Metis M.2

Hello! I’ve been trying to connect and use the Axelera Metis M.2 card with no success. The same M.2 on a different PC was recognized correctly by using:lspci | grep AxeleraBut now the card cannot be recognized at all on the new system. Here are some specs of the current host:Dell 5820 CPU: Intel(R) Core(TM) i7-7820X CPU @ 3.60GHz RAM: 32 GB GPU: 2 Quadro RTX 5000 OS: Ubuntu 24.04 I’ve also tried lspci | grep 1f9d with no success. The Metis driver is already installed and loaded with modprobe.I'd appreciate any help or insight on additional steps I could try.Additionally, this is the adapter I’m using for the Axelera card: 

Hi, We are considering adding an AI Accelerator for our board (i.MX8, yocto).Do you support yocto OS?Is there any reference design or example that runs Metis on yocto platform? Thanks,Afik

Have you seen our fruit demo? If not, ​@David explains it amazingly here.Basically, it is running 3 different yolov8 models on a single chip to process 4k camera input to segment fruit, while only using 640x640 models!I love this demo and use it to evaluate any system that comes across my desk! So, obviously, when ​@Victor Labian posted the setup for the Orange Pi here, I just had to try it out! And it works perfectly straight after following the instructions!I shaded the fruit to look different from the original to show the accuracy of the segmentation.If you want to run this demo for yourself, you can! It’s available in the examples folder of your SDK, called fruit_demo.py. Let me know what you think! Or, as ​@Spanner suggestedA Fruitful and A-peel-ing Orange Pi Experiment that Didn't Go Pear-ShapedHere's what's current on the Axelera grapevine - a core functionality of the Orange Pi with Metis is detecting fruit in real time! Super smoothie AI pipe-lime with bananas and apples being

Hi! I wanted to upgrade my M.2 card with latest firmware 1.3.0 but when trying to invoke command from instruction https://github.com/axelera-ai-hub/voyager-sdk/blob/release/v1.3/docs/tutorials/firmware_flash_update.md#flash-the-board-controller-firmware, I got an error:Do you know how this can be resolved? I’m using Intel i5 processor.

This may not be possible because of how the software and hardware work together. I’m new here, and not traditionally a SW engineer, but I have a question/ suggestion. I noticed when I attempted to run the example local LLM inference scripts in the SDK that I can’t run them with only 1GB of DDR memory on the card. I think the example needs to be edited so that the model can be run on 1 GB of DDR (a model like tinyllama or similar that’ll actually fit in the 1GB of space.)After thinking about it, I do have an intel Raptor Lake integrated GPU, and an external NVIDIA A500 from Adlink Technologies. Is there any way to use the memory in the integrated CPU or the thunderbolt connected A500 GPU to run the LLM inference demos by using the DDR memory in the other devices? That would be very very useful!

i am using Metis M.2 Eval System with Aetina RK3588 Industrial Motherboard . i connected everything to it keyboard, monitor and mouse and power supply . the monitor is showing blank screen . but yesterday it is working . why it is not working now . is there any solution. 

Hello everyone,I’m new to working with this hardware and recently installed the Metis M.2 chip (Firmware version: v1.2.0-rc2+bl1-stage0) on an ASRock Z170 Pro4S motherboard. The device was detected successfully.I’ve installed the Voyager SDK and attempted to run the following command from the Quick Start Guide:./inference.py yolov5s-v7-coco dataset --no-display However, I encountered an input/output error when executing the command. I’ve included the full error message below for reference.I would appreciate any guidance or suggestions you may have to help resolve this issue.Thank you in advance for your support. INFO    : Using default val dataset INFO    : Using device metis-0:5:0 INFO    : Network type: NetworkType.SINGLE_MODEL INFO    : Input INFO    :   └─detections INFO    : Imported DataAdapter ObjDataAdaptor from /home/tud/voyager-sdk/ax_datasets/objdataadapter.py Stream Playing:  71%|██████████████████████▏        | 5/7 [00:00<00:00,  6.26/s][AxeleraDmaBuf.cpp:234] UIO_IOCTL

