A Survey on Energy-Efficiency Mechanisms for Large Language Model Training: Measurement, Optimization Mechanisms, Evidence Boundaries, and Future Research Directions

The rapid growth of large language models (LLMs) has made training energy efficiency a major systems and sustainability challenge. Training requires substantial computation, memory, communication, and electrical energy, yet the literature often treats runtime, floating-point operations, memory, communication, cost, energy, and carbon as interchangeable indicators of efficiency. This survey critically reviews 54 research papers and one supporting software tool across measurement and carbon accounting, model and numerical efficiency, memory and communication, distributed planning, GPU power control, carbon-aware scheduling, lifecycle design, and reliability. We classify evidence as direct, modeled, simulated, reported-comparison, or enabling, and interpret results only within their stated workload, hardware, quality condition, and measurement boundary. The synthesis shows that individual mechanisms are mature, yet cross-paper comparison remains difficult because studies use inconsistent boundaries, quality targets, platforms, and accounting methods. We, therefore, propose a quality-aware reporting framework centered on energy to target quality, explicit boundaries, uncertainty, and reproducibility. The survey establishes an evidence-disciplined framework for evaluating energy-efficient LLM training and identifies integrated, quality-aware, whole-system energy optimization as the principal research direction.