Benchmark Dataset for Prompt-Shape- and Prefix-Cache-Aware CPU/NPU Routing on Edge SoCs

Résumé fourni par la source

This dataset contains the benchmark measurements and routing evaluation data supporting the study “Prompt-Shape- and Prefix-Cache-Aware CPU/NPU Routing for Large Language Model Inference on Edge SoCs”. The experiments were conducted on an RK3588 edge SoC using CPU and NPU inference backends. The deposited data include the calibration and held-out workloads used to evaluate Prompt-Shape-Aware Routing (PSR), as well as the development and final held-out workloads used to calibrate and evaluate Prefix-Cache-Aware Routing (PCAR). The dataset contains request-level CPU/NPU latency measurements, prompt-length and generation-budget configurations, prefix-cache reuse information, and repeated measurements used to calculate cumulative request latency, backend-selection accuracy, and relative oracle loss. Scripts for routing evaluation and reproduction of the principal results are also included where applicable. This repository is intended to support reproducibility of the experimental results reported in the associated manuscript.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.