Axelera AI Export and Deployment

Ultralytics partners with Axelera AI to enable high-performance, energy-efficient inference on Edge AI devices. Export and deploy Ultralytics YOLO models directly to the Metis® AIPU using the Voyager SDK.

Axelera AI provides dedicated hardware acceleration for computer vision at the edge, using a proprietary dataflow architecture and in-memory computing to deliver up to 856 TOPS with low power consumption.

Selecting the Right Hardware

Axelera AI offers various form factors to suit different deployment constraints. The chart below helps identify the optimal hardware for your Ultralytics YOLO deployment.

mermaid

graph TD
    A[Start: Select Deployment Target] --> B{Device Type?}
    B -->|Edge Server / Workstation| C{Throughput Needs?}
    B -->|Embedded / Robotics| D{Space Constraints?}
    B -->|Standalone / R&D| E[Dev Kits & Systems]

    C -->|Max Density 
 30+ Streams| F[**Metis PCIe x4**
856 TOPS]
    C -->|Standard PC 
 Low Profile| G[**Metis PCIe x1**
214 TOPS]

    D -->|Drones & Handhelds| H[**Metis M.2**
2280 M-Key]
    D -->|High Performance Embedded| I[**Metis M.2 MAX**
Extended Thermal]

    E -->|ARM-based All-in-One| J[**Metis Compute Board**
RK3588 + AIPU]
    E -->|Prototyping| K[**Arduino Portenta x8**
Integration Kit]

    click F "https://store.axelera.ai/"
    click G "https://store.axelera.ai/"
    click H "https://store.axelera.ai/"
    click J "https://store.axelera.ai/"

Hardware Portfolio

The Axelera hardware lineup is optimized to run Ultralytics YOLO26 and legacy versions with high FPS-per-watt efficiency.

Accelerator Cards

These cards enable AI acceleration in existing host devices, facilitating brownfield deployments.

Product	Form Factor	Compute	Performance (INT8)	Target Application
Metis PCIe x4	PCIe Gen3 x16	4x Metis AIPUs	856 TOPS	High-density video analytics, smart cities
Metis PCIe x1	PCIe Gen3 x1	1x Metis AIPU	214 TOPS	Industrial PCs, retail queue management
Metis M.2	M.2 2280 M-Key	1x Metis AIPU	214 TOPS	Drones, robotics, portable medical devices
Metis M.2 MAX	M.2 2280	1x Metis AIPU	214 TOPS	Environments requiring advanced thermal management

Integrated Systems

For turnkey solutions, Axelera partners with manufacturers to provide systems pre-validated for the Metis AIPU.

Metis Compute Board: A standalone edge device pairing the Metis AIPU with a Rockchip RK3588 ARM CPU.
Workstations: Enterprise towers from Dell (Precision 3460XE) and Lenovo (ThinkStation P360 Ultra).
Industrial PCs: Ruggedized systems from Advantech and Aetina designed for manufacturing automation.

Supported Tasks

The following tasks are supported across YOLOv8, YOLO11, and YOLO26 models.

Task	YOLOv8	YOLO11	YOLO26
Object Detection	✅	✅	✅
Pose Estimation	✅	✅	✅
Segmentation	✅	✅	⚠️ Voyager SDK only
Oriented Bounding Boxes	✅	✅	✅
Classification	✅	✅	✅

!!! note

YOLO26 segmentation is not yet supported through the Ultralytics `export` command. Users who need YOLO26-seg can deploy via the [Voyager SDK](https://github.com/axelera-ai-hub/voyager-sdk) using `deploy.py`, which provides a user-space workaround. Native compiler support will be added in a future release.

Installation

!!! warning "Platform Requirements"

Exporting to Axelera format requires:

- **Operating System**: Linux only (Ubuntu 22.04/24.04 recommended)
- **Hardware**: Axelera AI accelerator ([Metis devices](https://store.axelera.ai/))
- **Python**: Versions 3.10, 3.11, and 3.12
- **System dependency**: `sudo apt install libgl1` (required by OpenCV, not included via `pip`)

Ultralytics Installation

bash

pip install ultralytics

For detailed instructions, see our Ultralytics Installation guide. If you encounter difficulties, consult our Common Issues guide.

Axelera Driver Installation

Add the Axelera repository key:

bash

sudo sh -c "curl -fsSL https://software.axelera.ai/artifactory/api/security/keypair/axelera/public | gpg --dearmor -o /etc/apt/keyrings/axelera.gpg"

Add the repository to apt:

Choose the appropriate snippet from below to match the OS being used.

bash

# Ubuntu 22.04
sudo sh -c "echo 'deb [signed-by=/etc/apt/keyrings/axelera.gpg] https://software.axelera.ai/artifactory/axelera-apt-source ubuntu22 main' > /etc/apt/sources.list.d/axelera.list"

bash

# Ubuntu 24.04
sudo sh -c "echo 'deb [signed-by=/etc/apt/keyrings/axelera.gpg] https://software.axelera.ai/artifactory/axelera-apt-source ubuntu24 main' > /etc/apt/sources.list.d/axelera.list"

Install the SDK and load the driver:

bash

sudo apt update
sudo apt install -y metis-dkms=1.4.16
sudo modprobe metis

!!! tip "First run downloads the SDK automatically"

The first `yolo export format=axelera` or `yolo predict` with an Axelera model will automatically download and install the Axelera SDK packages. This may take several minutes depending on your connection speed, and no progress is shown during the download. To install manually beforehand:

```bash
pip install axelera-devkit==1.6.0 --extra-index-url https://software.axelera.ai/artifactory/api/pypi/axelera-pypi/simple
pip install axelera-rt==1.6.0 --extra-index-url https://software.axelera.ai/artifactory/api/pypi/axelera-pypi/simple
```

Exporting YOLO Models to Axelera

Export your trained YOLO models using the standard Ultralytics export command.