Hello!I am facing some problems tring to run a network inference on the Metis M.2 with Raspberry Pi via Docker. To try the inference I have followed the steps in the guide (https://support.axelera.ai/hc/en-us/articles/26362016484114-Bring-up-Voyager-SDK-in-Raspberry-Pi-5) but after trying to run the inference.py, I face the following errors:  (venv) root@iak:/home/voyager-sdk# ./inference.py yolov8s-coco-onnx ./media/traffic1_480p.mp4INFO    : Using device metis-1:1:0WARNING : Failed to get OpenCL platforms : clGetPlatformIDs failed: PLATFORM_NOT_FOUND_KHRWARNING : Please check the documentation for installation instructionsINFO    : Default OpenGL backend gl,3,3 overridden, using gles,3,1INFO    : Network type: NetworkType.SINGLE_MODELINFO    : InputINFO    :   └─detectionsStream Paused:  57%|██████████████████████████████████████████████████████▊                                         | 4/7 [00:00<00:00,  5.56/s][ERROR][axeWaitForCommandList]: Uio wait kernel failed with return c

Hello,I am currently working with a lane detection model named culane_res18.onnx, and I would like to know how I can run this model on the Metis M2 device.Here are some details about the setup:The model is culane_res18.onnx, which is a lane detection model. The training dataset is from the CULane dataset, and I have the link to the dataset here. The model file can be found here.Could you kindly guide me on how to run this lane detection model on Metis M2? I would appreciate any step-by-step instructions or basic guidance on how to deploy this model, considering that I have both the ONNX model and the dataset.Looking forward to your response!Best regards,Chinmay Mulgund

Hello!!Do you have any news about support for Ubuntu 24.04? I wanted to use axelera on raspberry pi cm5 however there is no support for 22.04 and I wanted to run it on the host, not in a docker container.Thank you 😊

Hello,I’m having trouble getting my AXELERA M2 to be recognized under Linux. Until recently everything was working fine; now when I try to run an inference the SDK can’t find the board. I’ve already tried rebooting, but the issue persists.aravind@aravind-H610M-H-V2:~/Desktop/voyager-sdk$ source venv/bin/activate(venv) aravind@aravind-H610M-H-V2:~/Desktop/voyager-sdk$ ./inference.py yolov5s-v7-coco chinmay/Vehicle-CV-ADAS/assets/challenge_video.mp4 --save-output ~/Desktop/voyager-sdk/chinmay/saved_videos/yolov5s-v7-coco.mp4ERROR : No target device found in lspci outputERROR : AXR_ERROR_CONNECTION_ERROR: No target device found in lspci output 

​@Victor Labian I’m also trying to setup rpi5 with metis m2. I have question what host OS you was using and have you changed anything in config.txt?

Hi, I am also working on same thing deploying the YOLOPv2 model on to Metis M2. I have downloadded the data set from bdd100k and arranged in this manner.voyager-sdk/├── data/│   └── yolopv2_dataset/│       ├── images/│       ├── labels/│       ├── cal.txt│       ├── val.txt│       └── data.yaml├── customers/│   └── my_yolopv2/│       └── yolopv2.pt└── yolopv2-custom.yamlso when i am trying to deploy the model using yaml I am getting this error(venv) aravind@aravind-H610M-H-V2:~/Desktop/voyager-sdk$ ./deploy.py customers/my_yolopv2/yolopv2.yamlINFO    : Using device metis-0:1:0INFO    : Detected Metis type as pcieINFO    : Compiling network yolopv2-custom /home/aravind/Desktop/voyager-sdk/customers/my_yolopv2/yolopv2.yamlINFO    : Compile model: yolopv2-customINFO    : Imported DataAdapter ObjDataAdaptor from /home/aravind/Desktop/voyager-sdk/ax_datasets/objdataadapter.py/home/aravind/.cache/axelera/venvs/93f45ae3/lib/python3.10/site-packages/torch/serialization.py:779: UserWarning: 'to

