Recent Topics

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

Hr department for BBA studys complete and after BBA for MBA start  

I am integrating Axelera Metis into a yocto-based project for a customer.In their case, this is a Zynq-UltraScale+ Petalinux 2025.2 project (yocto scarthgap).I have successfully integrated the meta-axelera layerroot@zcu104-axelera-2025-2:~# lspci -vvv00:00.0 PCI bridge: Xilinx Corporation Device 9131 (prog-if 00 [Normal decode]) …01:00.0 Processing accelerators: Axelera AI Metis AIPU (rev 02) Subsystem: Axelera AI Metis AIPU (rev 02) Control: I/O- Mem- BusMaster- SpecCycle- MemWINV- VGASnoop- ParErr- Stepping- SERR- FastB2B- D- Status: Cap+ 66MHz- UDF- FastB2B- ParErr- DEVSEL=fast >TAbort- <TAbort- <MAbort- >SERR- <PERR-- Interrupt: pin A routed to IRQ 89 Capabilities: [40] Power Management version 3 Flags: PMEClk- DSI- D1+ D2- AuxCurrent=375mA PME(D0+,D1+,D2-,D3hot+,D3cold-) Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME- Capabilities: [50] MSI: Enable- Count=1/32 Maskable+ 64bit+

Hardware: Axelera Metis PCIe  |  SDK: Voyager v1.6  |  OS: UbuntuContextI am trying to compile Cellpose cyto3 (a cell segmentation model) for the Metis AIPU as part of a benchmarking project comparing CPU and AIPU inference performance on kidney electron microscopy patches (512×512, uint8).What I have done so farThe ONNX export now works cleanly. I wrap the Cellpose network in a small PyTorch wrapper that replaces the make_style branch with a zeros tensor (the make_style branch uses Flatten/Pow/ReduceSum/Div which are unsupported). Export produces:Inputs: 1 — input, shape [1, 2, 512, 512] Outputs: 1 — output, shape [1, 3, 512, 512] Nodes: 215 — op types: Add (32), BatchNorm (33), Conv (41), Gemm (12), MaxPool (3), Relu (33), Resize (3), Unsqueeze (24), Constant (18) opset: 11  |  file size: 25.3 MBThe errorWhen I call compiler.quantize() from the Python API I get:QuantizationError: Input model cannot be quantized because the core graph has several inputs that are not currently supporte

Hi everyone,I am trying to get a Metis M.2 card working on Windows with Voyager SDK 1.6.0.Current status:Windows driver (MetisOmega 1.3.4) installed successfully Secure Boot disabled Test signing enabled Device detected correctly by axdevice Runtime firmware 1.6.0 loads successfullyaxdevice -v shows:Device 0: metis-0:0:0 1GiB m2flver=1.3.2bcver=1.4device_runtime_firmware=v1.6.0However, inference fails with both YOLOv5 and ResNet examples.Example:axrunmodel ...\resnet50-imagenet-onnx\1\model.json[ERROR][waitForQueueBinaryCompletion]: Wait kernel failed with return code 1.[libaxldev.c:140] Failed to wait for MSI 1: No errorFAIL: [WinError -1073741569] Windows Error 0xc00000ffThe device is detected and firmware loads, but execution fails when waiting for MSI interrupts.Has anyone seen this issue before on Windows with a Metis M.2 board? Any suggestions on firmware, driver, BIOS, PCIe/MSI settings, or compatibility requirements?Thanks!

Voyager SDK v1.6.1 is a patch on top of v1.6.0. No huge new feature set, but useful cleanup, real Windows performance gains, closed-loop power control on the upcoming Metis M.2 Max, and a proper docs portal at docs.axelera.ai.Windows is FasterThat’s not something you hear about Windows very often, is it? But this is referring to Voyager on Windows, so that might explain it. 😋The PCIe driver and libaxldev have both had performance work, closing the gap with Linux, which the Windows fans out there will appreciate. Logging and tracing infrastructure is also improved, with a utility for viewing traces.By the way, before we go too much further, this release requires Metis Driver 1.3.2 or newer on Windows. v1.3.1 and earlier are no longer supported, so v1.6.1 ships with driver 1.3.11.Metis M.2 Max Power ControlClosed-loop power control on M.2 Max has been improved ahead of its launch:No more unexpected performance degradation. The limit set via axdevice --set-power-limit now matches what ax

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 :)

Hello,There seems to be two different published benchmarks for ResNet50.The first is in the Voyager SDK’s documentation : 1756 FPS (M.2) / 1946 FPS (PCIe) :- https://github.com/axelera-ai-hub/voyager-sdk/blob/latest/docs/reference/model_zoo.mdThe second is on the Metis Marketing page : Up to 3200 frames per second (ResNet-50)- https://axelera.ai/ai-accelerators/aipu/metis I was able to achieve 2050 FPS, which is higher than the modelzoo documentation, but MUCH lower than the marketing claim:- https://mariobergeron.com/posts/edge-ai-power-p04-axelera-metis/ Where does the 3200 FPS benchmark claim come from ?   How was this benchmark achieved on Metis ?  Which hardware platform ?Which brings me to the next marketing claim : Industry-Leading FPS / Watt ratios on ResNet50v1 (graph showing 614 FPS/W- https://axelera.ai/ai-accelerators/aipu/metis I was only able to achieve 270 FPS/W :- https://mariobergeron.com/posts/edge-ai-power-p04-axelera-metis/ What does the 614 FPS/W benchmark come fro

How can I deploy WASB model on metis m2?

