Recent Topics

Hello.I recently received the m.2 metis accelerators and did a test on the boards provided by Axelera and got results close to the published benchmarks. It was succesful.I then connected these accelerators to the Orin (via the m.2 slot) in order to use them also for the my Jetson AGX Orin environment. But in the voyager-sdk installation i got “WARNING: Failed to refresh pcie and firmware” error. The installation completes but the accelerator device does not show up on my system. In lspci -tv command I get the output attached to this post. Can’t see metis there.I didn't see any information about the compatibility of these accelerators with Jetson systems (my m2 slot is M-key). If they are compatible, can you help me to fix the installation?Thanks :)

 I am considering installing a local AI system in my company.I kindly ask you to provide me with useful documentation to do this.Please consider a user who only has experience with C++ and Linux servers.Below is a brief project description.Project objectivesCreate an autonomous AI platform installed at your premises.Manage and monitor systems installed by customers (lights, sensors, cycles, errors, maintenance).Integrate Antholux3 and Topiaria2 into a single management architecture.Eliminate cloud dependencies → everything local: secure, scalable, proprietary.Create a technology hub for support, diagnostics, updates and remote assistance.General system architecture2.1 Main componentsLocal AI server (mini-PC, NUC or workstation)Internal databaseCommunication brokerRules engine + local AIUnified dashboardOTA module for updating remote systemsLogging and diagnostics system2.2 Technologies usedServer: Linux Debian/UbuntuDatabase: PostgreSQLBroker: MQTT (Mosquitto)UI: React or Angular + RES

Hi guys,After a small break, I was able to get the Metis M2 running on my Jetson Orin NX.However, is it possible that when using ‘axdownload’, the models are not precompiled? I tried to download ‘mobilenetv2-imagenet’, and got this error when trying to load it in. So, my guess is, I incorrectly assumed that the downloaded folder included a precompiled one? Or is there maybe a list of models which do include a precompiled model?Cheers (more info below),SebbeError-----------------------ERROR][zeModuleCreate]: Failed with error code: 0x70000004[ERROR][zeModuleCreate]: Build log:sh: 1: riscv64-unknown-elf-gcc: not foundCommand failed (32512)Failed to build module: sh: 1: riscv64-unknown-elf-gcc: not foundCommand failed (32512)[ERROR][zeKernelCreate]: Failed with error code: 2013265925Error at zeKernelCreate(module_handle, &desc, &kernel_handle): kernel_create: 409: Exit with error code: 0x78000005 : ZE_RESULT_ERROR_INVALID_NULL_HANDLEFailed to create handlesFailed to initialise AIC

Hello everyone,I’ve been facing problems with Ubuntu 22 where the BARs for my Metis M.2 board are not being assigned correctly. The kernel modules are installed and the device is recognized, but the BAR mapping seems inconsistent, which causes issues when running the Voyager SDK natively on Ubuntu 22.To work around this, I’m considering running the Voyager SDK inside a Docker container based on Ubuntu 22 while my host system remains on Ubuntu 24 (where the kernel modules work fine and the board is detected properly).Unfortunately, I cannot access the Axelera AI Help Center (https://support.axelera.ai/hc/en-us/articles/25953148201362-Install-Voyager-SDK-in-a-Docker-Container). Is there any alternative resource or guide that explains how to install and configure the Voyager SDK inside a Docker container?   

Red Dwarf’s really well known for how well it predicted (albeit comedicaly) VR tech, but it did a lot more than that. It heralded a lot of advance in AI that we’re now actually seeing emerge: conversational appliances, narcissistic holograms, forgetful mainframes, and borderline neurotic robot butlers not being the least of them.25th November = Gazpacho Soup Day. Look it up.The show was set in the 22nd Century, but it’s only 2025 and it’s already coming true. With edge AI powering robotics, smart devices, and real-time vision on the edge, so many of those hilarious sci-fi predictions are fast becoming reality. From Rimmer's hologram (an early take on digital consciousness) to Kryten's helpful, protocol-bound persona (think domestic service robotics with a strong opinion on laundry), Red Dwarf imagined AI not just as powerful, but as delightfully flawed. That might actually be the brilliance of what it predicted over some of its more popularised contemporaries like Star Trek or even The

Big update from the Metis world: the epic new 4-core PCIe card is now available to order over on the store.Huge performance shift that opens up seriously scalable edge AI.📊 Key specs: 856 TOPS (INT8) Choice of 16GB and 64GB DRAM ~12,800 FPS on ResNet-50 Runs dozens of streams or parallel models per board 30–58W typical power draw It’s not just the raw speed, it’s the ability to compress multi-camera inference, object detection, tracking, segmentation, and more into a single card that I find especially exciting. You can prototype in days, using the Voyager SDK’s YAML-based pipeline builder and model zoo (everything from YOLOv5 to small LLMs).The benchmarks are where it gets interesting. YOLOv5m hits 1539 FPS at 640x640! YOLOv7 manages 859 FPS, YOLOv8l gets 720 FPS. For context, that's enough to process dozens of camera streams simultaneously without breaking a sweat.If you’re testing edge-heavy ideas (multi-stream CV, robotics, smart cities), this is one to keep on your radar.

