New model reveals GPU energy efficiency sweet spots for A100, H100, H200

Energy consumption remains a critical bottleneck in high-performance computing, especially for heterogeneous CPU-GPU systems handling AI and scientific workloads. In a new preprint, researchers from the distributed computing community studied how hardware configuration affects energy-to-solution on modern NVIDIA GPUs (A40, A100, H100, H200) paired with an Intel Ice Lake CPU. Using realistic molecular dynamics benchmarks (GROMACS and AMBER) and a stress-test benchmark (FIRESTARTER), they varied frequency scaling and power caps to map out efficient and inefficient operating regimes. Their key discovery: GPU energy efficiency follows a workload- and architecture-dependent pattern governed by a nonlinear power-frequency relationship, with a distinct transition frequency beyond which efficiency sharply degrades.

The team developed an interpretable analytical model that decomposes GPU power into linear and nonlinear components. The model fits empirical data with low error and highlights three dominant parameters: baseline power, nonlinear power behavior, and the transition frequency. Importantly, the study finds that power capping is generally less effective for efficiency tuning than reducing frequency, especially for workloads operating far from thermal design power. This work provides model-driven guidance for selecting optimal operating points, offering a practical framework for HPC operators and AI engineers to cut energy costs without sacrificing throughput. The paper is available on arXiv under ID 2607.00819.