Radxa Fogwise Airbox edge AI box review – Part 1: Specifications, teardown, and first try

Radxa Fogwise Airbox, also known as Fogwise BM168M, is an edge AI box powered by a SOPHON BM1684X Arm SoC with a built-in 32 TOPS TPU and a VPU capable of handling the decoding of up to 32 HD video streams. The device is equipped with 16GB LPDDR4x RAM and a 64GB eMMC flash and features two gigabit Ethernet RJ45 jacks, a few USB ports, a speaker, and more.

Radxa sent us a sample for evaluation. We’ll start the Radxa Fogwise Airbox review by checking out the specifications and the hardware with an unboxing and a teardown, before testing various AI workloads with Tensorflow and/or other frameworks in the second part of the review.

Radxa Fogwise Airbox review

Radxa Fogwise Airbox specifications

The specifications below come from the product page as of May 30, 2024:

  • SoC – SOPHON SG2300x
    • CPU – Octa-core Arm Cortex-A53 processor up to 2.3 GHz
    • VPU
      • Decoding of up to 32x 1080p25 channels with H.265/H.264
      • Full processing of 32x 1080p25 channels with decoding and AI analysis
      • Encoding of up to 12x 1080p25 channels with H.265/H.264
      • JPEG up to 1080p600 (no typo, that’s 600 FPS) up to 32768 x 32768
      • Video post-processing such as image CSC, Resize, Crop, Padding, Border, Font, Contrast, and Brightness adjustments.
    • TPU – Tensor Processing Unit with up to 24 TOPS (INT8), 12 TFLOPS (FP16/BF16) and 2 TFLOPS (FP32) with support for TensorFlow, Caffe, PyTorch, Paddle, ONNX, MXNet, Tengine, and DarkNet
  • System Memory – 16GB LPDDR4X
  • Storage
    • 64GB onboard eMMC flash
    • M.2 M Key connector for 2230 NVMe SSD
    • MicroSD Card slot
  • Networking
    • 2x Gigabit Ethernet ports
    • Optional WiFi and Bluetooth via M.2 E Key module
  • USB
    • 2x USB 3.0 host ports
    • 1x USB Type-C Debug UART port
  • Power Supply – 20V via USB Type-C port, at least 65W
  • Dimensions – 104 x 84 x 52mm (metal case with active cooling)
  • Temperature Range – 0°C to +45°C
  • Compliance Certification – FCC / CE

The specifications and design are almost exactly the same as the Firefly AIBOX-1684X, but except for the SOPHON BM1684X (32 TOPS) used instead of the SOPHON SG2300x (24 TOPS), and the two M.2 sockets that don’t seem to be available in the Firefly AI box. At the time of the Firefly article (April 2024), I was told that “SG2300X supports open-source generative AI, while the BM1684X does not”, but it appears both chips are interchangeable for more on that below…

Based on the documentation, the Radxa Fogwise Airbox AI micro-server runs the CasaOS lightweight operating system offering a private cloud storage solution for home users. The company also offers a “Radxa Model Zoo” with ResNet-50, YOLOv5-det, YOLOv8-seg for object detection, recognition, and segmentation, and provides instructions to run LVMs or LLMs such as Stable Diffusion and Llama-3.

Fogwise “BM168M” unboxing

The Fogwise Airbox ships in a retail box that reads “Fogwise BM168M Edge Micro-Sever for AI”. Besides the typo, I was surprised by the size of the package as I expected something a bit larger similar to mini PC packages. The other thing is that it’s not called “Fogwise Airbox”, but “Fogwise BM168M” on the package.

Radxa Fogwise BM168M

In addition, if we look at the bottom side of the package, we can see the basic specifications that read “Power by SOPHON BM1684X” instead of “Powered by SOPHON SG2300X” as I would have expected. The package also lists the supported frameworks: PyTorch, ONNX, Baidu PaddlePaddle, Cafee, Tensforflow, MXNET, and Darknet.

