Hopefully this is not a stupid question but I came across Joshua-Riek/ubuntu-rockchip: Ubuntu for Rockchip RK35XX Devices (github.com) when searching for my video playback issues. Does anyone know how this compares with the current armbian release with vendor kernel? Are all the fixes from that release integrated in armbian or does it go the other way?

Give more info about your video playback issues.

 

I'm using Armbian Jammy and the only issues that I had when I installed and set everything up was in Kodi.  I was using Matrix, which was available at the time.  It had issues playing XVid and AVI files.  Recently I found that the next version, Nexus, was available, I believe in liujanfeng's PPA.  In any case, I backed up my SD card and set it up on another OPi5 I have.  I then upgraded Kodi to Nexus.  Low and behold, those XVid and AVI files now play properly.  I'm guessing there was an issue with codecs.

 

In any case, I'm sticking with my Armbian Jammy now.  I tried the vendor kernel and weird stuff was happening.  So I'll wait until the updated video/gpu drivers are more mature.

 

I've seen the Joshua Riek stuff mentioned frequently but as mine is now working, I have no incentive to continue.

Sadly Joshua's excellent work is finding it very tough and taking a toll - https://github.com/Joshua-Riek/ubuntu-rockchip/discussions/1104

 

He starts to feel the pain we feel for years already. It is hard to not tell people simply to fuck off. When I first came across his project I thought he must be crazy to do this all by himself. Kudos for doing this that long.

I just upgraded from Joshua's 5.10 based 22.04 to 24.8.4 (Noble with 6.10) and was hoping that my frigate would now run better... His suggestion was 24.04. but as he stopped his support (Which I totally get after reading the post @Werner linked) what would be the best foundation if I need the NPU?

As you have mentioned in chat you tried this with mainline 6.10.y which I don't expect to even work. Support for rk3588 soc is still under heavy development and I don't think anything NPU related has been pushed upstream yet.

 

You should try with 6.1.y vendor kernel. Why? https://docs.armbian.com/User-Guide_FAQ/#why-does-hardware-feature-xy-work-in-old-kernel-but-not-in-more-recent-one

2 hours ago, Werner said:

He starts to feel the pain we feel for years already. It is hard to not tell people simply to fuck off. When I first came across his project I thought he must be crazy to do this all by himself. Kudos for doing this that long.

That is sad to see,.. were there specific reasons that Joshua did not contribute to Armbian instead? Has the Armbian team had better luck with Rockchip? The issues he mentiosn with SDK sound quite serious.

IIRC he prefered to stay independent which I understood. Though that was a year ago where his motivation was still high.

 

Now things have changed. We told him he'd be more than welcome to join forces and share the burden.

Not sure what was his experience but Radxa isn't all to famous for their support:

https://www.jeffgeerling.com/blog/2023/rock-5-b-not-raspberry-pi-killer-yet

TLDR: So right now, I wouldn't recommend this board to just anyone, at least when it comes to software and support, due to its high price.

Edited October 28, 20241 yr by mircsicz

@mircsicz frigate with hw acceleration and npu object detection works nicely on joshua’s ubuntu 24.04 for me. i have inference speeds between  25 to 40 ms. i believe it would be same on armbian.

17 часов назад, janmat сказал:

frigate with hw acceleration and npu object detection works nicely on joshua’s ubuntu 24.04 for me. i have inference speeds between  25 to 40 ms.

Can you explain what these numbers mean? And what alternative can they be compared with?

It seems to me that this is probably a wrong translation.

it corresponds to the calculation time in milliseconds. The accuracy of the neural network usually correlates with the inference time, i.e. the more time is allowed for a decision, the more accurate the neural network can operate.

@going consider https://codingfield.com/blog/2024-01/install-frigate-in-proxmox-on-orangepi5plus/ for context - it is about how fast you can detect objects in order to be able to respond as fast as the video is changing. Depends on the model and key parameters.

 

If you wanted to process video at 30 frames per second, you'll need to make adequate detection sufficiently quickly: 1000 / 30 = 33.3ms allowing for zero other overheads.

 

These days you might have multiple cameras, so object detection speed can be pretty important.

Edited November 1, 20241 yr by ozacas

29.10.2024 в 21:18, janmat сказал:

it corresponds to the calculation time in milliseconds. The accuracy of the neural network usually correlates with the inference time, i.e. the more time is allowed for a decision, the more accurate the neural network can operate.

 

21 минуту назад, ozacas сказал:

If you wanted to process video at 30 frames per second, you'll need to make adequate detection sufficiently quickly: 1000 / 30 = 33.3ms allowing for zero other overheads.

 

These days you might have multiple cameras, so object detection speed can be pretty important.

 

Thank you both so much for these kind answers.
I've been wanting to learn how to use NPU for a long time.
But unfortunately I haven't figured out where I can start yet.

 

As far as I know from the documentation on the rk3588 SOC, there is a controller in it that is designed for I/O with NPU.
I still don't understand how I can program it and how to interact with the Linux kernel.

Can you tell me the direction to study?

 

@going a pretty broad question, but i'll try 😁

 

hardware accelerated video processing on rk3588 and linux - https://github.com/blakeblackshear/frigate/discussions/11502

hardware accelerated NPU via linux - https://github.com/airockchip/rknn-toolkit2/

large language models (LLM's) using NPU - https://github.com/airockchip/rknn-llm

 

Its important to understand you will need a compatible kernel driver to connect to the hardware - most likely the best option is the vendor kernel 6.1.75 with armbian. Also this is a rapidly evolving space - hard to keep up!

 

Depending on what you read though, often the CPU can offer better performance, although it is dependent on your requirements and the ability to utilise the hardware effectively.

2 часа назад, ozacas сказал:

pretty broad question, but i'll try 😁

 

hardware accelerated video processing on rk3588 and linux - https://github.com/blakeblackshear/frigate/discussions/11502

hardware accelerated NPU via linux - https://github.com/airockchip/rknn-toolkit2/

large language models (LLM's) using NPU - https://github.com/airockchip/rknn-llm

 

Thank you very much!
Now we can assume that I will disappear from the horizon for several months of study.🤗