Skip to main content

Recent Topics

45
Ioannis
Axelera Team
IoannisAxelera Team
Voyager SDK: Release Update with New Features
Voyager SDK: Release Update with New FeaturesBlog

At Axelera AI, we’re committed to advancing edge AI with every release, and Voyager SDK v1.3 is no exception. This quarterly update brings meaningful improvements that expand platform compatibility, introduce support for new models and use cases, and enhance system control and usability.Along with the new features, we have continued optimizing the execution of neural networks on our Digital In-Memory Computing engine to deliver higher performance and efficiency. Users will see measurable gains across major model families, another step forward in delivering a reliable, production-ready platform for AI innovation.We are also including experimental features to get feedback and share our innovations with you all - including the Tiled inference seen below in the video. Using YOLOv11, we are detecting people at high-resolution (4K). By leveraging our native tiled inference feature, developers can get an order-of-magnitude improvement in detections vs. the conventional approach of inferencing

4071
D
Cadet
G
Cadet
GiodstCadet
 Voyager SDK
Yolo11-coco-onnx of the model zoo low performance

Hi,I tried the Yolo11-coco-onnx models from Model Zoo and got terrible results. Can anyone give me some advice? ./inference.py yolo11n-coco-onnx dataset --no-display --show-device-fpsINFO    : Model:      yolo11n-coco-onnx                                         INFO    : Dataset:    CocoDataset-COCO2017INFO    : Date:       2025-08-21 12:33:18.028455INFO    : Inference Time: 67178.89msINFO    : Evaluation Time: 36222.56msINFO    : Evaluation Metrics:INFO    : =========================INFO    : | mAP_box       | 0.03% |INFO    : | mAP50_box     | 0.05% |INFO    : | precision_box | 0.08% |INFO    : | recall_box    | 1.70% |INFO    : =========================INFO    : Key Metric (mAP_box): 0.03%INFO    : Core Temp : 49.0°C                                                    INFO    : Metis : 804.0fpsINFO    : CPU % : 31.4%INFO    : End-to-end : 74.7fps./inference.py yolo11s-coco-onnx dataset --no-display --show-device-fpsINFO    : Model:      yolo11s-coco-onnx                            

725
H
Axelera Team
V
Axelera Team
Victor LabianAxelera Team
Factory technician monitors AI video inferencing detecting defects
A New Edge AI Hardware Option for BalenaOSBlog

Developers and engineers looking for new options for their fleets of devices using edge AI are getting them.BalenaOS has become popular precisely because it solves real deployment problems. The container-focused operating system enables over-the-air updates, remote management, and scalable fleets across embedded and edge devices.This approach works perfectly for large-scale solutions in smart retail, healthcare, autonomous systems, robotics, and manufacturing. The challenge has been finding AI hardware that matches balenaOS's deployment philosophy.Current solutions often create vendor lock-in and setup complexity that defeats the purpose of choosing balenaOS in the first place. New AI accelerators solve this problem by delivering right-sized performance that integrates naturally with container-native workflows. Businesses can optimize both system performance and power consumption without sacrificing the deployment flexibility that makes balenaOS valuable.Axelera® AI's Metis® AIPU is pu

623
D
Cadet
L
Cadet
LindeCadet
 Voyager SDK
Is there some low level NumPy / ONNX example?

I’m doing some testing with the PCI-Express based AI card. After some initial hiccups due to motherboard compatibility which required me to swap place for GPU and AI-card and then exposing the card into a ubuntu-22.04 docker container it’s working fine. I’m impressed by the performance of the card compared to using --pipe torch and having a CUDA based backend for the benchmark problems identifying traffic. I want to do inference with not only custom weights, but also custom nets and see how it performs. My wish is to have a sample that uses numpy f32 arrays as input and output and does inference on the card given a model from a ONNX file. To keep it simple, I made a PyTorch net [3, 64][ Relu ][64, 64][ Relu ][64, 1] and trained iton a mathematical function f(x1, x2, x3) = y.This toy net was exported to ONNX. From there I wanted to test inference on the AI card before moving to more complex things.I could not see that my wanted usecase was compatible with `deploy.py` and the required it

251
L
Cadet
mcunha
Ensign
mcunhaEnsign
 Project Challenge: Recognize & React
Summer Sidekick Update: Running AI on Mentis!