!!! example "Export to Axelera Format"

=== "Python"

    ```python
    from ultralytics import YOLO

    # Load a YOLO26 model
    model = YOLO("yolo26n.pt")

    # Export to Axelera format
    model.export(format="axelera")  # creates 'yolo26n_axelera_model' directory
    ```

=== "CLI"

    ```bash
    yolo export model=yolo26n.pt format=axelera
    ```

!!! warning "First export may fail after dependency update"

The Axelera compiler requires `numpy<2`. If your environment has `numpy>=2`, the first `yolo export` will auto-downgrade it but the export will fail due to the stale module state. Simply run the same export command again — it will succeed on the second run.

Export Arguments

Argument	Type	Default	Description
`format`	`str`	`'axelera'`	Target format for Axelera Metis AIPU hardware.
`imgsz`	`int` or `tuple`	`640`	Image size for model input.
`batch`	`int`	`1`	Specifies export model batch inference size or the max number of images the exported model will process concurrently in `predict` mode.
`int8`	`bool`	`True`	Enable INT8 quantization for AIPU.
`data`	`str`	`'coco128.yaml'`	Dataset config for quantization calibration.
`fraction`	`float`	`1.0`	Fraction of dataset for calibration (100-400 images recommended).
`device`	`str`	`None`	Export device: GPU (`device=0`) or CPU (`device=cpu`).

For all export options, see the Export Mode documentation.

Output Structure

yolo26n_axelera_model/
├── yolo26n.axm              # Axelera model file
└── metadata.yaml            # Model metadata (classes, image size, etc.)

Running Inference

Load the exported model with the Ultralytics API and run inference, similar to loading ONNX models.

!!! example "Inference with Axelera Model"

=== "Python"

    ```python
    from ultralytics import YOLO

    # Load the exported Axelera model
    model = YOLO("yolo26n_axelera_model")

    # Run inference
    results = model("https://ultralytics.com/images/bus.jpg")

    # Process results
    for r in results:
        print(f"Detected {len(r.boxes)} objects")
        r.show()  # Display results
    ```

=== "CLI"

    ```bash
    yolo predict model='yolo26n_axelera_model' source='https://ultralytics.com/images/bus.jpg'
    ```

Axelera AI Benchmarks

The Metis AIPU maximizes throughput while minimizing energy consumption.

Model	Metis PCIe FPS (frames per second)	Metis M.2 FPS (frames per second)
YOLOv8n	847	771
YOLO11n	746	574
YOLO26n	648.6	484.9

Benchmarks based on Axelera AI data. Actual FPS depends on model size, batching, and input resolution.

Real-World Applications

Ultralytics YOLO on Axelera hardware enables advanced edge computing solutions:

Smart Retail: Real-time object counting and heatmap analytics for store optimization.
Industrial Safety: Low-latency PPE detection in manufacturing environments.
Drone Analytics: High-speed object detection on UAVs for agriculture and search-and-rescue.
Traffic Systems: Edge-based license plate recognition and speed estimation.

Recommended Workflow

Train your model using Ultralytics Train Mode
Export to Axelera format using model.export(format="axelera")
Validate accuracy with yolo val to verify minimal quantization loss
Predict using yolo predict for qualitative validation
Deploy to a high-performance end-to-end pipeline without PyTorch dependency — see the YOLO on Voyager SDK examples for composable Python pipelines using axelera-rt

Device Health Check

Verify your Axelera device is functioning properly:

bash

# if axdevice cannot be found, please run at least one inference (see above) to ensure the required packages are installed
axdevice

For detailed diagnostics, see the AxDevice documentation.

Maximum Performance

This integration uses single-core configuration for compatibility. For production requiring maximum throughput, the Axelera Voyager SDK offers:

Multi-core utilization (quad-core Metis AIPU)
Streaming inference pipelines
Tiled inferencing for higher-resolution cameras

See the model-zoo for FPS benchmarks or contact Axelera for production support.

Known Issues

!!! warning "Known Limitations"

- **M.2 power limitations**: Large or extra-large models may encounter runtime errors on M.2 accelerators due to power supply constraints.

For support, visit the Axelera Community.

FAQ

What YOLO versions are supported on Axelera?

The Voyager SDK supports export of YOLOv8, YOLO11, and YOLO26 models. See Supported Tasks for per-model task availability.

Can I deploy custom-trained models?

Yes. Any model trained using Ultralytics Train Mode can be exported to the Axelera format, provided it uses supported layers and operations.

How does INT8 quantization affect accuracy?

Axelera's Voyager SDK automatically quantizes models for the mixed-precision AIPU architecture. For most object detection tasks, the performance gains (higher FPS, lower power) significantly outweigh the minimal impact on mAP. Quantization takes seconds to several hours depending on model size. Run yolo val after export to verify accuracy.

How many calibration images should I use?

We recommend 100 to 400 images. More than 400 provides no additional benefit and increases quantization time. Experiment with 100, 200, and 400 images to find the optimal balance.

Where can I find the Voyager SDK?

The SDK, drivers, and compiler tools are available via the Axelera Developer Portal.