OPTIMAS Experiment Dataset: Prompts, LLM Responses, and Runtime Measurements for Diagnostic-Informed GPU Code Optimization

This dataset accompanies the KDD 2026 paper "OPTIMAS: An Intelligent Analytics-Informed Generative AI Framework for Performance Optimization." It contains the prompts, LLM-generated optimized CUDA code responses, and runtime measurements from 3,410 experiments across 9 GPU benchmarks (Accuracy, Adjacent, BabelStream, Aop, Extrema, Shmembench, Sobol, SW4Lite, XSBench) evaluated against three LLMs: GPT-5, Gemini 2.5 Pro, and Llama 3.1 70B. Each (application, LLM) pair is stored in its own folder containing prompt and response .txt files per diagnostic configuration (PC sampling, Roofline, hardware counter interaction analysis, and combinations thereof). Runtime metadata and applied/ignored optimization rationales are provided in the included CSV files. All experiments were run on an NVIDIA H100 GPU using CUDA 12.3.