This week I made progress in custom dataset creation, model training, fine-tuning, and deployment. I also advanced on the pipeline schema and ONNX model compilation.Spoiler alert: using the Axelera pipelines was straightforward thanks to the provided examples 🚀. Progress in the PipelineI worked on model generation, definition, and pipeline construction, starting with the planned pipeline.The diagram below highlights this week’s main progress:Fine-tuning the YOLO11M model with extended classes. Training a MobileNet V2 model to identify bowl food levels.The highlighted blocks show the main focus of this week.  Dataset Creation and AnnotationI collected around 400 images of the environment I want to monitor, focusing mainly on my custom pet fountain and the automatic food dispenser.Example image: For annotation, I used Label Studio, creating six new classes:bowl_empty bowl_half bowl_full fountain_minimum fountain_middle fountain_maximumAfter manually annotating all 400 images, I exported

140
mcunha
Ensign
Spanner
Axelera Team
SpannerAxelera Team
 The Axelera Forum
👋 Say Hello – Introduce Yourself to the Axelera AI Community

Welcome to the Axelera AI Community – we’re so glad you’re here.This is a space for everyone from hardware engineers and AI devs to makers, enthusiasts, onlookers and partners, so don’t be shy. Whether you’re working with AI every day or just getting started, we’d love to know more about you.To get things going, let’s have a big round of hellos:Who you are and what you’re working on Your experience or interest in the AI/edge AI world What you’re hoping to learn, share or achieve in this communityCan’t wait to get to know you all, and to build something great together.Now – who’s going first? 👇

73220
D
A
Cadet
alsuttonCadet
 Metis PCIe
Problem with axdevice --refresh causing NVMe drives to be stopped on a running system

Putting this up for folk who may hit a similar issue.If you have a machine which has a single root PCIe device, where the output of lspci -tv looks something like this;alsutton@svr204:~$ lspci -tv-[0000:00]-+-00.0 Intel Corporation 8th Gen Core Processor Host Bridge/DRAM Registers +-02.0 Intel Corporation CoffeeLake-S GT2 [UHD Graphics 630] +-08.0 Intel Corporation Xeon E3-1200 v5/v6 / E3-1500 v5 / 6th/7th/8th Gen Core Processor Gaussian Mixture Model +-14.0 Intel Corporation Cannon Lake PCH USB 3.1 xHCI Host Controller +-14.2 Intel Corporation Cannon Lake PCH Shared SRAM +-16.0 Intel Corporation Cannon Lake PCH HECI Controller +-1b.0-[01]----00.0 Micron/Crucial Technology P2 [Nick P2] / P3 / P3 Plus NVMe PCIe SSD (DRAM-less) +-1f.0 Intel Corporation B360 Chipset LPC/eSPI Controller +-1f.3 Intel Corporation Cannon Lake PCH cAVS +-1f.4 Intel Corporation Cannon Lake PCH SMBus Controller

151
H
Axelera Team
D
Cadet
dario.finardiCadet
 Metis M.2
M.2 with more RAM or a smarter off-load

I'm evaluating your devices. I really love both the European heart and the great TOPS/Watt ratio.It would be amazing for edge computing.On the other side, our company is waiting for a GX10 from ASUS/NVidia with 128Gb of shared memory.The reason why I think that the Metis form factor is more interesting than the pci-ex would need a very long document :-)I read that your SDK is supporting off-load but reading on the forum I think that your off-loading support is limited only to data and not the core of the model.You know that when you use CUDA+torch you're able to off-load some part of the model out of the GPU memory and execute it running a small part of the model at once swapping coefficients from the GPU memory to the host RAM and vice versa.Are you planning to improve your off-loading system in your SDK?Are you planning to introduce a Metis with more RAM? 

