Analysis of AI Workloads on Nvidia Jetson Nano

This project analyzes the inference performance of several well-known AI models on the NVIDIA Jetson Nano under different stress conditions. The main goal is to assess the inference time of classification and object detection models in both ideal (baseline) and stressed environments.

Authors

• Antonio Boccarossa – a.boccarossa@studenti.unina.it
• Francesco Brunello – f.brunello@studenti.unina.it

Course: Real Time Systems and Industrial Applications
Institution: Università degli Studi di Napoli Federico II
Supervisors: Marcello Cinque, Andrea Marchetta
Date: 04.07.2025

---

Models Analyzed

• Classification Models:
  • Inception-v1
  • Inception-v4
  • MobileNet-v1
  • MobileNet-v2
• Object Detection Models:
  • SSD-MobileNet-v1
  • SSD-MobileNet-v2

---

Device: NVIDIA Jetson Nano

• GPU: 128-core Maxwell
• CPU: Quad-core ARM Cortex-A57
• Memory: 4 GB LPDDR4
• OS: JetPack SDK (pre-installed)
• Storage: External SSD used for image datasets

The environment was set up by flashing the OS to an external SSD and installing dependencies required for jetson-inference.

---

Setup & Tools

• Inference Framework: jetson-inference
• Stress Testing Tool: stress-ng
• Languages: Python (version 3.6)
• Libraries Used: jetson.inference, jetson.utils

Scripts were executed both in a clean "Golden Run" and under stress conditions.

---

Stress Scenarios

Stress tests simulate real-time load using the following stressors:

Test Type	Command
CPU	stress-ng --cpu 4
Virtual Memory	stress-ng --vm 4 --vm-bytes 2.5G
Mem Copy	stress-ng --memcpy 8
Interrupt	stress-ng --clock 4 --aio 4 --aio-requests 30
Open	stress-ng --open 4
Fork	stress-ng --fork 4
UDP	stress-ng --udp 4

---

How the Tests Work

Each test involves:

1. Running inference on 108 images per model.
2. Logging the average inference time for each batch.
3. Repeating the test 30 times per configuration to build a statistically valid sample.

Each script receives CLI arguments:

• filename: Path to image folder
• testtype: Name of the stress test (e.g., golden, cpu)
• --network: Name of the neural network model (e.g., googlenet, ssd-mobilenet-v1)

---

Sample Command

python3 execute_inference_obj_detect.py

python3 execute_inference_img_detect.py

---

Results

The inference times were measured and logged for each model in each scenario. Results are stored in .txt files in the format:

<network>_result_<testtype>.txt

Each file contains 30 average inference times (one per run).

---

References

• Jetson Inference GitHub
• Inception-v4 Paper
• SSD MobileNet Architecture
• Jetson Nano Specs