Radxa Fogwise Airbox Package SOPHON BM1684X

Radxa Fogwise Airbox is shown on the sticker at the bottom, but since the teardown below will reveal more BM1684X strings, I asked Radxa if it was normal I have received a BM1684X device instead of one with SG2300x. I was eventually told I had received the latter as when the Model contains the string “R31” the system is based on SG2300x, while if there’s R22 it is powered by BM1684X. SG2300x and BM1684X are essentially the same chips and the main difference is that SG2300x is a SOPHGO device, while BM1684X refers to Bitmain. The latter is now focused on (crypto) mining hardware.

Radxa Fogwise BM168M Unboxing
Raspberry Pi 5 for scale

There’s nothing much inside the case, as the device itself takes 99% of the space, and we only have a “QC passed” sticker and a Warranty card (back side not shown on the photo above) with QR codes on the other side pointing to documentation (link not working, but I found it with a web search, see specifications section) and the community forum. This explains why the package can be so small as users will need to get their own 65W+ USB-C power adapter.

Radxa Fogwise Airbox review

The rear panel includes two USB 3.0 ports, WAN and LAN gigabit Ethernet ports, and a USB-C port for power plus ventilation holes. The left side features a microSD card slot and a USB-C debug port.

Radxa Fogwise Airbox Speaker

The right side has a few holes for the built-in speaker.

Radxa Fogwise Airbox teardown

Let’s have a look inside.

Radxa Fogwise BM168M bottom cover

We’ll need to remove the four stick rubber pads and loosen four screws to remove the bottom cover.

Fogwise Airbox BM168M teardown

This reveals the M.2 Key M and Key E sockets listed in the specifications as well as the cables from two WiFi antennas. There’s a thick thermal pad that covers a chip in the middle.

ASMedia ASM2806

It happens to be an ASMedia ASM2806 PCIe Gen3 x2 switch with four downstream ports. Let’s remove four standoffs to take the main board out of the enclosure. I also had to disconnect the wire to the speaker (shown on the left in the photo below).

Radxa Fogwise Airbox teardown

The SOPHGO SG2300x processor is on the CPU module and in contact with the metal case through some thermal paste.

SOPHON BM1864 RealTek RTL8211F0

The top of the AIM_1684X_V1 system-on-module also features two Realtek RTL8211FG gigabit Ethernet transceivers and two Micron MT53E1G32D4NQ-053 32Gbit (4GB) LPDDR4 memory chips, meaning there are also two others underneath for a total of 16GB.

MPS2323 VL805

A Monolithic Power Systems (MPS) MP7475 PMIC can be found on the bottom right of the module, and a VL805 PCIe to USB 3.0 bridge is present on the mainboard for the two USB 3.0 ports.

APW8713

The last notable part on the board is the APW8713 8W step-down converter. I did not remove the CPU module which is attached through a single B2B connector to the mainboard.

AICore SG2300X B2B connector

First try

I reassembled the device to give it a try. None of my USB-C phone chargers will reach 45W and the Raspberry Pi 5 USB-C power supply is limited to 27W, so I used a 100W GaN USB-C power supply from MINIX. I also connected an Ethernet cable to the WAN port. The system automatically started upon applying the power.

Radxa Fogwise Airbox Ethernet Power

I searched for the device with nmap but nothing new showed up…


So I connected a USB-A to USB-C to the Debug port to access the console and see what was going on…


The system is stuck in a boot loop… So it looks like I have to install the image myself…

So I downloaded the Fogwise Airbox B4 image and we’re told to flash it to a microSD card with tools like Etcher, but USBImager won’t take the file… and looking into the tarball it’s not your typical img file…

sdcard radxa airbox b4 image

So I think I’ll stop for today and carry on once the documentation has been updated… So I prepared a microSD card with a FAT32 parition and copied the file on it. After that I turned off the device, inserted the microSD card, and restarted it to start the flashing process.


This will take a few minutes and end with:


Let’s turn off the device, remove the microSD card (the case is hot so I used a pencil to do so), and boot it again. This time I got to a login prompt:


And the Airbox also shows with an IP address:


But no port 81 opened as we have installed Ubuntu 20.04 and not CasaOS (as advertised in the documentation):


We can use the command line through the serial console or SSH using linaro username and linaro password, and run a few commands to get system information:


There’s also a web dashboard on port 80.

