SweetSpot Dataset: Energy and Performance Traces of LLM Inference Across Input and Output Lengths

A comprehensive benchmark dataset capturing power consumption, performance metrics, and timing traces for Large Language Models (LLMs) running on TensorRT-LLM inference engine using NVIDIA H100 GPUs. The dataset was produces for the paper SweetSpot: An Analytical Model for Predicting Energy Efficiency of LLM Inference (ICPE 2026), consult it for any additional information. The dataset contains detailed measurements of 13 different LLM models across various workload configurations, providing insights into token generation performance, energy consumption, and throughput characteristics. Each model was benchmarked under multiple input/output token lengths combinations to enable thorough analysis of inference behavior. The dataset covers 13 open-source LLM models ranging from 1.3B to 9B parameters from several families (LLaMa, OPT, Qwen, Granit, Gemma, Falcon). For each LLM, it contains traces collected with different workloads (Input/Output Lengths, Batch Size, Number of requests) and reports metrics on throughtput , latency and consumption (Total Energy (KJ), Energy Per Token (J), Efficiency (Tokens/J)).