150
D
Cadet
llrds
Ensign
llrdsEnsign
 Project Challenge: Recognize & React
Life after application.py - Iterative testing and learning

Wanted to start this discussion with other challenge members to see where did you go after application.pyWhat I’m doingWhat is natural to me is to modify the standard example, application.py, the first thing I’m trying to do is swap the example for another one to get a feel for how it works. The standard example uses yolov5m-v7-coco-tracker, which I found in a reference folder, so I checked for other similar examples in that location.The problem is that this takes a while between deployments, sometimes will errors right at the point where it is complete which i’ll open another thread for.What I want to knowIs there a better way? What did the other people in the challenge do from application.py?   

29035
shashi.chilappagari
Ensign
M
Cadet
manovellaCadet
 Metis M.2
Metis M.2 Firmware Update on Orange Pi 5 Plus

Hi ​@emustafa, please I need some helpI tried to upgrade firmware of Metis M2 but system (metis installed on orangepi 5 plus) gave me some errors and now after a few of reboots I obtain always same bad response from axdevice: libtriton_linux.c:1082] Device communication timed out: device did not respond within 1 seconds. (0)WARNING: Failed to get valid board type for device metis-0:1:0 got 6Device 0: metis-0:1:0 board_type=unknown (not responding) have I some possibilities to recover it? can I run some tools to diagnose bettere situation?thank in advance

605
V
Axelera Team
A
Cadet
alsuttonCadet
 Metis PCIe
Power requirements and PCIe slots

I’m putting an idea out to see if anyone has some data to back a theory.I’ve been talking with ​@Spanner about some problems I was having with a card, and, after looking over everything I’ve done, I’m wondering if a 75W slot is *required* rather than recommended, which would mean you need to use a PCIe x16 slot.One issue I had with the card was a small v shaped cut-out in the full height PCIe back-plate, which meant that the card wouldn’t sit fully into any PCIe x16 slot that I had (photo attached so you can see what I mean). This meant I had to use a 8x or 4x slot.Looking through the PCIe spec, only 16x slots are supplied with 75W, meaning that the 8x and 4x slots I tried were limited to 25W (which I confirmed), which I’m thinking could be the cause of the problems I saw.I have a 2nd hand machine turning up next week which I can test this theory on more easily (eBay still is useful for some things), but I’m wondering whether the PCIe card should be sized for a 16x slot in order to ens

572
A
Cadet
A
Cadet
alsuttonCadet
Problem with axdevice --refresh causing NVMe drives to be stopped on a running systemNew

Putting this up for folk who may hit a similar issue.If you have a machine which has a single root PCIe device, where the output of lspci -tv looks something like this;alsutton@svr204:~$ lspci -tv-[0000:00]-+-00.0 Intel Corporation 8th Gen Core Processor Host Bridge/DRAM Registers +-02.0 Intel Corporation CoffeeLake-S GT2 [UHD Graphics 630] +-08.0 Intel Corporation Xeon E3-1200 v5/v6 / E3-1500 v5 / 6th/7th/8th Gen Core Processor Gaussian Mixture Model +-14.0 Intel Corporation Cannon Lake PCH USB 3.1 xHCI Host Controller +-14.2 Intel Corporation Cannon Lake PCH Shared SRAM +-16.0 Intel Corporation Cannon Lake PCH HECI Controller +-1b.0-[01]----00.0 Micron/Crucial Technology P2 [Nick P2] / P3 / P3 Plus NVMe PCIe SSD (DRAM-less) +-1f.0 Intel Corporation B360 Chipset LPC/eSPI Controller +-1f.3 Intel Corporation Cannon Lake PCH cAVS +-1f.4 Intel Corporation Cannon Lake PCH SMBus Controller