Fogwise Airbox Dashboard

This time I’ll stop for now, and I have to figure out what to do next and learn how to use the system.

[One more small update… I’ve just realized CasaOS is not an OS, but a program installed on top of Ubuntu, Debian, etc…..


]

In the second part of the review, I plan to install the OS and run large language models (LLM) and large vision models (LVM) on the system. I’d like to thank Radxa for sending the Fogwise Airbox for review. It’s now available on Aliexpress for $331 plus shipping.

Continue reading Radxa Fogwise Airbox AI box review – Part 2: Llama3, Stable Diffusion, imgSearch, Python SDK, YOLOv8

Share this:

Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress

ROCK 5 ITX RK3588 mini-ITX motherboard
Subscribe
Notify of
guest
The comment form collects your name, email and content to allow us keep track of the comments placed on the website. Please read and accept our website Terms and Privacy Policy to post a comment.
19 Comments
oldest
newest
graffiti
graffiti
5 months ago

This has potential to be a very good home surveillance box if it can support frigate. Please can you test if frigate works on this box? Does the developer intend to support frigate.

Roberto
Roberto
5 months ago

lol but why? 30W idle when you can get 3W idle with an rk3566/rk3588 which is supported by frigate

urostor
urostor
5 months ago

To install the system, you might try to format an SD card as FAT32 and extract the tarball contents to it. Then it should flash to the eMMC on boot. You can monitor this via serial.

Anonymous
Anonymous
5 months ago

This thing does not seem to be good for anything. I dont get the reason why it exist. Too much heat, to less computing power, outdated SoC, closed source Software and Firmware, and so on.

There are bunch of boards out there that have same functionality, run mainline Linux and run with a fraction of the power.

fossxplorer
fossxplorer
5 months ago

Typo or does it really idle at around 30W?

back2future
back2future
5 months ago

[ “Idle power consumption is about 30.3 Watts”
What parts have most consumption influence? Configuration possibilities? (thx) ]

back2future
back2future
5 months ago

[ maybe because it’s useful for that conception?
“Fogwise AirBox is an embedded AI micro-server with arithmetic power up to 32TOPS@INT8, supports multiple precision (INT8, FP16/BF16, FP32), supports private GPT, text-to-image, and other mainstream AI model deployments, and is equipped with an aluminum alloy casing for deployment in harsh environments.”

and me often being astonished, why there are no configuration possibilities or monitoring tools, if that influence on power consumption on devices is getting more significant (compared to other SBCs)? (sounds unreasonable to You?)

and (btw) it’s not obvious what’s the default configuration? ]

back2future
back2future
5 months ago

[ “active by default”
another (unconventional) guess: comparison to temperature range from a ‘Mekotronics R57:
Temperature Range – -10 to 75°C’, while a Fogwise device can throttle above ~45°C (if not case temperature?, Dashboard describes with ‘chip junction temp’), there’s maybe limitation below 0°C without ‘heating’ from SoC (or partly from m.2 sdd) heatsink? For the Fogwise case this would be ~30-100W on an outer surface of ~57sq”/370cm² and additional airflow heat transfer. ]

John Smith
John Smith
5 months ago

I wonder if something like this would be good for a video box, like has enough A.I. TOPS to upscale dvd video 360p/480p to like 4k, one thing I doubt it has is quality video A.I. like many of the dedicated video boxes to improve video picture quality like getting rid of compression artifacts and lighting up shadows and improve overall watchability for 4k tv.

Zifeng ZHANG
5 months ago

Radxa have a demo named Real-esrgan-tpu_web can upscale any image or video to 2560×1920 with face enhance, actually it can upscale x4 by realesrgan model depend on the model input. please feel free to have a look or try this repo.

RaceFPV
RaceFPV
5 months ago

Please test this with Llama3, as one of its marketing perks was that it can run llama3 pretty smoothly.

Boardcon Rockchip and Allwinner SoM and SBC products