Benchmark datasets of GPU frequency switching and transition latencies across NVIDIA CUDA architectures

This repository contains the benchmark datasets and analysis outputs accompanying the paper “An In-Depth Study of GPU Frequency-Scaling Latency and Its Optimization on Modern Architectures” (Velicka et al., 2025). The datasets capture switching and transition latencies of GPU frequency scaling across three different NVIDIA CUDA architectures — RTX Quadro 6000 (Turing), A100 SXM-4 (Ampere), and GH200 (Grace-Hopper Superchip). The data were generated using the LATEST benchmarking tool, which implements a methodology for measuring both GPU switching latency (request + transition) and transition latency (execution time until a target frequency stabilizes). Measurements were repeated across a wide range of initial/target frequency pairs to ensure statistical robustness, while outliers were filtered using density-based clustering (DBSCAN). Contents: RTX Quadro 6000 – Baseline dataset with raw switching/transition measurements, scatter plots (with/without outliers), detailed zoom-ins, and latency heatmaps. A100 SXM-4 – Results from four individual devices, each stored separately with the same data structure as for the RTX Quadro 6000. GH200 (Grace-Hopper Superchip) – Includes standard direct transitions as well as experiments with intermediate-frequency switching strategies, with and without delays. The dataset provides both raw CSV data and processed visualizations (scatter plots, detailed plots, outlier-filtered plots, heatmaps). Related publication: Velicka, D., Vysocky, O., & Riha, L. (2025). Methodology for GPU Frequency Switching Latency Measurement. Proceedings of the 39th IEEE International Parallel and Distributed Processing Symposium (IPDPS), Heterogenity in Computing Workshop (HCW). 10.1109/IPDPSW66978.2025.00133 Velicka, D., Vysocky, O., Yasal, O., & Riha, L. (2026). An In-Depth Study of GPU Frequency-Scaling Latency and Its Optimization on Modern Architectures. Future Generation Computer Systems (FGCS), Volume 179. 10.1016/j.future.2025.10833