0
10
A
Cadet
M
Ensign
moorebrett0Ensign
 Project Challenge: Recognize & React
Elderly Guardian Update: Software Development Progress

Fall-Detection - Proof of ConceptWhat’s done Built a working end-to-end app on the Voyager SDK: camera > YOLOv8-Pose > decode > filters > overlay > action. Added OpenCV overlay (boxes + 17-key point skeleton) and optional MP4 recording. Integrated a Kasa plug trigger, I will be configuring an Alexa routine once I figure out some issues. Here is a screen recording of the fall detection working:     Pipeline YAML (high level) Model: yolov8lpose-coco  Preprocess: letterbox to 640×640, torch-totensor. Postprocess: decodeyolopose with NMS + confidence thresholds; outputs scaled back to original frame size. Outputs land in frame.meta['yolov8lpose-coco'] as: boxes (N×4, xywh), keypoints (N×17×3), scores (N, pose conf). How I decide what a “fall” is (current rules) Quality gates (to ignore furniture/noise): Min box area ≈ 90,000 px², max ≈ 250,000 px². Min pose confidence 0.65. At least 8 visible keypoints with per-keypoint confidence ≥ 0.35. Fall

221
Spanner
Axelera Team
M
Ensign
moorebrett0Ensign
 Project Challenge: Recognize & React
Elderly Guardian (My Pioneer 10 Project Challenge)

Hi everyone!My name is Brett, I am a full-time single dad, assistant soccer and softball coach, and software engineer based out of Portland, Oregon.For the Project Challenge, I’m building an AI-powered safety and wellness monitoring system for seniors living independently. The Elderly Safety Guardian will use computer vision with the Metis AI Accelerator to detect falls or unusual inactivity, and then check in with the resident through Alexa, turn on lights, and alert caregivers if needed. My goal is to give families peace of mind while helping elderly loved ones stay safe at home.This project matters to me because I’ve seen how difficult it can be for families to balance a loved one’s independence with their safety, especially as they get older. I want to create a solution that bridges that gap, empowering seniors to live confidently while still having support when it’s needed most.To start, I’ll be setting up the camera hardware and building out local fall and inactivity detection wi

34410
Spanner
Axelera Team
mcunha
Ensign
mcunhaEnsign
 Project Challenge: Recognize & React
Summer SideKick Update: AI Progress – Model Exploration and Pipeline Planning

This week, I focused on the AI aspects of the Summer SideKick project. I explored the model zoo to identify models already tested and validated on the platform, and I also ran experiments with the Voyager SDK to get comfortable with tasks such as model deployment, application development workflows, and more.For the data processing pipelines, I decided to start with a mid-size model for the first stage, which will handle the main classification task. For the next stages, which are more related to the application itself, I haven’t finished the exploration of algorithms and models yet. The SDK already provides an off-the-shelf tracker model, which I plan to use for monitoring pet activity. For other tasks, I’m considering fine-tuning CNN-based models such as MobileNet or a YOLO variation. The use of OpenCV also remains an option.Here is the pipeline I plan to implement next week:  Next week, I’ll move forward with implementing these models in the pipeline and testing how they perform on r

171
M
Ensign
M
Ensign
moorebrett0Ensign
 Project Challenge: Recognize & React
Update: Elderly Guardian

Hey everyone! Brett here, back from a much-needed family trip to Maui and excited to jump back into the Elderly Guardian Challenge.My Metis AI Accelerator System hardware arrived shortly before my trip, and I’ve successfully got the Metis board plugged in and recognized by my development machine. Huge thanks to the community and Axelera AI for making the setup process pretty straightforward, even after a few days away from the keyboard, it was simple to get started.Early Technical Progress:Right now, I’m focusing on the core of the project: real-time fall detection using computer vision. I’m currently wiring up YOLOv8 pose estimation on the Metis AI Accelerator System, leveraging its high-throughput performance to process camera feeds locally. Upcoming Integration Plans:Next steps will be integrating with Alexa for voice check-ins and automating responses (like turning on lights and notifying caregivers) based on the system’s alerts. I’ll also be refining my camera placement to maximiz

