The real‑world performance of edge NPUs depends on hardware, model transformation, and the whole deployment pipeline. To isolate these factors we propose a benchmarking methodology that defines three configurations: Baseline, Vendor‑Guided NPU‑Only (VG‑NPU) and Vendor‑Guided End‑to‑End (VG‑E2E). VG‑NPU keeps the model I/O and host‑side processing unchanged while using only the vendor’s NPU compiler; VG‑E2E follows the vendor’s native deployment flow, including model partitioning, quantization and scheduling.
We evaluated four mainstream edge NPUs on three vision workloads. Across the 12 NPU‑workload pairs, VG‑NPU improved accuracy and runtime call latency in 10 pairs, whereas VG‑E2E reduced system‑level latency in all pairs but improved accuracy in only four. Notably, rankings based on NPU‑level performance do not always match system‑level rankings, and VG‑E2E achieved an additional 10.4%–70.7% reduction in system latency over VG‑NPU.
These results highlight that when comparing heterogeneous commercial edge NPU platforms, one must assess vendor‑guided optimizations both at the controlled model‑transformation stage and at the full deployment stage to draw reliable performance conclusions. Review