GPU Power Super-Resolution: Dataset and Experimental Results

Dataset and experimental results for the GPU power super-resolution thesis. A lightweight MLP is trained to reconstruct 20 Hz sub-second GPU power traces from 1 Hz NVML sensor readings, augmented with request-level metadata from the vLLM serving engine. Ground-truth power measurements are collected using PowerSensor3 hardware instrumentation. Raw data archive (gpu-power-superres-raw-v1.0.tar.gz, 1.8 GB): Two raw data collections containing high-frequency PowerSensor3 sensor dumps, NVML metrics, vLLM request logs, and phase markers. Processed data archive (gpu-power-superres-processed-v1.0.tar.gz, 259 MB): Train/validation/test splits in NumPy format, out-of-distribution evaluation sets (ShareGPT and mixed workloads), CUPTI profiling variants, trained model checkpoints at multiple sampling frequencies and random seeds, and all experimental results (evaluation metrics, ablation studies, DVFS power capping comparisons, per-request energy attribution, and latency benchmarks). Figures can be regenerated using the plotting scripts in the source code repository.