904
A
mcunha
Ensign
mcunhaEnsign
 Project Challenge: Recognize & React
Fixing Slow Python Network Performance on the AETINA RK3588

Hi,I recently faced a network speed issue on the AETINA RK3588 while running Python code.It was not a general network problem, because everything worked fine with curl.However, when executing Python code, it was very slow — even for downloading very small packages.This made it impossible to use download scripts such as:./download_prebuilt.py --allIt seems that the problem comes from IPv6 handling in Python.Two possible solutions1) Disable IPv6 system-wide# Disable at runtimesudo sysctl -w net.ipv6.conf.all.disable_ipv6=1sudo sysctl -w net.ipv6.conf.default.disable_ipv6=1# Make persistentecho "net.ipv6.conf.all.disable_ipv6 = 1" | sudo tee -a /etc/sysctl.confecho "net.ipv6.conf.default.disable_ipv6 = 1" | sudo tee -a /etc/sysctl.conf2) Patch the scripts (e.g., inference.py, download_prebuilt.py) to force IPv4Add the following at the top of the script:import socketorig_getaddrinfo = socket.getaddrinfodef getaddrinfo_ipv4(host, port, family=0, type=0, proto=0, flags=0):    return orig_get

412
Spanner
Axelera Team
lorenzo.lamberti
Cadet
lorenzo.lambertiCadet
 Voyager SDK
Firefly ITX-3588J : i can not load the metis board not detected

Hello everyone, I am trying to use the Metis PCIe (4GB) in combination with the firefly ITX-3588J.Issue: I can not load the Metis kernel module (driver) into the Linux kernelSetup: I have the Metis PCIe with 4GB of ram  the Firefly ITX-3588J with 16GB of ram with a RockChip. OS: Ubuntu 22.04.4 LTS. I have a Gen3 Pcie with 4 lanes, details below```        LnkCap:    Port #0, Speed 8GT/s, Width x4, ASPM L0s L1, Exit Latency L0s <4us, L1 <16us            ClockPM- Surprise- LLActRep- BwNot- ASPMOptComp+``` So far:I installed the ubuntu OS as described here: https://support.axelera.ai/hc/en-us/articles/25556437653138-System-Imaging-Guide-Firefly-RK3588 I installed the latest 1.3.3 voyager sdk release.  i can see the metis board from lspci``` firefly@firefly:~$ lspci | grep accelerators01:00.0 Processing accelerators: Axelera AI Metis AIPU (rev 02)```however, i do not detect any drivers for the board, even after installing the sdk. Both the following commands do not return anything:```

262
Spanner
Axelera Team
M
Cadet
MehmetYigitCadet
 Metis M.2
Metis M.2 Accelerator Can't Recognized by Jetson AGX Orin

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 :)

22411
Spanner
Axelera Team
P
Cadet
PepijnCadet
 Metis PCIe
deploy.py takes really long yolov8 model

Hi, I’m trying to deploy and inference with a yolov8 segmentation model. I tried to follow the tutorial, but I just get:INFO: deploying model yolov8sseg-coco for 4 cores. This may take a while… How long should this run? I left it on for 3.5 hours and nothing happened (except the loading bar). This seems to be a model included in your library: yolov8sseg-coco.Any idea if I can turn on more extensive logging to see what is happening? Thank you in advance!  

9615
Spanner
Axelera Team
llrds
Ensign
llrdsEnsign
 Project Challenge: Recognize & React
How to solve "ModuleNotFoundError: No module named 'axelera.app'"

When trying to run the application.py example, I get this error. The vent is activated. Per this thread I thought I could solve with a removal of the venv and reinstall but it’s being persistent now. Has anybody else found and overcome this?I’m using the Aetina Eval Board that came with the Project challenge kit.

