AAEON NV8600-Nano AI Developer Kit Review – Part 2: Benchmarks, features testing, AI demos with Nx Meta

In the first part of the review, I had a look at the hardware of the NV8600-Nano AI developer Kit featuring an NVIDIA Jetson Orin Nano 8GB module, booted it to Ubuntu 22.04, checked some system information, and made sure both the USB camera and Raspberry Pi Camera Module 2 (MIPI CSI) module worked fine.

I’ve now spent much more time with the Edge AI devkit, and report my experience in the second part of the review with some benchmarks, key features testing, and AI vision demos using Network Optix Nx Meta IP video management platform with the provided USB camera and an ONVIF-capable network camera.

AAEON NV8600-Nano AI Developer Kit / Jetson Orin Nano benchmarks

It’s the first time I’m testing a platform based on NVIDIA Jetson Orin Nano, so I’ll go through all the benchmarks I typically use for such reviews.

Let’s start with sbc-bench.sh:

There was no throttling detected, and the ultra-quiet fan did its job perfectly, with the maximum temperature reported being 50.9°C in cpuminer.

The Jetson Orin Nano 8GB module achieved 14820 DMIPS in 7-zip. Let’s compare some of the results against other Arm platforms, namely the Raspberry Pi 5, a Rockchip RK3588 system like the Rock 5B, and Radxa’s Orion O6 mini-ITX motherboard.

In terms of memory bandwidth and multi-core performance, the NVIDIA Jetson Orin Nano is somewhat behind the Rockchip RK3588 octa-core SBC, and the CIX P1 Armv9 12-core CPU found on the Orion O6 is significantly faster, as one should expect.

Let’s evaluate web browser performance with Speedometer 2.0 on Chromium.

85.7 runs per minute (RPM) compares to 63.5 RPS on the Raspberry Pi 5, and 80.7 RPS on Khadas Edge 2 (RK3588S SBC). Note that this benchmark varies a lot on RK3588(S) targets depending on settings/optimization. I also ran Speedometer 3.0 (5.21 points) to have a reference point in the future since Speedometer 2.0 is now deprecated.

One of the main selling points of NVIDIA Jetson modules is their GPU, and the Jetson Orin Nano 8GB features an NVIDIA Ampere GPU with 1024 CUDA cores and 32 tensor cores.

To compare with other Arm platforms, I tried to run glmark2-es2-wayland, but I refused to run. Not exactly surprising, since the Ubuntu 22.04 OS on the board runs the X11 windowing system, instead of Wayland on most other boards running Ubuntu 24.04.

I could run glmark2-es2 without issues:

Results:

2,488 points in glmark2-es2. I can compare this result with one RK3588 benchmark on Ubuntu 22.04: 1042 points on Mixtile Core 3588E SoM. All other results I could find rely on glmark2-es2-wayland, and the results can’t be compared at all.

NVIDIA Jetson modules stand out in the Arm world because they embed a full-featured GPU with support for OpenGL 4.6 rather than only OpenGL ES used in the GPU of most other Arm platforms. So we can run glmark2 benchmarks with full hardware acceleration too:

For reference, I ran glmark2 on my laptop using the iGPU on an Intel Core i5-13500H Raptor Lake SoC and got 3571 points.

Another way to check how well GPU support is implemented is to run the WebGL Aquarium demo on Chromium.

6 FPS with 500 fish and 0% GPU utilization. This does not look good, and indeed, chrome://gpu confirms that OpenGL is disabled and graphics features rely on software rendering.

GPU acceleration/WebGL can be enabled on Chromium on the Jetson Orin Nano, but it requires rebuilding Chromium from source. I’ll skip that part as it’s out of scope of this review.

Storage and USB

Let’s now test the SSD that ships with the devkit:

The SSD delivers 2.14 GB/s reads and 1.08 GB/s writes.

I also tested the six USB 3.0 ports with lsusb and iozone command line utilities using an ORICO M234C3-U4 M.2 NVMe SSD enclosure.

