Windows Setup

This guide covers installing the Axelera software on Windows 11. The main path is native Windows — you install the driver and Axelera components directly on Windows without WSL.

Step 1: Install the Windows driver​

Windows Update might distribute an older version of the Axelera driver. To check which driver is installed, open a Command Prompt or PowerShell and run:

pnputil /enum-devices /deviceid "PCI\VEN_1F9D" /drivers | findstr /c:"Device Description" /c:"Driver Version"

The output lists each Axelera device and its driver:

Device Description:         Metis Device
Driver Version:             09/21/2026 1.4.5.0

If every Axelera device reports a Driver Version of 1.4.5, complete 1a and continue with Step 2. Otherwise, or if no driver is listed, follow the steps below to install it manually.

With a 150p card, install the Switchtec driver manually in any case, as described in 1e.

1a. Grant "Lock pages in memory" privilege​

This step is required once per machine. The Axelera runtime needs the Lock pages in memory privilege to operate correctly.

1. Press Win + S, type secpol.msc, and press Enter to open the Local Security Policy editor
2. In the left pane, navigate to Local Policies → User Rights Assignment
3. In the right pane, double-click Lock pages in memory
4. Click Add User or Group, enter your Windows username, and confirm

Log out and back in (or reboot) for the privilege to take effect.

1b. Download the drivers​

Download the driver archives from the Axelera software portal:

• Metis driver (required for all hardware): Metis-1.4.5.zip
• Switchtec PCIe switch driver (required only for the Axelera Server 150p (PCIe 4-AIPU) card): switchtec-kmdf-0.7_2019.zip

Extract each archive locally. Each extracted folder contains three files: .cat, .inf, and .sys.

1c. Remove any previous driver version​

If you already have an older Metis driver installed, remove it completely before installing the new one. Windows can keep multiple drivers for the same device.

In Device Manager, locate the Metis device under System devices, right-click and select Uninstall device:

In the confirmation dialog, check "Attempt to remove the driver for this device", then click Uninstall:

Verify the device now appears under Other devices as a generic PCI Device — this confirms the driver is fully removed:

1d. Install via Device Manager​

Open Device Manager (right-click the Windows Start button → Device Manager).

From the menu, open Action → Add Drivers:

Set the path to the folder where you extracted the Metis driver files:

Confirm that the Metis device now appears under System devices:

1e. Install the Switchtec driver (150p card only)​

Repeat the steps in 1d, this time setting the path to the folder where you extracted the Switchtec driver files. Confirm that the Switchtec device appears:

Step 2 (Optional): Install WSL2 for model deployment​

Network deployment is supported exclusively in Linux. WSL2 is the recommended approach for deploying models for Windows use.

2a. Install WSL2​

Open PowerShell as Administrator and run:

wsl --install -d Ubuntu-22.04

Restart your computer when prompted. After restart, Ubuntu opens and asks you to create a username and password.

2b. Clone the SDK in a Windows-accessible folder​

After the restart, open a regular Command Prompt (not PowerShell, not Administrator) and enter WSL:

wsl

Inside WSL, clone to a path under /mnt/c/ so both Windows and WSL can reach it:

# /mnt/c/ maps to C:\ on Windows
cd /mnt/c/Axelera
git clone https://github.com/axelera-ai-hub/voyager-sdk.git
cd voyager-sdk

This means:

• From WSL: /mnt/c/Axelera/voyager-sdk
• From Windows: C:\Axelera\voyager-sdk

2c. Follow the Linux setup inside WSL​

Install the SDK inside WSL, then use it for model deployment:

1. SDK Install — clone, install, and activate
2. deploy.py — compile and deploy models for use on Windows

Step 3: Install Python (Windows native)​

1. Download Python for Windows from the official website. The latest stable release tested by Axelera is 3.13.5.

2. During installation:

   • Enable "Add Python to PATH"
   • Select "Disable path length limit" when prompted

3. Add the Python installation path and the Python\Scripts path to the Windows Path environment variable:

   

4. Create a virtual environment:

   python -m venv venv-win
   

Step 4: Install Windows Axelera components​

Open a regular Command Prompt (not PowerShell, not Administrator) and navigate to your SDK directory:

cd C:\Axelera\voyager-sdk

Create a local download directory and download all packages:

rmdir /s /q windows-packages 2>nul
mkdir windows-packages
cd windows-packages

curl -L -O https://software.axelera.ai/artifactory/axelera-win/packages/1.9.x/axelera-win-device-installer.exe
curl -L -O https://software.axelera.ai/artifactory/axelera-win/packages/1.9.x/axelera-win-runtime-installer.exe
curl -L -O https://software.axelera.ai/artifactory/axelera-win/packages/1.9.x/axelera-win-services-installer.exe

cd ..

