LQLipeng Qian

Lipeng Qian

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Lipeng Qian has a total of 5 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Operator-Level Compute-to-Power Modeling framework for Fine-Grained Power Prediction in Computing-Power Systems

    by Lipeng Qian, Ronghua Zhang, Lin Zhang

    The rapid growth of large AI models highlights the need for compute–power collaboration in green computing. Accurate power prediction during AI task execution is critical for energy optimization and resource scheduling. Existing methods usually treat an entire AI task as a single black-box unit. They map task-level features directly to power consumption. This approach ignores individual operator behaviors and their power interactions. Consequently, they fail to capture the relationships among operator topology, temporal dependencies, and power evolution, leading to limited accuracy and poor interpretability.To address these limitations, we propose an Operator-Level Compute–Power Modeling framework (OLCPM,). First, we reconstruct each AI task into an Operator-Power Graph (OPG). The OPG models both execution topology and latent power correlations, linking computation directly to power responses. Second, a Lightweight Dual-Layer Network (LDLN) captures intra-task operator dependencies and inter-task temporal dynamics. For intra-task modeling, a power-coupled graph attention mechanism learns topological and power interactions. For inter-task modeling, a shared-state block Mamba tracks long-range power dynamics using a shared state space and block compression. Finally, a multilayer perceptron outputs continuous power waveforms.We evaluate OLCPM across 11 representative AI workloads, including CNNs, Transformers, ViTs, and state-space models. Experiments show that OLCPM consistently outperforms state-of-the-art methods in prediction accuracy while remaining highly efficient. Furthermore, operator contribution analysis and attention visualization prove that the framework accurately quantifies operator impacts on power variations.