Automated Hardware Complexity Profiling for Edge-Deployed YOLO Models

Edge deployment of object detectors fails when designers reason from theoretical FLOPs alone. We present the ModelComplexityProfiler, a deterministic post-training state of a distributed MLOps pipeline that fuses theoretical and empirical metrics into a single hardware cost matrix. Using ptflops it records GFLOPs; using CUDA-event timing and pynvml it records wall-clock latency and peak VRAM. Across YOLOv8n/s/m at 320, 640, and 1280 pixels the profiler exposes three repeatable facts: GFLOPs scale quadratically with resolution (observed ratio 4.00 +- 0.01 per input doubling); peak VRAM is astonishingly linear in GFLOPs (VRAM = 17.65 x GFLOPs + 200.2 MB, R^2 > 0.999); and latency grows sub-quadratically, from 2.38x to 3.86x per 4x FLOP step, because fixed memory overheads amortize as workloads grow. A conservative guardrail (25% VRAM headroom plus a 15 FPS floor) under a 2 GB edge budget rejects YOLOv8s and YOLOv8m at 1280px, two operating points a single-resolution 640px profile would deploy silently. The profiler emits a JSON complexity report consumed by the pipeline's LLM narrative stage, turning profiling into an auditable, versioned artifact instead of a manual spreadsheet.