Example output for the top left port (USB #1) on the front panel:

Results from left to right for all six ports:

• USB #1 (top) – USB 3.0 – 10,000 Mbps – Read speed: 946 MB/s
• USB #2 (bottom) – USB 3.0 – 10,000 Mbps – Read speed: 940 MB/s
• USB #3 (top) – USB 3.0 – 10,000 Mbps – Read speed: 953 MB/s
• USB #4 (bottom) – USB 3.0 – 10,000 Mbps – Read speed: 953 MB/s
• USB #5 (top) – USB 3.0 – 10,000 Mbps – Read speed: 951 MB/s
• USB #6 (bottom) – USB 3.0 – 10,000 Mbps – Read speed: 951 MB/s

Everything works as advertised.

Networking

AAEON NV8600-Nano comes with four gigabit Ethernet ports, which I tested with iperf3 full-duplex (bidir option) using a Ubuntu 24.04 laptop.

• LAN1 – Rx: 938 Mbps; Tx: 744 Mbps
• LAN2 – Rx: 939 Mbps; Tx: 936 Mbps
• LAN3 – Rx: 938 Mbps; Tx: 937 Mbps
• LAN4 – Rx: 938 Mbps; Tx: 936 Mbps

LAN2-4 ports are each connected to an Intel I210 Gigabit Ethernet controller and work at optimal speed. LAN1 will be fine for most purposes, but avoid using it if you plan to have heavy bidirectional traffic there, and rely on one of the LAN2-4 ports instead.

GPIOs

I didn’t test GPIO directly due to time constraints, but they are there with three gpiochip devices:

I could also list them with gpioinfo:

You’ll find the pinout diagram for the 40-pin GPIO header, and all connectors on the board for that matter, in the user manual, where there are also some basic instructions explaining how to use the GPIOs.

AI testing – Network Optix Nx Meta

While it’s possible to test AI workloads using Jetson Platform Services from the JetPack 6.2 SDK on Jetson Orin hardware, the NV8600-Nano DevKit comes with Network Optix (Nx) Meta platform and NX AI Manager Plugin, which can help users get started faster and accelerate development using a USB camera. So that’s what I will use in this review.

Starting Chromium will automatically enter the Nx Meta dashboard. Let’s click the “Setup New System” button.

Now we’ll be asked to enter a system name (devkit-aaeon-cnx) and create a password in the next two windows.

The system is not ready for use, and we can log in to it with the admin user and our password.

The dashboard’s left menu features four main parts: System Administration, Cameras, Users, and Servers.

When I went to the “Cameras” section, there was already one camera. I thought it might have been a camera used by AAEON before they shipped the sample, but actually the system automatically detected my Reolink TrackMix PoE security camera because it supports the ONVIF standard. I can’t quite remember the password, but maybe I’ll try it later.

AAEON documentation provides further instructions to get started with Nx Meta. We first need to create an account on meta.nxvms.com.

Click the “Create Account” button on the website to create an account with your email. After that, download the Nx Meta client for your OS. I have a laptop running Ubuntu 24.04, so I was redirected to the “Ubuntu x64 – Client installer”, but it’s also available for Windows, Mac OS, Arm (NVIDIA Jetson, Qualcomm Laptop, Raspberry Pi…), and mobile devices running Android or iOS.
After installation, I could start the “Nx Meta” client program, and the devkit-aaeon-cnx server was automatically detected.

I clicked on it, and was asked to enter the credentials I created earlier in the web dashboard.

Then, I had to right-click on the server and select “Server Settings”.

The important part is to tick the box “before Autodetect built-in and USB webcams” and click OK.

The UP USB webcam I received with the kit is now detected. But we are not done just yet. Right-click on the camera and select Camera Settings.

Navigate to the Plugins section, and enable “NX AI Manager”.

At this point, the client complains that “This system is not registered to a cloud user. Add a cloud user through the System Administration”. That’s because I did not log in to the cloud. See the notifications section from previous screenshots, reading “Check out Cloud – Connect to your system from anywhere”. So I went back there, clicked the Connect link, and logged in using the credentials (email + password) from the cloud services I had just registered with.

That’s better. I can now install Nx CPU or ONNX-CUDA. I selected the latter to make use of the Jetson Orin GPU. This will automatically load the “80-Classes Object Detector [320×320]” pipeline, and you can configure it to select which type of objects to detect from the list.

We can now play… Click on Object Search, and rounded boxes will now show around known objects. A keyboard and two TV monitors were detected in the screenshot below, and you can also see screen captures for various objects on the right panel.

