ZCU104 DPU Inference Benchmark Dataset

Description: Raw sensor data collected at 50Hz from 76 sensors from 23 custom 1D CNN models on the ZCU104 MPSoC, including power rails, temperature sensors, CPU frequencies, and more. The collected data is available as Parquet. Hardware: Xilinx ZCU104 evaluation board, DPU core B4096 (DPUCZDX8G ISA 2), Vitis AI runtime 3.5.0 (pytorch-nndct 3.5.0+60df3f1+torch1.13.1), Ubuntu 22.04 (PYNQ 3.0.1). Models compiled from PyTorch 2.x training checkpoints using the Vitis AI Docker image (xilinx/vitis-ai-pytorch-cpu, Vitis AI 3.5.0, opset 17). Structure: Files are organized as model-{params}/rep{n}_*.parquet. File Contents rep{n}_batches.parquet Per-batch latency, energy (J), per-rail power (W), temperatures, IIO voltages, CPU frequencies rep{n}_hw_samples.parquet Raw HW telemetry time-series (50 Hz) per batch window rep{n}_predictions.parquet Per-sample ground-truth label, predicted label, and anomaly score rep{n}_summary.parquet Per-repetition aggregate metrics (accuracy, F1, latency percentiles, mean power, total energy)