Compare YOLO, ResNet, MobileNet, and EfficientDet side-by-side across Jetson, Coral, and Hailo hardware. Configure target platform, runtime, and precision to benchmark FPS, latency, memory, accuracy, and power per frame.
The comparator loads hardware-specific benchmark data for your selected platform and runtime, then displays FPS, latency, memory, and accuracy metrics side-by-side for all available model architectures. Benchmark rows use measured data from representative deployments; interpolated rows extrapolate from available benchmarks using resolution and batch-size scaling; theoretical rows are planning-only estimates. Accuracy (mAP50-95) is currently sourced from official vendor documentation and standard COCO val2017 datasets.
Model efficiency is determined by architecture complexity (layer depth, parameter count), input resolution (FPS scales roughly inversely with pixel count), runtime precision (INT8 ≥ FP16 ≥ FP32 in terms of throughput), and hardware accelerator type (Jetson GPU, Coral Edge TPU, Hailo-8 VPU). Memory footprint scales with model size and batch count; inference latency is primarily determined by FPS and layer topology. FPS/TOP efficiency combines hardware utilization (TOPS) with actual throughput (FPS) to normalize across architecturally different accelerators.
Detection accuracy and inference throughput are often in tension: smaller models (MobileNet, ResNet-50) achieve higher FPS but lower mAP; larger models (YOLOv8x, EfficientDet-L) reach higher accuracy but lower throughput. The optimal model balances detection quality (mAP needed for your use case) against inference capacity (streams × FPS required). Always measure accuracy on representative data from your deployment environment, since domain shift and scene complexity can significantly impact real-world mAP.
FPS per TOPS of AI performance — a standard cross-hardware efficiency metric used by hardware vendors including Hailo, TI, and Xilinx in benchmark papers. Higher means more efficient use of hardware AI capacity. Useful for comparing architecturally different models on the same accelerator.
Power/frame is calculated from the hardware's nominal power mode envelope divided by measured FPS. It is not measured draw — actual power varies with workload, thermal state, and pipeline. Use a power meter (e.g., INA219) for accurate measurements.
mAP50-95 on COCO val2017 is currently only available for YOLO11 and YOLO12 variants (sourced from Ultralytics). ResNet, MobileNet, and EfficientDet mAP values would require separate benchmark runs on the same dataset for a fair comparison.
A model is Coral-compatible if it supports INT8 precision and fits within Edge TPU on-chip SRAM (~8 MB of INT8 weight storage). MobileNet V1, V2, and SSD variants qualify. Most YOLO variants are too large for full on-chip execution.
Benchmark-backed rows are ±10–15% of real measured throughput. Interpolated rows are ~65% accurate. Theoretical rows are planning-only (±30–50%). Memory estimates are ±30%. Always validate on device before finalising model selection.