A huge box arrived today! It’s thoughtfully packed full of hardware for getting going with Axelera.The most interesting bit of course is the compact (131 x 120 mm) Axelera M.2 Eval System (PDF Product Brief). The Eval System contains the  Metis AI Inference Acceleration Card on top of a standard single-board computer (SBC). You can run the Metis Acceleration Card using standard interfaces (PCIe via an M.2 hardware socket), therefore you can use legacy SBCs with M.2 that you may already have deployed, as well as new compute hardware, depending on needs.The Metis M.2 Eval System SBC uses a Rockchip RK3588 (PDF datasheet) system-on-chip (SoC). Internally, the RK3588 contains two quad-core Cortex-A series processors plus a GPU, and lots of useful interfaces including the required PCIe.  The board supports built-in storage (eMMC) pre-installed with Ubuntu 22.04  and for additional storage there’s a micro SD card slot, as well as an M.2 socket (M-Key) for inserting an optional solid-state dr

Non c’è modo di aggiungere RAM a questa bellezza 😁?E’ perfetta per il mio Laptop o su Orange PI ( Embedded ).PCIe 5.0 ? LLM non gira su M.2 ? 😣Leggendo l’SDK… Interessante… 🤔

Hello,I am very new to this so this question might be very basic. I am trying to run some very basic matrix multiplication application on the AIPU. I have a model forward doing only GEMM and exported to onnx. I tried doing as in the tutorial with a .yaml file to then deploy it and use it (maybe with axrunmodel?) but I am failing to do so as I don’t know what to put in the model info in the .yaml file. For example I don’t know what to put in the Task_category.Could you please explain how one would do that or tell me if I am doing all this completely wrong? Thank you very much

Hi,How can I implement the following preprocessing to run the end-to-end C/C++ pipeline?I've already implemented the build_gst function of the custom decoder with C++.def override_preprocess(self, img: PIL.Image.Image | np.ndarray) -> torch.Tensor:#trasformazione immagineimg = np.array(img)if img.ndim == 2:img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB)else:img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) img = np.float32(img/255.)#Stima rumoresigma_est = estimate_sigma(img, channel_axis=-1, average_sigmas=True)sigma_est = np.sqrt(sigma_est)sigma_est = min(sigma_est, 0.08) img = cv2.resize(img, (640,480), interpolation=cv2.INTER_LINEAR)img_tensor = torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float()noise_tensor = torch.full_like(img_tensor[:1, :, :], fill_value=sigma_est, dtype=torch.float32)final_tensor = torch.cat([img_tensor, noise_tensor], dim=0).to(torch.float32)return final_tensor

Buongiorno. sto valutando di inserire nella mia azienda un sistema locale di AI.Vi chiedo gentilemente di indicarmi documentazione utile ad fare questo.Per favore pensate ad una utenza che ha solo esperienza di C++ e Linux server.A seguire un breve documentazione di progetto.Obiettivi del progetto Creare una piattaforma AI autonoma installata presso la tua sede. Gestire e monitorare impianti installati dai clienti (luci, sensori, cicli, errori, manutenzione). Integrare Antholux3 e Topiaria2 in un’unica architettura di gestione. Eliminare dipendenze cloud → tutto locale: sicuro, scalabile, proprietario. Creare un polo tecnologico di supporto, diagnostica, aggiornamento e assistenza remota Architettura generale del sistema2.1 Componenti principali Server locale AI (mini-PC, NUC o workstation) Database interno Broker di comunicazione Motore regole + AI locale Dashboard unificata Modulo OTA per aggiornare gli impianti remoti Sistema di logging e diagnostica 2.2 Tecnolog

I just discovered the Smarter Spaces Challenge and I realize I am 48 hours past the deadline.However, I have a very strong project proposal ready ("EdgePark: Privacy-First Parking Monitor") and the specific technical skills to execute it immediately (experience with NVIDIA Jetson, Raspberry Pi, and custom YOLO models).Since I am confident I can deliver a high-quality project within the timeline, is there any possibility to still submit my application? I can send my proposal details immediately.Thank you for considering a passionate latecomer!Cheers, Tino

Project Title:NEMA SmartCare – Edge AI Assistant for Safety and Health in Homes, Schools, and WorkplacesShort Description:NEMA SmartCare is an edge-based AI system running on Orange Pi 5 Plus and the Metis M.2 accelerator. It monitors posture, falls, abnormal movements, environmental risks, and biometric data from external ESP32 sensors. The goal is to make homes, schools, gyms, hospitals, and offices safer by preventing accidents and supporting well-being.How It Works:Cameras analyze posture, movements, and the environment using AI models optimized for Metis.The AI detects falls, fainting, unsafe movements, prolonged bad posture, and risky behavior.Wearable sensors (ESP32 / MadBot) send real-time vital and motion data via Wi-Fi/Bluetooth.A local AI avatar (NEMA Edge) provides real-time feedback and guidance (e.g., “You are bending your back incorrectly”).In emergencies, it sends instant alerts to caregivers, relatives, or staff.Why It’s Innovative:Fully offline, privacy-preserving edg

I wanna use axelera metis on alinx board and I wonder that is there any yocto support? 