Let’s see what else the NX AI Plugin has to offer. In the Plugin tab, click “Manage Devices” and the Models section. If nothing shows up, click “All available models” or “Nx Demo Models”.

Three models are available:

• Face locator
• 80-Classes Object detector [320×320] – Currently selected
• 80-Classes Object detector [640×640]

So click on the Sites icon, select your server (devkit-aaeon0cnx), and you can now configure pipelines from there. I kept the 80-classes object detector, and selected “Add a new Pipeline”…

… before clicking “Added to usb_cam-FHD Camera: FHD Camera pipeline”.

Now there are two model pipelines assigned to the camera: the object detector and the face locator. We’ll leave “no postprocessor selected” as it is for now.

Face and object detection pipelines are working at the same time. The screenshot below shows the detection of two faces from the Face locator, two persons, and one teddy bear from the object detector.

Nx Meta allows for more complex tasks as well. Let’s go to Camera Rules to create an Analytics Event to count objects and overlay the results on top of the camera output in the client.

We also adjusted the Postprocessor as shown below, with Object counting enabled for the object detector, but not the Face locator. Other options include Illegal Dumping, Line Crossing, and Loitering Detection (standing somewhere for no obvious reason).

We held a few objects in front of the camera, and although not all were detected at the time of the screenshot, we can see the overlay on the bottom right with two persons, one bottle, a keyboard, and one cell phone (the cup).

At this point, I decided to connect my PoE security camera over ONVIF (after finding its password), and it worked too. I could assign the NX AI Manager Plugin to it and detect objects like cars, persons, and chairs. The PTZ function works too, and presets from the Reolink app (e.g., gate and porch) are accessible from the Nx Meta client. Pretty neat!

We also tried the Line Crossing postprocessor. I was confused at first, since I could not find where to draw the line, but it’s in the pipeline configuration window, and I just had to scroll down a bit. After everything was properly configured, it could detect line crossing and show results overlaid on the video.

However, it’s not practical to do that on a tracking camera, since the line is drawn at a fixed location in the window, and will not automatically stay there. The screenshot below better explains what I mean :).

Conclusion

AAEON NV8600-Nano AI developer kit is a solid NVIDIA Jetson Orin Nano development platform with support for multiple cameras through MIPI CSI (e.g., Raspberry Pi Camera Module 2), six USB 3.0 ports, and four gigabit Ethernet ports, as well as three M.2 sockets for expansion, one fitted with an NVMe SSD for this review. It’s essentially a NVIDIA Jetson Orin Nano Super Developer Kit with extra features.

Everything I tested just worked, except for a few details. First, GPU acceleration is not enabled in Chromium. Note that it’s the same for all Jetson modules using JetPack 6.2, and can be enabled by manually recompiling Chromium (not tested here). Then the GbE RJ45 port on the left attached directly to the NVIDIA Jetson Orin Nano module does not perform as well as the other three ports using Intel I210 controllers for bidirectional transfers, not that it would matter if you are just going to connect a camera to it.

The highlight of the review was the Network Optix Meta (Nx Meta) platform, which makes it easy to run AI workloads on USB or networked cameras. I could test the object detector and face locator models on the UP USB camera in a few minutes, and set up more complex scenarios to count objects in the frame. I did the same with an ONVIF-enabled Reolink TrackMix PoE camera and also configured a Line Crossing pipeline for it. While getting started with Nx Meta is relatively straightforward, I found the user interface to be not always intuitive, so I wasted some time navigating part of it.

I’d like to thank AAEON for sending the NV8600-Nano AI developer kit along with a USB 2.0 camera for review. The devkit can be purchased on the UP shop for $649 plus taxes (if any) and shipping, and the UP HD camera sells for $35. That’s quite more expensive than the NVIDIA Jetson Orin Nano Super Developer Kit, but it ships with extra accessories (256GB SSD with JetPack 6.2 SDK + Raspberry Pi Camera Module 2), offers extra features and ports, is designed to operate in the -25°C to 70°C temperature range, and availability may not be as restricted as for the low-cost devkit from NVIDIA.

Jean-Luc Aufranc (CNXSoft)

Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.

Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress. We also use affiliate links in articles to earn commissions if you make a purchase after clicking on those links.