I'm trying to deploy MitoNet v1 (a mitochondria segmentation model based on PanopticDeepLab) on the Axelera Metis AIPU using the Voyager SDK v1.5, and I'm running into issues exporting the model to ONNX.MitoNet is available via the empanada-napari library and its weights are stored as a TorchScript archive (.pth). When loaded, the model is a `torch.jit.RecursiveScriptModule` that takes a (1, 1, H, W) float32 tensor as input and returns a dict with keys `sem_logits`, `ctr_hmp`, and `offsets`.The problem is that the model uses `aten::__interpolate` with an `int[]` size argument, which blocks all ONNX export paths I've tried:1. `torch.onnx.export` with opset 11 and 17 → `INTERNAL ASSERT FAILED: We don't have an op for aten::__interpolate`2. `TS2EPConverter(model, args).convert()` → same `aten::__interpolate` error3. `axelera.compiler.quantize()` directly on the TorchScript model → `ValueError: Failed to get input shape from calibration dataset`Environment:- Voyager SDK v1.5- PyTorch 2.10.

Hello,I am trying to compile an IBM Granite embedding model for the Metis AIPU.Current test model:```textibm-granite/granite-embedding-311m-multilingual-r2https://huggingface.co/ibm-granite/granite-embedding-311m-multilingual-r2```My final target model is:```textibm-granite/granite-switch-4.1-3b-previewhttps://huggingface.co/ibm-granite/granite-switch-4.1-3b-preview```Target hardware and software:```textMetis SBC16 GB board RAM4 GB AIPUPython 3.12Voyager SDK 1.6.0```I tried to compile the embedding model with `axcompile`, but it fails during quantization.Command:```bashaxcompile \  --input granite_embedding_metis_work/granite_embedding_311m_static.onnx \  --config granite_embedding_metis_work/granite_embedding_metis_4gb.json \  --output granite_embedding_metis_work/axcompile_out \  --overwrite \  --dataset-len 2 \  --log-level DEBUG \  --quantize-only```The ONNX export succeeds.Current ONNX details:```textONNX opset: 17Input: inputs_embeds [1, 128, 768]Output: embeddings [1, 768]```The

Hi,We are using the Axelera Metis PCIe AI accelerator card with a ZCU102 evaluation board as the host system. We are facing a critical compatibility issue when running the Voyager SDK on the ZCU102.---Setup Details:- Host Board: AMD Xilinx ZCU102 (Zynq UltraScale+ MPSoC)- Host CPU: Quad-core ARM Cortex-A53 (ARMv8.0-A)- OS: Custom Yocto Linux (kernel 6.12.10-xilinx)- Axelera Card: Metis PCIe AI Accelerator- Voyager SDK Version: Latest (installed via install.sh --no-driver --media)- metis Driver Version: 1.4.16 (loads successfully)---Issue:When running inference.py or importing axelera.app, the process crashes immediately with:  Illegal instruction (core dumped)  SIGILL (signal 4)---Root Cause Identified:After detailed debugging, we found that the Voyager SDK Python packages (scipy, torch, cv2, grpc, onnx etc.) are compiled with LSE (Large System Extensions) atomic instructions such as ldadd, stadd, swpa, cas.The ZCU102 Cortex-A53 is ARMv8.0-A which does NOT support LSE instructions. LSE

Hello Axelera support team,I am running the Axelera Metis M.2 accelerator on a Radxa Rock 5B (RK3588 SoC) with Armbian and kernel 6.18.10-current-rockchip64. After extensive debugging over several days, I have identified a compatibility issue between the metis driver and the ARM SMMU-v3 on this platform. I would appreciate your guidance on resolving this.HARDWARE & SOFTWAREBoard: Radxa Rock 5B (RK3588, 8GB RAM, PCIE3.0-4X) OS: Armbian, kernel 6.18.10-current-rockchip64 (mainline) Accelerator: Axelera Metis M.2 Driver version: metis 1.4.16 SDK/Runtime version: voyager-sdk / runtime-1.6.0-1 PSU: 65WWHAT WORKSThe Metis is detected correctly via PCIe: lspci shows "Axelera AI Metis AIPU (rev 02)" at 0000:01:00.0 After patching the Device Tree to expand the PCIe memory window (the default 14MB window was too small; expanded the 64-bit prefetchable range at 0x900000000 to 128MB), the driver loads successfully: /dev/metis0 -> metis-0:1:0 is created "Axelera AIPU PCIe Driver, version 1.

Dear Axelera Support,I am currently following the Voyager SDK quick start guide (YOLO inference demo) on an RK3588-based Compute Board (received out-of-the-box), using Voyager SDK version 1.5.3. However, I am unable to successfully run inference due to persistent device communication issues with the Metis AIPU.--------------------------------------------------System Overview- Platform: RK3588 (Compute Board, out-of-the-box setup)- Voyager SDK: 1.5.3 (and also tested with 1.4.0)- Connection: PCIe (Metis AIPU detected as 01:00.0 [1f9d:1100])--------------------------------------------------Initial IssueRunning the quick start inference:./inference.py yolov5s-v7-coco usb:/dev/video1Result:ERROR: No devices foundAt that moment:- lspci returned no devices, indicating a PCIe enumeration issue.--------------------------------------------------After Cold BootAfter performing a full power cycle:- lspci correctly shows:  01:00.0 Processing accelerators: Axelera AI Metis AIPU (rev 02) [1f9d:1100]

My colleagues don't remember the Ubuntu password and I have to use the miniPC. I have tried some passwords they have given me, but no luck.

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

I honestly proud for the AI community in that they will be good functions and I creatively appreciate for all AI humans to create some earning benefits thanks for a lot of 💝