Dynamic Voltage and Frequency Scaling (DVFS) allows GPUs to adjust their power usage dynamically based on workload intensity, reducing energy consumption during low-demand periods while maintaining performance when needed. NVIDIA GPUs, such as those in DGX systems, support DVFS through tools like NVIDIA Management Library (NVML) and nvidia-smi, enabling fine-tuned power management. This approach balances efficiency and performance, critical for diverse AI workloads like training (high compute) and inference (variable demand), aligning with NVIDIA's energy-efficient computing initiatives.
Consolidating workloads onto a single GPU (Option A) risks overloading it, degrading performance and negating energy savings due to inefficiency. Scheduling workloads at night (Option C) addresses cost but not total consumption or sustainability, and it may delay time-sensitive tasks. Reducing clock speed universally (Option D) lowers power use but sacrifices performance across all workloads, which is impractical for an AI data center. DVFS is the most effective NVIDIA-supported strategy here.