Supporting data for: Human activity recognition on a NearLink-capable WS63 RISC-V microcontroller: An empirical study of deployment resources

Supporting measurements and reproducibility materials for local human activity recognition inference on the HiHope HH-D02 development board with a WS63 RISC-V microcontroller. The deposit separates P0 subset logit validation, P1 full-test INT8-weight replay over 2,947 windows across 24 firmware chunks, and P2 same-board FP32/INT8-weight comparison over six separately built, flashed and started firmwares. P2 repeats the same 48 distinct windows three times per mode, yielding 144 calls per mode. Runtime clock state was not independently recorded, so timing differences are descriptive rather than calibrated causal speedups. Materials include archived UART records, P1 chunk-level classification and timing summaries, P2 per-call timings, host references, resource-accounting tables, a checkpoint, its parameter-matched 44,906 B MindIR export, recovered quantization settings, exported parameter headers, the attributed 48-window UCI HAR input subset, model-specific C inference sources and analysis/validation code. The firmware executes the C engine, not the MindIR file. Historical Lite exports are excluded and are not used in the final manuscript comparison. All board inputs are prerecorded UCI HAR windows. No live-IMU, wireless or energy measurements are supplied. Individual P1 call timings and full-test device logits are not available. SDK trees, SDK-dependent task/build components, complete linker maps, ELF binaries, firmware packages and build logs are not included. The public validator recomputes archived records and checks resource-CSV consistency; it does not independently recount withheld linker maps or recreate the board build. Licensing is file-specific: author-created data, documentation, configuration and model/parameter artifacts outside code/ are licensed under CC BY 4.0; author-created software in code/ is licensed under MIT. The adapted UCI HAR subset retains its upstream CC BY 4.0 terms, source citation and adaptation notice. See LICENSE, code/LICENSE and RIGHTS_AND_ATTRIBUTION.md. The upstream dataset is Human Activity Recognition Using Smartphones, UCI Machine Learning Repository, https://doi.org/10.24432/C54S4K. Dependencies and the SDK must be obtained from their respective providers under their own terms.