Recent Topics

Hi,I just finished setting up everything and now I’m trying to run an example pipeline using.\inference.py yolov7-640x480-coco cars_clip.mp4Building the model seems to work fine but when the inference starts, it either crashes after some images withDetecting... : 2%|█▋ | 37/1802 [00:01<00:39, 44.29frames/s][ERROR][axeCommandQueueExecuteCommandListsAsync]: Wait kernel failed with return code -1.terminate called after throwing an instance of 'std::runtime_error' what(): axr_run_model failed with Error at zeCommandQueueExecuteCommandLists(cmdqueue, n_cmdlists, cmdlists, nullptr): cmdqueue_run_cmdlists: 319: Exit with error code: 0x7FFFFFFE : ZE_RESULT_ERROR_UNKNOWNAborted (core dumped)or instantly withINFO : Could not exec vainfo: Command '['vainfo']' returned non-zero exit status 3.WARNING : Timeout for querying an inferenceERROR : InferenceStream terminated due to T

Hi community! It's been a while since my last update - lots happened! The short version: MotionFlow now runs fully on the Orange Pi with the Metis accelerator, and I've been iterating on making the whole thing actually usable in daily life. Moving to MetisI ripped out the OpenVINO inference path entirely and committed to the Axelera Voyager SDK as the only backend. I initially fell into the trap of pulling frames through OpenCV, which gave nice per-frame control, but a real performance bottleneck. Switching to the SDK's streaming API and letting its GStreamer pipeline handle everything made a huge difference! It manages video decoding and pipelined inference across the 4 Metis AIPU cores.For this, I stitched together a simple custom pipeline combining yolov11l-pose and ByteTrack for tracking. Multi-stream support was actually easier than expected - the SDK's create_inference_stream accepts multiple RTSP sources and tags each frame with a stream_id, so I can process several cameras in o

I try to use Metis M.2 on Jetson Nano, I followed all the steps to install the board but when I try to run the first example from Voyager-SDK I get the following error:(venv) adrianza@ubuntu:~/voyager-sdk$ ./inference.py yolov5s-v7-coco media/traffic1_1080p.mp42026-02-05 09:54:10.578538286 [W:onnxruntime:Default, device_discovery.cc:164 DiscoverDevicesForPlatform] GPU device discovery failed: device_discovery.cc:89 ReadFileContents Failed to open file: "/sys/class/drm/card1/device/vendor"WARNING : Failed to get OpenCL platforms : clGetPlatformIDs failed: PLATFORM_NOT_FOUND_KHRWARNING : Please check the documentation for installation instructionsINFO : Deploying model yolov5s-v7-coco for 4 cores. This may take a while...|████████████████████████████████████████| 7:53.1 Opening in BLOCKING MODE NvMMLiteOpen : Block : BlockType = 261 NvMMLiteBlockCreate : Block : BlockType = 261 WARNING : capsfilter0: BUS: not negotiatedWARNING : (../libs/gst/base/gstbasetransform.c(1431): gst_base_tr

Completed the initial architecture implementation. The device can now track standard therapy exercises and show the count. Next step is to detect the pattern from an unknown exercise and then use it to help practice. I was also thinking of adding audio support - any recommendations on TTS models that can run with low latency?

We have continued advancing the development of MedBot both in terms of design and functional integration, with a strong focus on improving its medical-oriented appearance and strengthening the connection between hardware, sensors, and artificial intelligence.One of the first visible changes is the redesign of the MedBot case. The enclosure has been updated with a more medical look, including a white border and cleaner lines, reinforcing its identity as a healthcare-oriented device.On the hardware side, we have connected the display provided by Axelera to the electronic mainboard and completed an initial software implementation that interfaces with our NEMA artificial intelligence through web-based communication. This enables MedBot to connect online to NEMA and start real-time interaction with its AI services.We are currently testing AI functionalities that will support assistance and interaction within MedBot, validating communication flows and response behavior.Starting from this wee

