TSYS School Research Team of Yi Zhou, Prudhvi Raj Nelapatla and Elijah Starkey Critically Review Research on Large Language Models

A new study by TSYS School computer scientist Yi Zhou and TSYS School students Prudhvi Raj Nelapatla and Elijah Starkey asserts that the rapid growth of large language models has made training energy efficiency a major systems and sustainability challenge. As they explain, 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. To examine these issues, the trio of researchers critically reviewed 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. They classified evidence as direct, modeled, simulated, reported-comparison, or enabling, while they interpreted 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. Their study, which appears in the current issue of Electronics, therefore, proposes 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 large language models training and identifies integrated, quality-aware, whole-system energy optimization as the principal research direction.