Analytical Method Estimates LLM Inference Energy on Modern GPUs

This research introduces a novel analytical framework for estimating the energy consumption of large language model (LLM) inference, specifically targeting NVIDIA H100-class GPUs. Unlike traditional methods that rely on direct hardware measurements, this approach provides a structured, empirically calibrated methodology. It integrates parameter-scaled FLOP accounting, memory traffic factors, and hardware-specific energy coefficients for various computational operations. The methodology offers a granular breakdown, distinguishing between prompt prefill and autoregressive decoding energy, and further dissects total energy into compute, parameter access, key-value cache write, and attention read components. This detailed analysis allows for understanding how energy scales with model size, context length, and generated token count. The goal is to provide transparent and reproducible approximations for comparative studies, green-coding initiatives, and early-stage design evaluations of LLM workloads, rather than replacing physical power measurements.