281 lines
14 KiB
Markdown
281 lines
14 KiB
Markdown
|
---
|
||
|
|
comments: true
|
||
|
|
description: Deploy Ultralytics YOLO models on Axelera AI's Metis hardware. Learn how to export, compile, and run high-performance edge inference with up to 856 TOPS.
|
||
|
|
keywords: Axelera AI, Metis AIPU, Voyager SDK, Edge AI, YOLOv8, YOLO26, Model Export, Computer Vision, PCIe, M.2, Object Detection, quantization
|
||
|
|
---
|
||
|
|
|
||
|
|
# Axelera AI Export and Deployment
|
||
|
|
|
||
|
|
!!! tip "Experimental Release"
|
||
|
|
|
||
|
|
This is an experimental integration demonstrating deployment on Axelera Metis hardware. Full integration anticipated by **February 2026** with model export without requiring Axelera hardware and standard pip installation.
|
||
|
|
|
||
|
|
Ultralytics partners with [Axelera AI](https://www.axelera.ai/) to enable high-performance, energy-efficient inference on [Edge AI](https://www.ultralytics.com/glossary/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](https://www.ultralytics.com/glossary/computer-vision-cv) at the edge, using a proprietary dataflow architecture and [in-memory computing](https://www.ultralytics.com/glossary/edge-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 <br> 30+ Streams| F[**Metis PCIe x4**<br>856 TOPS]
|
||
|
|
C -->|Standard PC <br> Low Profile| G[**Metis PCIe x1**<br>214 TOPS]
|
||
|
|
|
||
|
|
D -->|Drones & Handhelds| H[**Metis M.2**<br>2280 M-Key]
|
||
|
|
D -->|High Performance Embedded| I[**Metis M.2 MAX**<br>Extended Thermal]
|
||
|
|
|
||
|
|
E -->|ARM-based All-in-One| J[**Metis Compute Board**<br>RK3588 + AIPU]
|
||
|
|
E -->|Prototyping| K[**Arduino Portenta x8**<br>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](https://docs.ultralytics.com/models/yolo26/) and legacy versions with high FPS-per-watt efficiency.
|
||
|
|
|
||
|
|
### Accelerator Cards
|
||
|
|
|
||
|
|
These cards enable AI acceleration in existing host devices, facilitating [brownfield deployments](https://www.ultralytics.com/glossary/edge-computing).
|
||
|
|
|
||
|
|
| Product | Form Factor | Compute | Performance (INT8) | Target Application |
|
||
|
|
| :---------------- | :------------- | :----------------- | :----------------- | :----------------------------------------------------------------------------------------------------------------------------------------- |
|
||
|
|
| **Metis PCIe x4** | PCIe Gen3 x16 | **4x** Metis AIPUs | **856 TOPS** | High-density [video analytics](https://docs.ultralytics.com/guides/analytics/), smart cities |
|
||
|
|
| **Metis PCIe x1** | PCIe Gen3 x1 | **1x** Metis AIPU | **214 TOPS** | Industrial PCs, retail [queue management](https://docs.ultralytics.com/guides/queue-management/) |
|
||
|
|
| **Metis M.2** | M.2 2280 M-Key | **1x** Metis AIPU | **214 TOPS** | [Drones](https://www.ultralytics.com/blog/build-ai-powered-drone-applications-with-ultralytics-yolo11), 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](https://www.ultralytics.com/solutions/ai-in-manufacturing).
|
||
|
|
|
||
|
|
## Supported Tasks
|
||
|
|
|
||
|
|
Currently, Object Detection models can be exported to the Axelera format. Additional tasks are being integrated:
|
||
|
|
|
||
|
|
| Task | Status |
|
||
|
|
| :----------------------------------------------------------------- | :----------- |
|
||
|
|
| [Object Detection](https://docs.ultralytics.com/tasks/detect/) | ✅ Supported |
|
||
|
|
| [Pose Estimation](https://docs.ultralytics.com/tasks/pose/) | Coming soon |
|
||
|
|
| [Segmentation](https://docs.ultralytics.com/tasks/segment/) | Coming soon |
|
||
|
|
| [Oriented Bounding Boxes](https://docs.ultralytics.com/tasks/obb/) | Coming soon |
|
||
|
|
|
||
|
|
## 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**: Version 3.10 (3.11 and 3.12 coming soon)
|
||
|
|
|
||
|
|
### Ultralytics Installation
|
||
|
|
|
||
|
|
```bash
|
||
|
|
pip install ultralytics
|
||
|
|
```
|
||
|
|
|
||
|
|
For detailed instructions, see our [Ultralytics Installation guide](../quickstart.md). If you encounter difficulties, consult our [Common Issues guide](../guides/yolo-common-issues.md).
