Tutorial 14 — YOLO11 vs YOLOv8 Benchmark on Jetson Orin

Run a head-to-head benchmark of YOLO11 and YOLOv8 directly on your Jetson hardware. Measure real FPS, inference latency, and mAP accuracy — and decide which model is right for your application.

What you will learn

• Key architecture differences between YOLOv8 and YOLO11
• How to run a standardised benchmark on Jetson hardware
• How to measure FPS, latency, memory usage, and mAP50-95
• When to use YOLO11 vs YOLOv8 for your specific use case
• How TensorRT FP16 vs FP32 affects speed and accuracy

Step 1 — Run the benchmark script

cd ~/tutorials/14-yolo11-benchmark
python3 benchmark.py --models yolov8n yolov8s yolo11n yolo11s --imgsz 640

The script runs each model on 500 frames from a standard test video and prints a comparison table.

Step 2 — Benchmark code

from ultralytics import YOLO
import time, cv2

def benchmark_model(model_path, source, frames=500):
    model  = YOLO(model_path)
    cap    = cv2.VideoCapture(source)
    times  = []

    for _ in range(frames):
        ret, frame = cap.read()
        if not ret: break
        start = time.perf_counter()
        model(frame, verbose=False)
        times.append(time.perf_counter() - start)

    avg_ms  = sum(times)/len(times) * 1000
    avg_fps = 1000 / avg_ms
    print(f"{model_path:30s}  {avg_fps:.1f} FPS  |  {avg_ms:.1f}ms/frame")
    return avg_fps, avg_ms

benchmark_model("yolov8n.engine", 0)
benchmark_model("yolo11n.engine", 0)

Expected results on Jetson Orin Nano (TensorRT FP16)

✅ Next: Tutorial 15 — TensorRT Export Guide | Back to Jetson Kit