Taming the GStreamer Beast: From 3 FPS to Real-Time1️⃣ The BottleneckWe hit a wall.Our Python-based torch-aipu pipeline was smart, but slow. It was handling video decoding, preprocessing, and visualization all on the host CPU.The result? 3.4 FPS on the host. Frames were piling up. Latency was unacceptable.To get real speed, we had to go lower level. We had to move the entire pipeline to GStreamer. 2️⃣ Building Custom OperatorsStandard GStreamer elements weren't enough. Optical flow is unique. It needs to see "time."We built a custom C++ Preprocess Operator. It acts as a short-term memory. It caches the previous frame. It stitches the "past" and "present" into a 6-channel tensor (NHWC) for the NPU.We also built a Postprocess Operator. The model outputs raw numbers. This operator converts that math into a colorful RGB flow map and pushes it downstream as a FlowImage. 3️⃣ Fighting the Crashes (The Real Work)Migration was not smooth. We fought issues across the entire stack. Here is the ba

Hi Axelera inferences and compiler available https://hub.degirum.com

Hello,I am trying to open ssh server in my Metis SBC board (with 1x AIPU, 4 GB of RAM, 16GB of host RAM and active cooling, Rev1), but right now I’m in the sh $ console with the “antelao” user, and whenever I need to run a sudo command, I’m asked to enter the sudo password, which I don’t know. Can you help me with this or point me to the right tutorial/guide?Thanks a lot

Hi,I have 3 axelera metis m2 device. I could not enable update for all of them. I have Voyager SDK 1.5.2 version that installed successfully. When i enter the command “./enable_bootloader_update.sh" it gives an error such as “Device communication timed out” and “Failed to load firmware”. I disabled the IOMMU option in BIOS and disabled the secure boot. My device is seen as “03:00.0 Processing accelerators: Axelera AI Metis AIPU (rev 02)" in the lspci output. Also when i enter the “axdevice” command it says : "[libtriton_linux.c:1313] Device communication timed out: device did not respond within 1 seconds. (0)WARNING: Failed to get valid board type for device metis-0:3:0 got 8Device 0: metis-0:3:0 board_type=unknown (not responding)"How can i solve this problem ? Can you help me ? The full terminal output of the command is that : (venv) ysf@ysf:~/Desktop/voyager-sdk$ ./enable_bootloader_update.shFirmware Flasher - Extracting files...Extracting firmware files...Files extracted successful

Hi Community!I want to share the architectural progress I've made with the Voyager SDK since my last post about Project Gaius.My journey began with a plan to utilize a pre-trained YOLOv11, which showed great promise for hand keypoint detection. However, after encountering persistent matmul errors during deployment to the AIPU, I pivoted to YOLOv8n-pose. I trained a custom model on the Ultralytics Hand Keypoints Dataset, which successfully compiled and is now running smoothly on the Metis.To address the resolution challenge where downscaling camera streams makes hands too pixelated for accurate tracking, I am implementing a "Smart Cropping" 2-stage inference pipeline. Two networks work in tandem: the first detects the global context and roughly locates the hand, while the second extracts precise 21-point keypoints. Instead of resizing the whole image to the standard 640x640 input, the logic uses the first stage to crop the hand from the high-quality stream before feeding it into the sec

Huge updates today for ABC Project. We are starting the development of the first part of object recognition based on Metis of the entire process pipeline, in order to make the OCR simpler and more accurate (i.e. less capable of delivering wrong results). All these operations are both to be done in real world and on software, as we need to start with a proper and repeatable hardware setup prior to properly set software up. The first operation in the real world is to bound a specific area with masking tape, which we call the loading area, where the book to be processed will be placed; this area is just below the camera. After this, we draw an X on each corner of the loading area's rectangle, with a black marker on the masking tape and a calibration procedure based on OpenCV library detects the rectangle to crop later the acquired images.  The next action, in order to make the setup repeatable, is to measure the height of the camera from the tabletop where the loading area is located; thi

