Lightweight large language models (LLMs) are increasingly deployed on personal computers, particularly in resource-constrained edge and mobile environments. In these settings, energy consumption, execution time, and memory usage directly affect usability. However, existing evaluations of LLM efficiency rely heavily on proxy descriptors such as parameter count or FLOPs, often decoupled from task precision. This paper introduces a PTME-based experimental framework for precision-aware profiling of lightweight LLM inference, jointly measuring Precision, execution Time, peak Memory usage, and Energy consumption through direct hardware-level measurements.
The methodology is applied to a representative set of lightweight LLMs executed locally within edge-class resource envelopes on a controlled desktop platform, using benchmarks that span code generation, mathematical reasoning, and multi-task understanding. We find that static proxy descriptors approximate inference cost well but fail to predict precision. Tightening the resource envelope increases cost without affecting precision, amplifying execution time more strongly than energy and penalizing larger models the most. Moreover, no single model dominates across all PTME dimensions, and a Pareto analysis reveals non-dominated configurations that would be hidden by accuracy-only or efficiency-only assessments, providing practical guidance for selecting models under different resource envelopes. These results indicate that selecting lightweight LLMs based solely on size, FLOPs, latency, or accuracy may lead to the wrong deployment candidate; PTME profiling uncovers configurations that preserve useful accuracy at lower physical costs.
Blogger's Review: This paper offers a fresh perspective on profiling lightweight large language models, particularly in resource-constrained environments, emphasizing the importance of a comprehensive assessment of precision, time, and energy consumption, which is of significant practical and academic value.