In mid-August this year I took an interest in the Radxa Linkr KVM – and all I have to say on that is that it works perfectly over WiFi, offering KVM features in a slim stick. HDMI in, USBC out to the target machine – USB power.
It’s just as well that it works well, because the company then sent me their Q8B board (complete with heatsink but no box or PSU) to play with. The latter sat on a shelf until this morning because it turns out what they mean by USB C is PD (Power Delivery) USB C – and 20v at that. I said about the PSU: “no problem – I have loads of them” and indeed I do have loads of QuickCharge (QC3) and PD USB supplies with USB-C outputs.
And so it was that I ended up with an SBC with which I could do zilch, because it simply would not turn on. Eventually I twigged that not all PD power supplies are the same. There are some (but by no means all) which have 20 volts maximum, at up to 3 amps or so. But none of mine went past 12 volts.
So, two days ago, I ordered a reasonably priced—I think about 10 euros—power supply from Amazon with up to 20 volts at 3.5 amps output out of the USB-C connector. While waiting, I went off on my PC to grab a Linux distribution recommended by Radxa. I then put that image onto a micro SD card all ready to go. Of course, without a working Radxa computer, I couldn’t really prove it was working.
Next morning I popped out to the local tobacconist to pick up my post and there was the power supply. That afternoon, I plugged the power supply into the wall and into the Radxa unit. I plugged the HDMI output from the Radxa unit into the KVM along with one of the USB connectors (not USB-C) on the SBC into the KVM. I plugged the SBC into the power supply and a green light came on. Magic!
I went to my PC from where I could control and view the output of the SBC.
And I noticed it was waiting for a login from either Radxa or Rock. I went to the Radxa website on the information on this SBC and sure enough, login user radxa with password radxa.
And here we are. Of course, just to be awkward, while writing this, the thing went to sleep so I’ll try again.
Looks pretty standard up to now.
Once into Firefox (bottom left) I went off to settings to ensure the PC never times out on me again – so far, so good. I note the unit has no Bluetooth – not a problem as I have several BT dongles… – (I’m hooked into the SBC by Ethernet incidentally – that’s why I didn’t get pestered for passwords).
Next – the Bluetooth – making sure of no surprises there. I’m using one USB connector – there are 3 more and then a pair of USB-C connectors. I have no idea what model my Bluetooth dongles are. I’ve had them lying around for ages.
They say Bluetooth 4, and on the side, CSR4. That’s it. Wish me luck. Under power – settings I found all manner of settings including BT. Can it really be this easy?
Once I inserted the dongle, a couple of devices magically appeared which I didn’t even know had BT (WiFi IOT gadgets one of which I just created in ESPHome this week).
Ok, that’s Bluetooth sorted. Next, I needed to get the device running on Spanish time.
Not the best of starts – the board was definitely hooked into the Internet but only USA seemed to work… AHAH, I’d not actually selected one of the Ethernet ports, and so I wasn’t connected to the Internet, but now I am.
Erm, not so fast..
Its definitely hooked up – Wikipedia works – so why can’t I change location? I’m in a completely different time zone to the USA – and China – so asking for help mid-afternoon is a non-starter.
Ok it turns out there are various ways to get there.
This part at least is a fairly standard Linux setup… normal Ethernet (2 of them) or USB Ethernet… not seeing anything about WiFi – not essential but it would have been nice. It seems WireGuard VPN is built in – or you can add others.
Oh, hardware, before you ask.
AHAH, terminal – SUDO RSETUP – I found GB (UK) needed adding – seconds later… wheee.
I did think of asking ChatGPT for help with this – but as it’s a new board – I decided to go old-school initially (ie hunt and peck)..
Ok, reboot it seems..
Next challenge – WIFI – I could see no signs of WiFi even with my WiFi dongle plugged in – I won’t bore you with the details only half of which I understood – but on my PC I pulled up my new friend ChatGPT who had me open up the Linux terminal in the main menu – type in a bunch of commands into the Q8B including sudo apt update – sudo apt upgrade – and some specific installs – and lo and behold – in the terminal….
You may not find that exciting – I did as the WiFi with my cheap USB dongle clearly works..