655
A
Axelera Team
H
Cadet
HermioneCadet
 Voyager SDK
inference.py quick start - import Gst failed ModuleNotFoundError

Hi,My goal is to try out a few tutorials with the latest SDK release. Looks like a python package is not found, any clues what I could try?System: Axelere M.2 Card with Aetina Eval SystemVoyager SDK v1.3.1 - a fresh install of the SDK this timeFirst installed the latest driver https://software.axelera.ai/artifactory/axelera-apt-source/metis-dkms/metis-dkms_1.0.2_all.debthen I ran the installer with./install.sh --all --media --user <email addr> --token <token>I got a few suspicious warnings and hit ‘y’ a number of times. Then install of 186 packages completed.[185/186] Install gstreamer1.0-rtspbuilding operatorsrefreshing pcie and firmware0000:01:00.0 : DeviceDevice 0: metis-0:1:0 1GiB m2 flver=1.2.0-rc2 bcver=1.0 clock=800MHz(0-3:800MHz) mvm=0-3:100%Installation complete, but with unresolved issues (see above) firefly@aetina:~/Documents/g2-testing/voyager-sdk$ source venv/bin/activate I proceeded to try inference.py with one of the included videos as input(venv) firefly@aet

806
M
M
Ensign
moorebrett0Ensign
 Project Challenge: Recognize & React
Elderly Guarding update: KISS

Elderly Guardian – Project Challenge Progress Update This week, I’ve caught myself coming up with overly complex solutions for device connectivity problems. I’ve had to step back and reframe my approach several times, staying true to the KISS principle: Keep It Simple, Stupid. My focus remains on building something practical, reliable, and quick to deploy.The Metis board’s operating system only includes built-in drivers for certain Realtek chipsets. The TP-Link plug I originally considered wasn’t the issue, it’s that the Wi-Fi dongle I first tried used a chipset without native driver support, meaning I would have had to manually compile and install a driver before I could even get online. For a hackathon project, that extra setup time wasn’t worth it.The Edimax EW-7822UAC Wi-Fi Adapter uses a supported chipset, making it far closer to plug-and-play on the Metis system. Paired with a USB hub to solve the limited port problem, it let me get connected quickly and focus on building the act

211
Spanner
Axelera Team
S
Cadet
SiegfriedCadet
uv buildNew

uv (pip alternative) build for the voyager-sdk. Many organizations are switching over to uv due to its many benefits over pip. This would speed up setup of the sdk and allow easier merging with development environments  

4
70
S
Cadet
llrds
Ensign
llrdsEnsign
 Project Challenge: Recognize & React
Can we control the fan speed on Aetina RK3588?

I have SDK Version 1.3.3, there is a comment in release notes about fan but it seems to be for the Metis. I saw already threads discussing the PCIe and M.2 fan speed on the Metis, this question is about the Aetina evalkit running at 100% fan all the time.There are none of these types of details in here.https://support.axelera.ai/hc/en-us/articles/22846665291794-AETINA-RK3588-Industrial-Motherboard-M-2I tried wider searching Aetina and Firefly on Google, there are some suggestion that control might be possible in /sys/class/hwmon but I could not find a match, or I should be looking for something else in there.Can I make the fan speed of the SBC by dynamic?

1589
llrds
Ensign

Badge Winners

Show all badges
Code of Conduct

Address

HTC5, High Tech Campus
5656 AE Eindhoven
The Netherlands
Email: info@axelera.ai

Menu

AI Accelerators

AI Software

Metis Evaluation Systems

Partners

Product Inquiry

Company

Our Story

Our Team

Careers

News

Contact

Reducing CO₂ with
Axelera’s Forest

Follow Us

LinkedIn

Twitter

Sign Up for Our Newsletter

EU flag Axelera AI is co-funded by the European Union’s Horizon research and innovation programme.

Privacy Policy  |  Cookie Policy  |  Terms of Sale  |  Vulnerability Disclosure Policy

Copyright 2025 © All rights Reserved.