Hi again, long time to see ^^After 6 months I saw a post about ORIN support and I wanted to try my Metis M.2 again.First I wanted to try it on my Desktop computer, ORIN will come after. But unfortunately, I got some errors.Note: I was able to run Metis M.2 on my desktop before.Now I am using same operating system (Ubuntu 22.04) but with different kernel version I suppose.I can install the repo succesfully but when I try to run “axdevice” command it gives me an error like this:[libdmabuf.c:1712] Device communication timed out: device did not respond within 1 seconds. (0)WARNING:axelera.runtime:Failed to get valid board type for device metis-0:2:0 got 5Device 0: metis-0:2:0 board_type=unknown (not responding)What I try to solve problem but didnt help:Disabled secure boot from BIOS Disabled Intel_iommu from grub entry Disabled pcie_aspm from grub entry Installed latest version of voyager (1.5.2) Installed previous version of voyager which I was able to run on April 2025 (1.2.5)Can you hel

I am looking for a way to infer Qwen VL 32b instruct A3b on the edge using Axelera m 2 form factor hardware. But I would need a lot of Ram to make this happen I am talking at least 48Gb. Could anyone share some light on what are my options. I need to do video inferencing and scene understanding on request from the user. My idea was to grab a few frames whenever the request was initiated and then choose the best frame to do inference. 

Hello everyone! This is our first thread in the community, and we’re excited to share what we’re building with the challenge kit! Project titleA.B.C. ( A.I. Book Cataloguer ) What you’ll be buildingWe’re developing a next-generation tabletop station that automatically scans a book’s cover, uses an on-device AI model to extract the title, author, and publisher, and instantly displays the results on screen. The operator simply places the book on the surface beneath a camera connected to the Orange Pi running Metis and reviews the output-fast, intuitive, and hands-free.The system combines an object-recognition model, used to distinguish the operator’s hands from the book itself, with an OCR model that extracts the text of interest directly from the cover, enabling accurate, real-time identification of title, author, and publisher.If the information is correct, the book is cleared; if not, the operator simply crosses their index fingers to form an X under the camera. This gesture is detect

For anyone building custom embedded Linux images with Yocto, there's now an official BSP layer available for Metis hardware.It’s pretty niche stuff, but if you're working on embedded systems where you need a stripped-down, customised Linux image (think industrial deployments, kiosks, dedicated edge boxes), this is for you.What's in it:The layer includes recipes for the Metis PCIe driver and the udev rules. Add it to your Yocto build, include the packages in your image, and your custom Linux will recognise and talk to Metis hardware out of the box.Currently supports Yocto 5.0 (Scarthgap), with branches for Kirkstone and Dunfell if you're on older versions.One important thing to note:The VoyagerSDK itself still doesn't run natively on Yocto yet. You'll need to run it inside a Docker container on your target system. The kernel driver runs on the host and handles the hardware communication, but inference and SDK tools happen inside the container. Just FYI 👍.Quick look at the setup:Clone t

Hello everyone,Abdi here. I am a part-time working student at Axelera. I joined in December.I am having a blast playing with the Metis devices. Our task is focused on testing the Voyager SDK documentation. Is it clear enough for the customer? Where can it be improved? We test the documentation by actually following it, such as working through the tutorials provided in the documentation.Furthermore, we test how easy it is to perform tasks that users may actually want to do, such as counting how many objects are in a video frame using YOLO. Deploying models,running different kinds of models.We evaluate how useful the documentation is in this context and identify where the gaps are.Currently, I am working on integrating an external tool, Whisper, with the Metis Dev System on the Arduino Portenta X8, and  routing Whisper’s output into an LLM running on the device. -Abdi   