Hi,I’m trying to deploy a custom YOLO ONNX model on the Metis compute board using the Voyager SDK.The model fails during deployment because it expects (at least that’s what I understood^^) a fixed input shape of [8, 3, 128, 256], while the Voyager pipeline always provides a batch size of 1: Got: 1 Expected: 8My question is simple:👉 Is it possible to modify or extend the Voyager SDK (operators, pipeline, custom code, etc.) so that I can create a batch of 8 images before inference?Or is Voyager strictly unable to handle ONNX models with a fixed batch > 1, meaning the model must be re-exported with batch=1 or dynamic batch?Any guidance or examples would be appreciated.Thanks!

Hey everyone! 👋I wanted to share what I’m planning to build for the challenge and hopefully connect with others doing similar things. I’m working on Aurorion Edge — a small, on-premise AI “story hub” that museums, galleries, and heritage spaces can plug into their environment to make exhibits smart without needing cloud services or internet. I already have Aurorion  which gives people narrated stories and cultural insights as they explore places. The part I’m building here is the local hardware unit that does all the sensing and triggering on-site for institutions.  Using the Orange Pi 5 Plus + Metis M.2, plus the Sonoff cameras from the kit, I’m creating a mini edge server that can: 🔹 Detect which exhibit or object someone is near🔹 Run vision models locally (no cloud streaming!)🔹 Broadcast local “context triggers” to the Aurorion app🔹 Play the right micro-story or audio tour instantly🔹 Work completely offline and preserve privacy🔹 Turn any small room into a “smart cultural spac

Last time we ran a project challenge here on the Axelera AI community, we sent out ten kits to the most inspiring project ideas, and what came back were some genuinely exciting, real-world applications of edge AI.Projects that didn’t just prove what’s possible, but showed the potential to grow into something so much bigger. The kind of ideas that could spark careers, or even become the next generation of AI-driven companies.Now it’s time to do it all again! Smarter Spaces is about taking that same creativity and applying it to the places we live, work, and visit every day.Think AI that makes your environment smarter, safer, more efficient, or simply more amazing, all running locally, privately, and completely offline.If that sounds like your kind of challenge, we’d love to see your idea. You’ve still got time to apply, and it’s totally free!👉 axelera.ai/smarter-spaces

Hello,I’m currently trying to install the Voyager SDK (I tried v1.2.5 and the last version) inside a Docker container running Ubuntu 22.04, so I can later transfer the complete SDK onto my Metis Compute Board, which is currently empty (no sudo, no preinstalled SDK).However, no matter what I do, I always end up with a partial SDK installation — the folder size never exceeds ~80 MB, and key components such as ax_runner, ax_compiler, ax_engine, and axelera-runtime are missing.Here’s what I’ve already tried: Cloned the SDK from the official GitHub repo (https://github.com/axelera-ai-hub/voyager-sdk), as described in the installation guide. Successfully ran install.sh (after removing sudo calls since I’m root inside Docker). The installation completes without critical errors but stops with the message:  WARNING: Failed to refresh pcie and firmware The final folder size is about 80 MB, and python3 -m ax_runner --help fails with: No module named ax_runner I understand from the logs that

Hello again ! I’m currently exploring the Voyager SDK on my Metis Compute Board and everything works great with the pre-installed models.Now I’d like to integrate my own ONNX model into the SDK.I’ve seen that each model in Voyager uses a corresponding YAML configuration file that defines preprocessing, inference, and postprocessing parameters.Is there any official guide or documentation explaining: how to properly add a custom ONNX model to the SDK, what each section of the YAML file means (e.g., pipeline, models, datasets, operators), especially tricky ones like weight_md5 or else, and any important details or requirements when adapting an external model (like quantization, input format, etc.)? Thanks a lot in advance for your help!If there are any example templates or recommended practices, that would be really helpful.Best regards,Thibaut

Hey again !Finally I was able to install the voyager-SDK on my metis compute board ! Turns out I was debugging too much with chat gpt and it really does nonsense things I should have read more thoroughly the user guide for the metis or  the github repository… But now I have the SDK and I can test the different models directly in the SDK :)I successfully ran the yolov5s-v7-coco.yaml via ./inference <modelname> <media> But I can only run this model…I wonder if it’s normal and the other models in the folder ax_models/ are here but not compiled and need a certain action by the user (me), or if it is a bug and I’m doing something wrong ? When I try to run for example the yolov7 model with the argument --no-display just to see if it works : ./inference.py yolov7-coco media/traffic1_1080p.mp4 --no-display I have this error : 2025-11-05 15:46:12.084642319 [W:onnxruntime:Default, device_discovery.cc:164 DiscoverDevicesForPlatform] GPU device discovery failed: device_discovery.cc:89