AI Lounge

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.

Hi Axelera community,I’ve just started working with the Metis SBC and Voyager SDK, and I’m looking for some structured guidance on how to get up to speed.My background:~10 years of hardware experience (board-level design, complex FPGA design using VHDL/Verilog) Comfortable with Python and have worked with AI models at an academic level Limited embedded experience beyond bare-metal concepts (I’ve usually worked alongside dedicated embedded engineers)I’d love suggestions on: A learning path for Metis + Voyager SDK Where to start (simple “hello world”–style projects) How to progress toward more complex designs (e.g., integrating peripherals, using accelerators, optimizing performance, etc.) Career/skills development advice For those who have followed a similar path, from low-level hardware/FPGA work into higher-level embedded Linux and SoC-based systems:

Hi everyone,Has anyone here collaborated with remotely hired AI or edge AI engineers on Axelera Metis or Voyager SDK projects, for example, through a remote recruitment agency like CrewBloom?I’d love to hear how you handled onboarding, access to hardware, SDK setup, and day-to-day collaboration in a fully remote setup. Any lessons learned or tips that helped projects run smoothly would be greatly appreciated.

The NanoPC-T6 from FriendlyElec is now supported by Voyager SDK. This is one that's been requested a few times (shout out to the folks who’ve previously asked for it), so I figured it was worth a proper write-up.What's the NanoPC-T6?It's an RK3588-based SBC that's been popular in the community for a while, particularly for networking and NAS builds. The board itself is pretty compact and comes with dual 2.5GbE ports, an M.2 slot for NVMe storage, and somewhat unusually for an SBC, an HDMI input alongside the outputs. Pricing starts around $100 for the 4GB/32GB variant and goes up to about $150 for the 16GB/256GB version. So it’s an affordable piece of kit and a smart choice for the right projects.

hi there,i am currently writing up an application with friends for a grant on “ai&art” which is due by tomorrow.we want to do music reactive visuals on edge with multimodal inputs and want to use the Metis for live inference. i think of visuals being an interesting playground since they allow to very freely construct the loads for the chip. if you are interested in collaborating please feel free to reach out! there is no requirement on expertise.  ideas only are also very much welcome.(as well as additional sources for funding or sponsors)hope to have caught your interestbest carlo 

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

Have you seen that ​@Victor Labian’s successfully brought up the Voyager SDK on the Orange Pi 5 Plus with a Metis M.2 module? It took a bit of work—Ubuntu flashing, driver rebuilds, and tweaking the PCIe device tree—but it’s running, and it’s fast. 😃This setup is seriously impressive for something so compact and affordable. Inference runs smoothly, and with some tuning, it could potentially be running advanced models like YOLOv8 without trouble.It’s just an experiment and a test at this point. No big plans around the Orange Pi, but it looks like a really promising start.If you’re experimenting with AI on SBCs—or thinking about it—I’d love to hear what you’re working on. Orange Pi, Raspberry Pi, custom boards, anything edge. 👍

All the release notes for Voyager updates go straight out on the GitHub repo, so I don’t tend to parrot them here, but this is a pretty big one and feels like it deserves a spot here on the community!If you’ve been using 1.3 already, there are some nice changes that make life easier.A few things that particularlt stand out to me:🔹 Cascade pipelines feel smoother. Chaining models together (like running a classifier on detections) takes less fiddling around than before. 🔹 More operator and metadata options. You can now shape pre/post-processing and inference results in ways that fit better with your own app. 🔹 Model Zoo is easier to adapt. Swapping in your own weights on the reference pipelines is less hassle, so you can get something tuned to your dataset faster. 🔹 Custom evaluators are handy if you want accuracy measured in a way that matches your own project rath

Dear Axelera Team, We are Master’s students in Business Engineering: Technology and Entrepreneurship at KU Leuven. As part of our course on Strategic Management of Technology, we are exploring how innovative companies manage and foster innovation. Through Axelera’s LinkedIn we noticed your commitment to supporting innovative ideas. We would be very interested in learning more about your approach to innovation management. Our project would involve studying your innovation practices and providing an analytical report as an outcome. Would you be open to collaborating with us on this project? Thank you very much for your time and consideration. Best regards,Anna VellaKU Leuven – Faculty of Economics and BusinessLeuven, Belgium 

Buongiorno a tutti, Mi chiamo Vito e sono il titolare di una società Italiana, una start-up. Insegno anche in un istituto tecnico di Brescia Robotica, Elettronica e Informatica. Abbiamo avviato un progetto per la realizzazione di un robot umanoide nella mia società con alcuni dei miei studenti, qualche mese fa nell'ambito dell'assistenza. Ho visto l'intervista dell'Ing. Del Maffeo e sono rimasto colpito. Spanner, ​@Spanner dobbiamo implementare il robot con AI e quindi di una scheda potente, possiamo collaborare? Il potenziale è grande..  come potremmo approfondire l'argomento? Vi ringrazio a tutti in anticipo.. 

Thought you might want to check out the live stream from YOLO Vision that’s happening today in London. Some of the Axelera AI team are there, and will be up on stage soon! 

