Ovaday

The one from MR / my previous message

just as a statistics/stability report: the DTB with the NPU works very well. Uptime of the device is already a week, it's running armbian and immich on it. While immich lags sometimes because the board can be weak for video processing, but in general it already processed ~20k images with OCR and face detection based on RKNN. I consider it pretty good results.

Created a Pull Request: https://github.com/armbian/linux-rockchip/pull/500 Final DTB: rk3566-box-X88PRO20-npu.dtb @Hqnicolas do you know how I can add the board as selectable from the Armbian website? Do I have to also add DTS that will work for latest kernel?

Patched the .dtb for v6.1, currently not yet in final format. I will probably create a commit to Armbian with a .dts for our board. I am attaching the dtb/dts just to keep track of the progress. Here are results of the RKNN test launch: rknn_api/rknnrt version: 2.3.2 (429f97ae6b@2025-04-09T09:09:27), driver version: 0.9.8 model input num: 1, output num: 1 input tensors: index=0, name=input, n_dims=4, dims=[1, 224, 224, 3], n_elems=150528, size=150528, fmt=NHWC, type=INT8, qnt_type=AFFINE, zp=0, scale=0.007812 output tensors: index=0, name=MobilenetV1/Predictions/Reshape_1, n_dims=2, dims=[1, 1001, 0, 0], n_elems=1001, size=2002, fmt=UNDEFINED, type=FP16, qnt_type=AFFINE, zp=0, scale=1.000000 custom string: Begin perf ... 0: Elapse Time = 8.06ms, FPS = 124.12 ---- Top5 ---- 0.935059 - 156 0.057037 - 155 0.003881 - 205 0.003119 - 284 0.000172 - 285 So: CPU: OK GPU: OK via Panfrost NPU: OK, RKNN inference works on RKNPU rk3566-box-X88PRO20-npu.dtb rk3566-box-X88PRO20-npu.dts

I got a small progress here with a RKNN (NPU). 1. note from previous posts: rk3566-box-demo.dtb looks like does not support our USB. Use rk3566-box-X88PRO20.dtb that other shared. I tried to use the NPU, looked across the forum. Managed to get our .dtb running with the correct settings - however the Kernel didn't had the RKNN packages + it could not link the 'rocket' to our NPU. While the NPU was recognized after my tweaks. Read that it runs only on kernel v. 6.1 from VENDOR - recompiled the kernel for our board, installed it over the existing installation - .dtb didn't run... When I switched to box-demo from the kernel the NPU and RKNN were recognized instantly. I will not attach the kernel itself for now, as it is ~300 Mb, however you can compile it from Armbian: ./compile.sh kernel BOARD=station-m2 BRANCH=vendor \ KERNEL_BTF=no KERNEL_GIT=full it have to be installed like that: dpkg -i linux-image-vendor-rk35xx_*.deb linux-dtb-vendor-rk35xx_*.deb update-initramfs -c -k 6.1.115-vendor-rk35xx Probably will be a task for today to tweak .dtb for this kernel version and get board running with RKNN. I want to try immich with a local image classification.

I've created a guide here how to run a Home Assistant OS suprisingly it run mostly natively. Didn't do anything except replacing the DTB. Also right away flashed it into a emmc. On my board emmc is from Samsung and has a read speed on 180Mb/s...

I extracted DTB/DTS from the Android FW (files attached). For whatever reason they are signed in the image as rk3568. No ideas why. 02_dtbdump_rockchip,rk3568-evb.dtb 02_dtbdump_rockchip,rk3568-evb.dtb.dts Attention: They are for Android! Linux might not boot with them!

I know it was already written here in some of the messages that these devices are not secure. But now I got a message from Telekom that they detected a Chineese spyware connection: The timeframe is exactly when I have switched the device on with an Android on board.

Got the same board running Armbian with ChatGPT. Because we performed a lot of experiments, here are summary of how we actually got it running tl.tr: upgrade_tool is wonderful, but running it on my board destroyed everything and I had to recover from backups. We backuped board completely, then intentionally corrupted android bootloader from adb and managed to get U-Boot running via modification of boot.scr Here is a chat-gpt generated summary: