24 Orange Pi Zero Boards Cluster and ArmWizard’s Debian Image

ArmWizard forum member “-W.-” needed a cluster to test his firmware build and deployment solution for Orange Pi Zero boards which can deploy the firmware to multiple boards from one single board.

So he procured 24 Orange Pi Zero boards, a 24-port switch, some USB hubs and debug board, as well as some other accessories, power supplies, and two wooden planks. That’s the results after assembly.

Click to Enlarge

So we have 4 tower of six Orange Pi Zero boards fairly neatly connected to the 24-port Ethernet switch with 30 cm Ethernet cables, and four black USB hubs for the USB to TTL debug boards. The side view below shows the boards are powered by micro USB cables connected to three different power supplies, cooling achieved via four fans,  and the power extension is hidden right underneath close to the Ethernet switch.

Click to Enlarge

The cluster will also be used for machine learning latter on,  at least to detect potential bottlenecks due to the relatively slow network (100M Ethernet).

What’s also interesting is that he won’t be running an Armbian image like many of us do, but as I found out on Orange Pi forums, ArmWizard’s Debian 9.4 image with mainline Linux 4.15 instead.

ArmWizard is a French community for hobbyists and professional interested in Arm development boards. Everything in French language, and they describe the goals for the reference images in details, but basically the firmware images target various embedded cards and are build periodically (typically daily) with three types offered:

  • minimal – Bootable minimal image with access to SSH.
  • desktop – Debina with Xfce environment.
  • server – Based on minimal image, but adds better security.

At this time, only the minimal version is available, and you’ll find images, not only for Orange Pi Zero, but also for Orange Pi R1, and Cubieboard 2/3 boards on the Download page. There are other boards listed, but no image is currently available, support is only planned.

Leave a Reply

9 Comments on "24 Orange Pi Zero Boards Cluster and ArmWizard’s Debian Image"

avatar
  Subscribe  
newest oldest most voted
Notify of
Jerry
Guest

Raspberry folks did this long ago. The only difference is, RPi boards have plenty of GPU power for all sorts of stuff. Is that cheap board as energy efficient as the new RPi 3B+?

tkaiser
Guest

> Is that cheap board as energy efficient as the new RPi 3B+?

LOL! Efficient? The 3B+ is wasting way too much energy for its lousy performance. An Orange Pi Zero with appropriate settings (Armbian) idles at a third of what the RPi 3B+ needs:

https://libre.computer/2018/03/21/raspberry-pi-3-model-b-review-and-comparison/

My Orange Pi Zero as fully working web server consumes less than 700mW *active* while the RPi 3B+ wastes almost 2W when doing nothing. But of course such an OPi Zero cluster is pretty useless especially when running with unpatched mainline kernel since then not even cpufreq/DVFS scaling is working.

The above cluster is a nice example how inefficient and weird SBC clustering is (way too much space needed just for cabling) and there are just a few boards out there where it would make some sense depending on the use case. E.g. ODROID MC1 or the NanoPi Fire3 for the same 35 bucks as an RPi but way faster, smaller and providing better heat dissipation: http://www.friendlyarm.com/index.php?route=product/product&path=69&product_id=206

Jerry
Guest

It can basically run any OpenCL program. For example demanding photography processing. https://www.raspberrypi.org/forums/viewtopic.php?t=194952

You could even do something like anti-virus scans with the GPU
https://www.extremetech.com/computing/267716-intel-microsoft-plan-to-offload-antivirus-scanning-to-the-gpu

tkaiser
Guest

Did you visit the links you posted (or even try to read and understand)?

This OpenCL implementation is nothing more than a proof of concept, not even completed and far from being usable. It’s the result of a single person’s ‘masters thesis’ and of no use to anyone else. Also it seems you don’t know OpenCL. You don’t write ‘programs’ but ‘kernels’ trying to solve a specific problem and with a broken/incomplete OpenCL implementation you won’t be able to solve any real problem. Performance of the PoC is also low as hell due to VC4 being that old and limited.

And the virus-scan link is about Intel GPUs used in PCs/Macs and not slow BroadCom toys like the RPi. In the PC/Mac world GPGPU and OpenCL is already working pretty well. On my main platform (macOS) OpenCL has been promoted to developers starting ~9 years ago. It is widely adopted by Apple in the OS itself (with every new OS version more and more stuff runs on the GPU and frees up the CPU cores) but 3rd party developers mostly still ignore it.

The only real use of ‘GPU power’ on the RPi I know of is one case of ‘deep learning’ on the QPU cores. But that’s far away from using ‘GPU power for all sorts of stuff’ as you claimed.

TL;DR: Unless you are a bare metal programmer you can *not* make use of ‘GPU power for all sorts of stuff’ on a Rasperry Pi. And as usual getting closer to these RPi devices means only having to deal with an insane amount of ignorance and stupidity.

Singman
Guest

Ok, everyone got your point, you’re an absolute fan of Orange board. Now, calm down and maybe take your opinion to somewhere else.

tkaiser
Guest

> Ok, everyone got your point, you’re an absolute fan of Orange board.

At least you failed to get my point it seems. 😉

I was not talking about Orange Pis at all. How could you’ve missed that?

Just spreading some information about the freshly released RPi 3B+ not being ‘energy efficient’ but just showing the worst performance/watt ratio of any SBC currently known. And then adding some information about ‘GPU power’ on these Raspberries (you can *not* make use of their GPU capabilities in a general purpose way. For this you need more expensive SBC for example based on NVIDIA Tegra SoCs).

wens
Guest

cpufreq for H3 was pushed out to -next today. Though support for SY8106A is still pending.

tkaiser
Guest

> RPi boards have plenty of GPU power for all sorts of stuff

Care to provide at least one example of how one can make use of this ‘GPU power’ in cluster applications, especially ‘for all sorts of stuff’?

theguyuk
Guest

Nice fire proof wood they have their.

Apart from blowing dust around what are the open air fans doing?