EnergyLens: Solving LLM Inference Energy Optimization Challenges with an Interpretable Closed-Form Model

EnergyLens uses symbolic regression to derive a closed-form energy consumption model with only 12 parameters from a small number of samples. It achieves an 88.2% accuracy in configuration selection, far exceeding the traditional method's 60.9%, providing a physically interpretable and practical solution for energy optimization in LLM inference. This study addresses the limitations of existing energy optimization methods and represents a significant advancement in the field of energy optimization for large model deployment.

With the diversification of large language model (LLM) architectures (dense models, MoE models, state space models) and their deployment on heterogeneous accelerators to handle multimodal workloads, inference energy optimization is as important as latency and throughput optimization. Existing methods have limitations: either they use latency as a proxy for energy consumption (in over 20% of configurations, the latency-optimal and energy-optimal configurations do not overlap), or they rely on data-hungry black-box models (requiring hundreds of samples to generalize across models and hardware).

The core innovation of EnergyLens is using symbolic regression to derive a 12-parameter closed-form model from a small amount of profiling data, expressed entirely using system attributes (parallelism, batch size, sequence length, etc.), achieving three decouplings: separation of contributions from tensor parallelism and pipeline parallelism, separation of energy consumption between prefill and decoding stages, and cross-hardware portability. In terms of technical details, the 12 parameters cover energy consumption of compute-intensive operations, memory access overhead, parallel communication energy consumption, changes in batch processing efficiency, the impact of sequence length on bandwidth, etc. The structure is automatically discovered via symbolic regression without manual specification.

The research team fitted the EnergyLens model using only 50 performance profiling measurements. The Top-1 configuration selection accuracy reached 88.2%, far exceeding the previous analytical baseline of 60.9%, and the prediction accuracy is comparable to ensemble machine learning methods that require 10 times more samples. This reduces performance profiling overhead by an order of magnitude, and the closed-form nature makes the optimization results physically interpretable.

The practical value of EnergyLens includes: reducing data center operating costs (minimizing energy consumption while meeting latency SLAs), supporting green AI initiatives (reducing carbon footprint), accelerating new hardware adaptation (no need to re-collect large amounts of profiling data), and optimizing resource allocation in multi-tenant scenarios (energy-aware scheduling decisions).

Limitations and future directions of EnergyLens: 1. Dynamic workload adaptability (currently for static configurations; needs to be extended to scenarios with drastic changes in request patterns); 2. Complexity of multimodal workloads (energy consumption characteristics of video, audio, etc., differ significantly from pure text); 3. Interaction with compiler optimizations (coordinating model predictions with compiler decisions like XLA and TVM).

EnergyLens demonstrates that through symbolic regression and physically interpretable modeling, high-precision energy consumption prediction can be achieved with very few samples, providing a practical tool for the actual deployment of LLMs and new ideas for the sustainable development of AI systems and green computing. As the scale of LLM deployment expands, such energy optimization technologies will become an indispensable part of the infrastructure.