EdgeFlowNet Optical Flow. Successfully Deployed On Board1️⃣ EdgeFlowNet Is Now Running on the BoardWe’ve successfully compiled and deployed EdgeFlowNet,a modern encoder–decoder optical flow model,directly on the Orange Pi 5 Plus + Metis M.2 platform.2️⃣ Calibration for Optical Flow Is Not “Plug-and-Play”Unlike single-image models, optical flow operates on consecutive frame pairs.To make this work inside the Voyager SDK calibration flow,we implemented a custom DataAdapter dedicated to optical flow:Takes two consecutive RGB frames Concatenates them into a 6-channel tensor Ensures layout, normalization, and ordering match model expectationsThis adapter is used only for calibration,but it is critical for correct quantization and deployment.3️⃣ A Mysterious Shape Mismatch — and How We Found ItDuring deployment, we encountered a confusing Relay type error:Error: The Relaytype checker is unable to show the following types match:Tensor[(192,260,4,64),int8]Tensor[(192,256,1,64),int8]This dimens

If you've wanted to peek under the hood of how your Metis hardware talks to Linux, there’s a new Axelera AI repo with the PCIe driver source code.Repo: https://github.com/axelera-ai-hub/axelera-driverFor most of us running the standard Voyager SDK setup, you won't likely need to touch this. The driver gets installed automatically. But if you're doing something a bit more adventurous (custom kernel builds, embedded systems, that sort of thing) this could be really useful. I know at least a few people on here have been experimenting with Debian and Yocto, for instance.Quick overview of what's there:It's the metis.ko kernel module. The one that handles all the PCIe communication between your system and the Metis hardware. Currently it supports Ubuntu 22.04 and 24.04 (LTS).Building it yourself is pretty straightforward if you should want to. Make sure you've got your kernel headers installed (sudo apt-get install linux-headers-$(uname -r)), then it's just the usual make and sudo make insta

I’m interested in hearing from everyone here about which models you’re successfully using with your Metis device and Voyager SDK, and which ones you’d like to see being officially supported.This could be really useful info in terms of prioritisation when it comes to adding more support in the model zoo.

On an AMD CPU with ubuntu 24.04, I see these warnings. Inference does open, but it seems to render poorly.(venv) my@machine:~/voyager-sdk$ ./inference.py yolov8l-coco-onnx media/traffic1_1080p.mp4 media/traffic1_1080p.mp4 media/traffic1_1080p.mp4 media/traffic1_1080p.mp4INFO    : Could not exec vainfo: Command '['vainfo']' returned non-zero exit status 3.WARNING : Failed to get OpenCL platforms : clGetPlatformIDs failed: PLATFORM_NOT_FOUND_KHRWARNING : Please check the documentation for installation instructionsOpenCL doesn't find any platforms(venv) my@machine:~/voyager-sdk$ clinfoNumber of platforms                               0ICD loader properties  ICD loader Name                                 OpenCL ICD Loader  ICD loader Vendor                               OCL Icd free software  ICD loader Version                              2.3.2  ICD loader Profile                              OpenCL 3.0

 Hi everyone!We’re Zijie Ning ​@mm0806son   and Enmin Lin ​@Enmin  from KU Leuven (Belgium).We’re thrilled to share that our project FlowSentry-Wake has been selected as one of the top 10 projects in the Axelera AI Smarter Spaces Project Challenge! 🎉Thank you to the reviewers for the recognition, and to the community for creating a space where edge-AI ideas can truly come to life.FlowSentry-Wake is an on-device intelligent space-guarding system. Our goal is to explore how a space can be understood locally—without cloud dependency—while keeping the system low-power, reliable, and resistant to simple adversarial tricks. 🔧 What Problem Are We Addressing?Many conventional security systems struggle with a few persistent issues: High power usage Running heavy AI models 24/7 is expensive, hot, and often unrealistic for long-term deployment. By improving energy efficiency in our embedded system, we can significantly reduce overall power consumption, which directly lowers the system’s carbon