Authors: Qi Luo and Kunlin Li (Hong Kong University of Science and Technology, Guangzhou); Changxi Liu (National University of Singapore); Dongsheng Wang (Tsinghua University); and Yun Chen (The Hong Kong University of Science and Technology)
Abstract: GPUs enable large-scale LLM training, but high energy consumption and sharp power transients hurt efficiency and reliability. Existing GPU power-management methods are largely reactive and do not exploit the phase structure of training iterations. We present PHANTOM, a predictive, phase-aware dynamic voltage and frequency scaling (DVFS) framework for sustainable and reliable multi-GPU LLM training. We show that iteration-level power traces exhibit repeatable, fine-grained phase patterns. PHANTOM uses a lightweight temporal convolutional network (TCN) to predict short-horizon GPU power from synchronized hardware events and phase signals, then applies a Pareto-based multi-objective optimizer to choose frequencies ahead of phase transitions, balancing energy efficiency and power stability under real-time constraints. This design is lightweight enough for online deployment without intrusive changes to the training workflow. On multi-GPU LLM workloads, PHANTOM reduces energy by 15%–16%, normalized power-slew rate (NPSR) by 9%–15%, and energy–delay product (EDP) by 10%–11% compared with the default hardware DVFS policy.
Presentation: file