelerator-only energy efficiency, proposing an optimum alignment criterion. On our device, the optima of both metrics coincide. However, the criterion reveals that if server and accelerator power scale differently on other devices, optimizing accelerator efficiency alone may not reflect true server efficiency.