Thats the fun bit done – and now – the Dragon Q8B, based on the Qualcomm Snapdragon 8cx Gen 3 (SC8280XP) has an 8-core Kryo CPU, Adreno 690 GPU and Qualcomm Hexagon AI Engine rated at up to 29+ TOPS. Radxa specifically positions it for local/edge AI, including compact LLMs. Or so it says.
After following the simple instructions on the Radxa website – I ran a test in the terminal…
fastrpc_test -a v68
which produced this:
Well, it LOOKS encouraging.. next step is QAI APPBuilder – I could mess with lanuage models and get myself lost – or use Real-ESRGAN which takes an image – messes with it and Raxda even have a specific example…
I’m not going to go into the whole build thing – the Radxa site is comprehensive and tells you what to install and how to install it – for reference for initial messing around I’m using a 64GB microSD – I suspected even then that when I get into this I’m going to need an M2 drive – the board supports up to 2 drives.
Next I followed instructions from the Radxa site to download half the world onto my little 64GB micoSD – I should point out the heatsink is running quite cool throughout all of this.
And…
python run_inference.py –model real_esrgan_x4plus –args “–chipset 6490 –bin –input_image_path models/computer_vision/super_resolution/real_esrgan_x4plus/assets/input.jpg –no_show”
Shortly thereafter – a 512*512 image (input.jpg file and a 2048*2048 image (output.png). Nothing major but proves the whole thing works…
And now I can move on to something much more interesting than super-resolution: YOLO object detection.
That will let me point the Q8B at an ordinary photograph and have its NPU identify things such as people, cars, dogs, chairs, etc., with bounding boxes.
That’s a much better demonstration of what the Q8B’s advertised AI capability can actually do.
This was right at the edge of my abilities – but in that terminal – I managed to get this to pop up..
Some genuine detection taking place on my new RADXA board… Well, it’s a start….
And now, hand it one of my real photos – here’s the test – I’m looking forward to this…
SO I put in a totally random photo from one of my holiday trips – and this is what popped up in the Linux terminal seconds later.. (apparently the actual detection took all of 9ms).
For my next trick – not prompted – I stuck an old USB microphone dongle I had lying around. arecord -l showed it to be present… so next I tried arecord -D plughw:1,0 -d 10 -f cd test.wav and said “This is Pete testing the Dragon Q8B. If this works, we’re going to make this little thing do some clever stuff.”. After 10 seconds, the prompt returned.
To cut a long story short – I found my old Logitech H800 Bluetooth/USB headset. I lost the USB dongle years back but the Bluetooth is intact and I’d already dropped a USB BT dongle into the computer – went to te normal Linux set up and added the headset – no problem.
Next I had to add a couple of packages to handle audio processing but I’m afraid all the audio models I tried failed. But then it would help if I knew what I was doing…
Ok, image classification on the above image…
python run_inference.py –help-model googlenet
python run_inference.py –model googlenet –args “–chipset 6490 –image /home/radxa/photos/myphoto.jpg”
restaurant 46.4%
dining table 6.9%
potter’s wheel 6.3%
plate 2.7%
patio 2.2%
1.82 ms. Not bad.
Still using my microSD – no big M2 drive…
python run_inference.py –model inception_v3 –args “–chipset 6490 –image /home/radxa/photos/myphoto.jpg”
ah, well, no laptop in my photo… no WOK that I’m aware of… Restaurant – well, ok.
So, all you have above is the summary – ChatGPT and I did a lot of work installing dependencies which would have proved tiresome for you had I gone into it – but as this is my first attempt at looking at an SBC with local AI hardware – up to now I’m having fun and all on a simple microSD. Today I got the urge to fit an M2 drive and went off to Amazon to see that prices have gone up recently – then just by fluke opened up a USB drive enclosure to find a tiny M2 drive in it – 256GB. We did some checks to find very disappointing speeds over USB-C.
This morning I pulled out the little drive from its enclosure – half the length of the ones Radxa recommend… but the connections looked right so I took a chance and inserted it. 15 minutes later I had the entire microSD copied over to the M2 and the latter mounted after which I powered off, removed the microSD – powered up and NO PROBLEM with read speeds of 1.5G/second and write speed 1.2G/s.
This is fun and this is WAY more than just a Linux SBC.