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

Abstract

This dataset supports the experiments reported in “Prompt-Shape- and Prefix-Cache-Aware CPU/NPU Routing for Large Language Model Inference on Edge SoCs.” It contains CPU/NPU latency measurements, benchmark prompts and prompt manifests, prompt-generation scripts, measurement runners, and analysis code for Prompt-Shape-Aware Routing (PSR) and Prefix-Cache-Aware Routing (PCAR) on the RK3588 platform. The main 1.5B experiments include 56 PSR calibration workloads, 49 independent PSR held-out workloads, 60 PCAR development workloads, and 45 independent cache-aware final workloads, with three repetitions on both CPU and NPU. A secondary Qwen2.5-0.5B sensitivity dataset is also included. The released datasets distinguish actual emitted output length (output_tokens) from backend-native generation counters (runtime_generate_counter). Detailed field definitions, token-accounting conventions, experimental configuration, and reproduction instructions are provided in DATA_DICTIONARY.md, config/experiment_config.md, and README.md.