hello I hijack this a little bit … once I type “lspci” it gave me following output on a pi 5 0001:00:00.0 PCI bridge: Broadcom Inc. and subsidiaries BCM2712 PCIe Bridge (rev 30)0001:01:00.0 Processing accelerators: Axelera AI Metis AIPU (rev 02)0002:00:00.0 PCI bridge: Broadcom Inc. and subsidiaries BCM2712 PCIe Bridge (rev 30)0002:01:00.0 Ethernet controller: Raspberry Pi Ltd RP1 PCIe 2.0 South Bridge how to activate the device and getting to work with my python3 script  

Hello my name is Kevin, i am trying to work on a project that the m.2 is going to interfere with my cpu and later on maybe my gpu.is there a way, and another question whats the command to go on rtsp?  my system is:AMD 9950X3D CPUAMD 7900XTX GPU48 gb DDR5 on 8200MHZMetis Axelera AI on an m.2 port. 

This is a very interesting thread as I’m testing and M.2 card on an AMD system as well and am having what seems to be the exact same problems. Things added / tried: kernel command line includes md_iommu=off pcie_aspm=off update-pciids → resulted in card having the correct name in lspci : Axelera AI Metis AIPU (rev 02) axdevice --refresh -v results in: INFO:axelera.runtime.axdevice:Removing 0000:20:04.0INFO:axelera.runtime.axdevice:PCIE rescan0000:25:00.0 : Axelera AI Metis AIPU (rev 02)INFO:axelera.runtime:Found PCI device: 25:00.0 Processing accelerators: Axelera AI Metis AIPU (rev 02)INFO:axelera.runtime:Found AIPU driver: metis                  90112  0WARNING:axelera.runtime:4PCI device count mismatch: lspci=1, triton=0(removed the traceback) just tried rerunning install.sh which mentions at the end: building operators refreshing pcie and firmware WARNING: Failed to refresh pcie and firmware  tried removing metis-dkms / rmmod metis and reinstalling results in same warning as ab

Hello,I know that axelera does not work in the attention mechanism. For example, can we convert the attention mechanism in the yolov11 architecture to convolution? Have you done any research or read on this subject? Does Softmax structure work in Axelera? (as far as I know, no) Is there a way to turn Softmax into a working structure?

Hello, is there a detailed document where I can learn about the Axelera chip architecture and how to do the quantization in it?Thanks

Hello,I wanted to test the Yolo11n model. But I couldn't do this because I got an error in the calibrating part. Is this because of the use of attention mechanism in Yolo11 and Yolo12?Error: ValueError: Some modules couldn’t be simplified: <… activation.Softmax> <…. matmul.MatMul>In addition, when I looked at the ready yaml files on voyager sdk github, I saw that the yolov8x model was not available. Is there a special reason for this? Because when I prepared and tried a new yolov8x yaml file, I got the similar error again. But as far as I know, the attention mechanism is not used in Yolov8x.Thanks

I’m scoping out the Metis M.2 module for a compact edge AI project and wondering how capable it really is. I’m mainly interested in real-time object detection on 1080p video streams, maybe 2–4 cameras max.Can it handle multi-stream inputs in real-time, or is that better suited to the PCIe version?

It’s actually not that complicated getting started with an M.2 module, and there’s a good chance that the Metis isn’t the first M.2 you’ve ever had in your hands.But it’s still good policy to familiarise yourself with both the hardware and the software when you’re ready to unbox that little AI beauty.To help with that, we’ve got a series of articles to walk you through…Installation: Handling and installing the M.2 in your host system, and a great tip on how to confirm there’s suitable power getting to it once it’s in place Integration: Top tips on figuring out whether the host PC (or system or platform) is nicely compatible with the needs of the Metis accelerator. Voyager SDK: Once it’s installed and ready, all you need is the Voyager SDK.This is all just to get you started, and then the real work begins! Make sure you share what you’re building right here on the Axelera community 👍