4a. Install the executable packages​

Using Windows Explorer, navigate to C:\Axelera\voyager-sdk\windows-packages and run each .exe installer in sequence.

After installation, verify via Settings → Apps & features (or Add or remove programs):

All three packages must be listed:

You must see: Axelera Device Package, Axelera Runtime, and Axelera Services. Exact version numbers will vary.

Close all Command Prompt and PowerShell windows after the installers complete, then open a new Command Prompt before continuing.

4b. Install Python packages​

Activate the virtual environment and install the packages:

venv-win\Scripts\activate.bat

pip install --extra-index-url https://software.axelera.ai/artifactory/api/pypi/axelera-pypi/simple axelera-runtime==1.9.0 axelera-llm==1.9.0 axelera-types==1.9.0 axelera-runtime2

axelera-runtime2 provides the Pipeline Builder API; install the version matching your SDK release. It requires the Microsoft Visual C++ Redistributable (x64), which most systems already have.

venv-win\Scripts\activate.bat

4c. Install a GLSL compiler (for the Vulkan renderer)​

The Vulkan renderer compiles its shaders at runtime and needs a GLSL compiler. Skip this step if you only use --display opencv, console or none; on Linux the equivalent is the glslang-tools package.

Recommended - the Vulkan SDK via winget:

winget install KhronosGroup.VulkanSDK

The SDK sets VULKAN_SDK, which the renderer searches directly, so no Path change is needed. winget tracks upstream releases, so there is no version to keep up to date here.

Alternative - glslang on its own. The SDK is a large install for a single executable. To fetch just the compiler, in PowerShell:

$asset = (Invoke-RestMethod https://api.github.com/repos/KhronosGroup/glslang/releases/latest).assets |
    Where-Object name -like '*windows-x86_64-release.zip'
Invoke-WebRequest $asset.browser_download_url -OutFile "$env:TEMP\glslang.zip"
Expand-Archive "$env:TEMP\glslang.zip" -DestinationPath C:\glslang -Force

Then add C:\glslang\bin to the Windows Path environment variable.

Verify. Open a new Command Prompt so the environment change is picked up, activate the virtual environment, and run:

python -c "from axelera.runtime2 import display_vk_shaders as s; print(s.compiler())"

This performs the same lookup as the renderer and prints the compiler it found.

Step 5: Run examples​

First, obtain a deployed model. If you completed the WSL2 setup in Step 2 and deployed a model there, it is accessible from Windows at C:\Axelera\voyager-sdk\build\. Otherwise, download a pre-built model:

axdownloadmodel resnet50-imagenet-onnx

The model downloads to C:\Axelera\voyager-sdk\build\.

Run inference using the axrunmodel utility (measures performance):

axrunmodel C:\Axelera\voyager-sdk\build\resnet50-imagenet-onnx\resnet50-imagenet-onnx\1\model.json

Run inference using the Python API:

Download an image to C:\Axelera\voyager-sdk\examples\axruntime\images, then:

cd examples\axruntime
python -m axruntime_example -v --aipu-cores 1 --labels imagenet-labels.txt ..\..\build\resnet50-imagenet-onnx\resnet50-imagenet-onnx\1\model.json images

Run a model recipe with the Pipeline Builder API​

Model recipes under model_recipes/ run natively on Windows against a pre-deployed model. Deployment itself stays in Linux/WSL2: from WSL (Step 2), build the model once -

python model_recipes/pose/yolo11n_pose_coco_axrelay.py --build

• which writes build/yolo11n-pose-coco.axm into the shared checkout.

Then, back in native Windows with the virtual environment active, run the same recipe against that artifact:

python model_recipes\pose\yolo11n_pose_coco_axrelay.py --axm build\yolo11n-pose-coco.axm --run image.jpg --display none

Detections are printed to the console. To render them in a window instead, use --display vulkan (the preferred renderer, requiring the GLSL compiler from Step 4c), --display opencv, or --display auto to pick the best available. Inputs can be image files, video files, or network streams (RTSP/HTTP) - cameras are not supported on Windows.

Step 6: Run LLM models​

A set of precompiled LLM models can run natively on Windows. To see available models:

axllm --help-network

Run a model in single-prompt mode:

axllm llama-3-2-1b-1024-4core-static --prompt "Give me a joke"

Run interactively in chat mode:

axllm llama-3-2-1b-1024-4core-static

For all options, see the LLM Inference Guide.

Using PowerShell​

After completing installation, you can use PowerShell instead of Command Prompt. The only difference is how you activate the virtual environment:

.\venv-win\Scripts\Activate.ps1

Once activated, all the same commands work as in Command Prompt.

Get-ExecutionPolicy

Set-ExecutionPolicy RemoteSigned

Troubleshooting​

See also​

• SDK Install — the main Linux/WSL installation guide
• Verify Setup — confirm hardware and SDK are working
• LLM Inference Guide — full LLM options including Windows