Hello,I am currently running performance tests on the Metis M.2 card with different models (e.g., YOLOv8n).I observed a behavior that I would like to clarify: When running with 1 AIPU core, I get lower latency and the same FPS as when using multiple cores (4). However, with 4 AIPU cores, the latency per image is higher, even though the FPS does not increase. In addition, CPU utilization is noticeably higher with 4 cores (around 17%) compared to only ~0.8% with 1 core. I would like to know: Is this expected behavior (due to synchronization overhead between AIPU cores)? Is there a recommended configuration to optimize either latency or throughput (FPS) depending on the use case? Is it normal that CPU usage increases significantly when using 4 cores, given

In case anyone on here hasn’t spotted it in the wild yet, the Metis M.2 Max was just announced. It’s basically managed to pack PCIe card-level performance into an M.2 stick. So that’s better performance for LLMs and vision transformers on much smaller host devices (RPi, Orange Pi, etc), all while keeping power draw super low (around 6.5W). It’s been bumped up to 16GB of memory, slimmer design options, and added security features for tougher environments.Feels like a big step forward for anyone wanting to run GenAI, computer vision, or LLM workloads locally, on devices with an impressively small footprint, without needing a data centre to do it 😆.Shipping starts later this year.

Heads-up if you’re building a new AI project from scratch – the new Metis Compute Board is now up for pre-order!It’s a compact SBC with the Metis AI accelerator baked right in, powered by the ARM-based RK3588. And it runs the Voyager SDK out of the box. Loads of interfaces, solid memory specs, tiny footprint (170x170mm), and very dev-friendly.Integrated Metis AIPU with 214 TOPS for multi-stream, multi-model AI workloads 16 GB RAM for the CPU + 4 GB for the AIPU, with storage and connectivity for almost any setup Comes with Voyager SDK and BSP for fast prototyping and deploymentThere’s a product brief attached if you want to see a few more details, too.Bonus: pre-orders get 15% off. Shipping is slated to begin in June 2025, and you get 15% off for pre-ordering.Check it out: 

We are happy to announce the release of our latest PySDK (version 0.18.0) that supports Axelera runtime 1.3.3. See our docs: Overview | DeGirum Docs for more details. Examples and user guides coming soon.

I suppose it’s inevitable that there’ll always be a new model in the works, but hushed conversations about the Orange Pi 6 Plus are starting to surface…Considering how well the Orange Pi 5 Plus has (rather unexpectedly) performed in low-cost edge AI applications, it’s hard not to feel some small pang of excitement around the next version!https://www.reddit.com/r/OrangePI/comments/1mbd5zc/the_upcoming_orangepi_6_plus/We need to put one in the hands of ​@Victor Labian and ​@Steven Hunsche!

We're excited to announce that Axelera is now integrated with DeGirum AI Hub—a platform that lets you run edge AI models directly in your browser without any setup or installations.🔍 What You Can DoExplore a wide range of ready-to-use AI models for real-time inference across various use cases: 🧑 Face Detection – detect and track faces in video streams ✋ Hand Detection – enable gesture recognition and interaction 🔥 Fire and Smoke Detection – build early warning systems 🚘 License Plate Detection – power smart parking and traffic analytics 🚗 Car Detection – monitor traffic flow, parking spaces, and more All of these can now be tested instantly using Ax

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

I've put together a quick walkthrough of how to set up the Voyager SDK with a Metis M.2 card on the Orange Pi 5 Plus, and wanted to share the full process with you all. This has turned out to be a fantastic low-cost dev setup to start experimenting with edge AI using our Metis platform.Here’s what we cover:Installing the Metis M.2 card Setting up the Orange Pi with the proper OS image Installing the Voyager SDK (using SSH, so you can do it headless) Running a Yolov8L demo with live inferenceEverything is done using standard tools like Balena Etcher, SSH, and a few terminal commands. Whether you’re building a prototype or just exploring Metis on embedded platforms, this setup is a great place to start on a device that's cheap, powerful, and very easy to use.

Just FYI, I’ve been tinkering with the platform design a bit, as I always felt like the code block text input burned your retinas out, being such a big block of white against the dark theme 😆I’ve mimicked GitHub's colour pallette for it, but we can tweak it further if anyone has a preference. (Or if you spot anywhere that’s now messed up because of the CSS changes, let me know),As you were!

TechCrunch piece about a nonprofit that had GPT and Claude models run a fundraising op on their own—no human help except setup. The AIs picked a charity (Helen Keller Intl), made docs, set up socials, emailed people, and ended up raising $257 in a week!Can’t tell if this is good or bad...https://techcrunch.com/2025/04/08/a-nonprofit-is-using-ai-agents-to-raise-money-for-charity/ 

I’m using Axelera’s Voyager SDK to run a YOLOv8s object-detection pipeline on video files, and I need to capture every detection into a JSON array for downstream processing. For each detected object, I want to record: frame_index: the zero-based frame number timestamp: the frame’s PTS in seconds bbox: [x1, y1, x2, y2] coordinates class_id: numeric COCO class index class_name: readable label (e.g. “person”) score: detection confidence What’s the cleanest way to hook into the Voyager inference loop, pull out the ObjectDetectionMeta (task name “detections”), convert its boxes, scores, class_ids, and labels into native Python types, attach the FrameResult.src_timestamp, and dump the results into a sin