Jetson Nano/NX vs. Raspberry Pi 5 Performance

ECE MAE 148 Final Project

Team 4 Winter 2025

Table of Contents

1. Team Members
2. Abstract
3. Promises and Stretch Goals
4. Hardware
5. Software
6. Final Metrics
7. Accomplishments
8. Challenges
9. Documentation
10. Potential Improvements
11. Course Deliverables

---

Team Members

Jake Honma - MAE Controls & Robotics - Class of 2026 - jhonma@ucsd.edu

Harsh Salva - MAE Controls & Robotics - Class of 2026 - hsavla@ucsd.edu

Andrew Dunker - ECE ML & Controls - Class of 2025 - adunker@ucsd.edu

Jingli Zhou - Math/CS- Class of 2027 - jiz228@ucsd.edu

---

Abstract

Our project aimed to measure the performance of deep-learning models trained on DonkeyCar on the Jetson Nano, Jetson Xavier NX, and the Raspberry Pi 5 (w/ & w/o Hailo AIHAT+).

---

Promises and Stretch Goals

Promised

• Benchmarked performance for:

  • Jetson Nano w/ Tensorflow/TensorRT
    
    
    

  • Jetson NX w/ Tensorflow/TensorRT
    
    
    

  • RPI w/ Tensorflow/HailoRT
    
    
    

• Documentation for the Process

Stretch Goals

• Benchmarked performance for models with different resolutions
• Fully integrate the Hailo Model with DonkeyCar

---

Hardware

---

Software

---

Final Metrics

---

Accomplishments

• Document the process for setting up DonkeyCar on the RPI 5.
• Document the process for converting linear (.h5) models into TensorRT directories and Hailo Executable Files (.hef through AIHAT+).
• Benchmarked performance for 160x120 resolution model across all hardware.
• Load and integrate Hailo Executable Files (.hef) into DonkeyCar.
• Create a .h5 to .hef converter in Python

---

Challenges

• Going into the project our Jetson SD card was corrupted. Initially, we used a backup image, but this was using a lower version JetPack that did not allow us to test models. We then had to completely reflash our SD card, reinstall all the dependencies, and setup up DonkeyCar again.
• There is out-of-date documentation for converting linear models to utilize TensorRT GPU acceleration. This needed to be updated and varied based on different dependencies across the Jetson Nano & Jetson NX.
• Without GPU access in a previously set-up docker container, TensorRT conversion could not be completed.
• Due to the recency of which Hailo Executable Files were introduced and the limited documentation, attempting to integrate .hef files into DonkeyCar was difficult.
• We could not physically test the performance of models through manage.py drive due to VESC issues at low speeds.
• Raspberry Pi 5 requires more power to run with the AI Hat+ than we were able to provide with the soldered on USB-C connection from the DC-DC converter on our car.

---

Documentation

• TensorRT Directory Conversion
• RPI Setup & AIHAT+ Benchmarking

---

Potential Improvements

• Finish optimizing the integration of Hailo in DonkeyCar
  • Though we were able to get a .hef file to run in DonkeyCar, based on the results and what we expected, the integration could be performed more efficiently. Although we did make a more efficient integration, this persistent model integration remains untested at this point.
• Test Models at Different Resolutions
  • Test and optimize models based on OAKD lite resolution and hardware (Jetson vs. Raspberry Pi 5) to find the highest-performing resolutions for deep learning.
• Continue to Optimize DonkeyCar Training
  • Finish optimizing memory management to allow training of larger resolution models

Course Deliverables

• Autonomous Laps
• GPS Laps
• OpenCV Laps
• Final Presentation Slides

---