Recent Topics

Hello Axelera AI Team,I’m Nathan Davies, Founder and Chairman of KOPO, a U.S.-based technology company developing an integrated digital ecosystem for the U.S. and African markets.We are currently evaluating edge-AI hardware and acceleration technologies for Nathan, KOPO’s AI platform, across several classes of devices and infrastructure.Our interest in Axelera AI is broader than a single development board. We would like to understand whether your technology could support a scalable edge-AI architecture across applications such as smart devices, merchant/payment hardware, intelligent terminals, mobility, vision-enabled systems and other KOPO hardware platforms.In particular, we are interested in understanding:Which current Axelera products/platforms would be most appropriate for embedded and edge inference deployments. Supported AI frameworks, models and development toolchains. Performance and power characteristics for on-device inference. Availability of SDKs/APIs and developer hardwar

I have filed a few issues on Github. The first one is a week old and hasn’t been looked at. I’ve uncovered a few plumbing issues today as well. Most of what Wingman ships is already readable: the VS Code base and the bundled Codex engine are open source, and the wingman-dashboard extension ships as unminified JavaScript. Publishing the source here would cost little and would help in a few concrete ways:Contributions instead of reports. Several of the open issues are one-line fixes. With the source here, they could arrive as pull requests. Packaging for more distributions. Packagers for Arch, Fedora and others could build Wingman properly instead of repackaging the .deb. Trust. For a tool that runs an agent with shell access on developers' machines, being able to review the code is a strong selling point, especially for a European company competing with much larger vendors.If publishing the whole codebase isn't an option, opening just the Linux launcher (/usr/bin/wingman) and the packag

My Axelera Embedded 111c (Metis Compute Board) cannot communicate with the Metis AIPU. I need help solving the issue .Board details- BSP: 1.3.3, flashed over USB with upgrade_tool (kernel 6.1.148-rockchip-standard)- Driver: patched to 1.6.2 by installing kernel-module-metis-6.1.148-rockchip-standard_v1.8.0-r0_arm64.deb and axelera-container_1.0-r0_arm64.deb (as described in the SDK 1.8.0 release notes). /sys/class/metis/version shows 1.6.2.- Voyager SDK: 1.8.0 (axversion reports 1.8.0, axversion --driver reports 1.6.2)- Did axdevice ever work on this board before? noSymptoms- axdevice reports: "Device communication timed out", "Failed to get valid board type for device metis-0:1:0 got 12", and board_type=unknown (not responding).- axdevice --refresh -v reports: "USR_DMA_XFER failed: Connection timed out", "Failed to write ELF to device memory", and "Failed to load runtime stage0".- lspci shows the device: 01:00.0 Processing accelerators [1200]: Axelera AI Metis AIPU (rev 02) [1f9d:1100

Hi,I’m planning to install a Metis M.2 accelerator inside a Lenovo ThinkPad T480.The T480’s main M.2 storage connection is PCIe Gen3 x2, and I’m considering either the Embedded 110m or the 113m 8 GB.I mainly need to confirm:Do both the 110m and 113m work correctly on PCIe Gen3 x2? Can they be power-limited to around 7–8 W if necessary? If yes, which one would you recommend for a laptop setup focused on ROS2, robotics, computer vision, and some VLM/LLM inference? Is a custom heatsink + small fan enough for this kind of installation?The laptop runs Ubuntu on x86-64.Thanks!

Hi,I am running voyager-sdk v1.8 and trying to compile a single GeMM model for 4 AIPU cores so that all 4 cores cooperate on a single inference. According to the CompilerConfig reference, MulticoreMode lists cooperative as a valid value, described as cores cooperating on a single inference.When I set multicore_mode to cooperative with aipu_cores_used set to 4 and resources_used set to 1.0, the compilation fails. I will attach the full traceback.Using batch mode instead compiles successfully, however the input shape gets promoted from batch 1 to batch 4.My questions are: what is the intended use of cooperative mode and how does it differ from batch mode in practice? Is it available in v1.8 or is it something still in development? And if cooperative is not the right tool here, is there a recommended way to have all 4 cores work together on a single inference? Attached the python script i am using for compiling. I am using axrunmodel for running the compiled ”model”. Output of axdevice:De

Big day at Axelera today. Two things landed this morning and, while the hardware is the headline, I reckon the software is every bit as interesting, especially to everyone here in the community.AxScriptThe short version is, you’ll be able to compile your own models. Not your weights on one of our models. Your actual model, your operators, your data movement, your scheduling, up and running on Axelera silicon.If you’ve ever hit the end of a supported model list and thought "well, that's that", this is the fix. AxeleraScript, or AxScript, is a Python-based DSL that programs the AI cores directly. Two ways to use it:Write it by hand. There's documentation and an agentic harness that reads your source model and helps produce the AxScript for it, with you keeping control of what comes out. Evaluation and quantization tooling ships alongside, so you can check accuracy at each step rather than at the end. Or don't. AxGraph, the graph compiler, still takes ONNX and PyTorch and handles it autom

In case you haven’t bumped up to the latest Voyager yet, let’s take a look at the highlights. 😃 Of which there are plenty.TL;DR: Axelera Server 150p (4-chip PCIe, 8GB) and Metis chip-down designs are supported, semantic segmentation is now a first-class task alongside face recognition and licence plates, Pipeline Builder gets a much more finalised API, and there are standalone Python model recipes you can copy and edit.Three sleeper hits I think are actually the good stuff:The RISC-V toolchain swap is a free speed-up. "LLVM 20 replaces GNU newlib" reads like housekeeping. It's actually a wheel going from ~200 MiB to ~20 MiB, about 900 MiB back on disk, and up to ~10% better inference on small, control-bound networks. Faster for doing nothing. Sounds like a Dire Straits parady song.The Scheduler stops grabbing your card just because you looked at it. Connection is lazy now, and a pipeline with no op.load node never touches the device. Previously, merely referencing the scheduler claime

 I am using the Axelera Embedded 110m module (Metis AIPU) on a NanoPC-T6 Plus board.Please note that I am using the NanoPC-T6 Plus (LPDDR5 architecture), not the standard NanoPC-T6 (LPDDR4X). The official image provided for the standard NanoPC-T6 does not resolve the PCIe memory allocation failure (code:error -12) on the T6 Plus due to hardware and Device Tree differences between the two boards.I have tested the official Ubuntu 22.04 and Ubuntu 24.04 OS images provided on the official NanoPC-T6 Plus wiki page, as well as the Armbian 26.8.3 NanoPC LTS image, but the issue persists on all of them.Here are the troubleshooting steps I have attempted so far:Added the recommended kernel parameters (scode]pci=realloc=on[/code] and ncode]pcie_aspm=off[/code]) to the boot configuration. Decompiled and manually updated the hcode]ranges[/code] property in the Device Tree (ecode].dtb[/code]) to expand the non-prefetchable MMIO window for the PCIe controller node. Performed a full cold power cycle

Hello Axelera community,I wanted to share a user story of how Wingman allowed me to increased the performance of my benchmarking harness from 695 FPS to 1601 FPS. Same AI module, same model. One missing property.Edge AI Coding — Part 3: Axelera Wingman | Mario BergeronIf you are wondering what Axelera’s Wingman is bringing to the table, compared to general-purpose coding assistants, these are my 2 cents.I love where Axelera is going :)Cheers !