|
||
|
|
|
||
|
|
### Axelera Driver Installation
|
||
|
|
|
||
|
|
1. 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"
|
||
|
|
```
|
||
|
|
|
||
|
|
2. Add the repository to apt:
|
||
|
|
|
||
|
|
```bash
|
||
|
|
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"
|
||
|
|
```
|
||
|
|
|
||
|
|
3. Install the SDK and load the driver:
|
||
|
|
|
||
|
|
```bash
|
||
|
|
sudo apt update
|
||
|
|
sudo apt install -y axelera-voyager-sdk-base
|
||
|
|
sudo modprobe metis
|
||
|
|
yes | sudo /opt/axelera/sdk/latest/axelera_fix_groups.sh $USER
|
||
|
|
```
|
||
|
|
|
||
|
|
## 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
|
||
|
|
```
|
||
|
|
|
||
|
|
### 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 |
|
||
|
|
| `int8` | `bool` | `True` | Enable [INT8 quantization](https://www.ultralytics.com/glossary/model-quantization) for AIPU |
|
||
|
|
| `data` | `str` | `'coco128.yaml'` | [Dataset](https://docs.ultralytics.com/datasets/) 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](https://docs.ultralytics.com/modes/export/).
|
||
|
|
|
||
|
|
### Output Structure
|
||
|
|
|
||
|
|
```text
|
||
|
|
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](https://docs.ultralytics.com/integrations/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'
|
||
|
|
```
|
||
|
|
|
||
|
|
!!! warning "Known Issue"
|
||
|
|
|
||
|
|
The first inference run may throw an `ImportError`. Subsequent runs will work correctly. This will be addressed in a future release.
|
||
|
|
|
||
|
|
## Inference Performance
|
||
|
|
|
||
|
|
The Metis AIPU maximizes throughput while minimizing energy consumption.
|
||
|
|
|
||
|
|
| Metric | Metis PCIe x4 | Metis M.2 | Note |
|
||
|
|
| :------------------ | :------------ | :----------- | :---------------------- |
|
||
|
|
| **Peak Throughput** | **856 TOPS** | 214 TOPS | INT8 Precision |
|
||
|
|
| **YOLOv5m FPS** | **~1539 FPS** | ~326 FPS | 640x640 Input |
|
||
|
|
| **YOLOv5s FPS** | N/A | **~827 FPS** | 640x640 Input |
|
||
|
|
| **Efficiency** | High | Very High | Ideal for battery power |
|
||
|
|
|
||
|
|
_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](https://docs.ultralytics.com/guides/object-counting/) and [heatmap analytics](https://docs.ultralytics.com/guides/heatmaps/) for store optimization.
|
||
|
|
- **Industrial Safety**: Low-latency [PPE detection](https://docs.ultralytics.com/datasets/detect/construction-ppe/) in manufacturing environments.
|
||
|
|
- **Drone Analytics**: High-speed [object detection](https://docs.ultralytics.com/tasks/detect/) on UAVs for [agriculture](https://www.ultralytics.com/solutions/ai-in-agriculture) and search-and-rescue.
|
||
|
|
- **Traffic Systems**: Edge-based [license plate recognition](https://www.ultralytics.com/blog/using-ultralytics-yolo11-for-automatic-number-plate-recognition) and [speed estimation](https://docs.ultralytics.com/guides/speed-estimation/).
|
||
|
|
|
||
|
|
## Recommended Workflow
|
||
|
|
|
||
|
|
1. **Train** your model using Ultralytics [Train Mode](https://docs.ultralytics.com/modes/train/)
|
||
|
|
2. **Export** to Axelera format using `model.export(format="axelera")`
|
||
|
|
3. **Validate** accuracy with `yolo val` to verify minimal quantization loss
|
||
|
|
4. **Predict** using `yolo predict` for qualitative validation
|
||
|
|
|
||
|
|
## Device Health Check
|
||
|
|
|
||
|
|
Verify your Axelera device is functioning properly:
|
||
|
|
|
||
|
|
```bash
|
||
|
|
. /opt/axelera/sdk/latest/axelera_activate.sh
|
||
|
|
axdevice
|
||
|
|
```
|
||
|
|
|
||
|
|
For detailed diagnostics, see the [AxDevice documentation](https://github.com/axelera-ai-hub/voyager-sdk/blob/release/v1.5/docs/reference/axdevice.md).
|
||
|
|
|
||
|
|
## Maximum Performance
|
||
|
|
|
||
|
|
This integration uses single-core configuration for compatibility. For production requiring maximum throughput, the [Axelera Voyager SDK](https://github.com/axelera-ai-hub/voyager-sdk) offers:
|
||
|
|
|
||
|
|
- Multi-core utilization (quad-core Metis AIPU)
|
||
|
|
- Streaming inference pipelines
|
||
|
|
- Tiled inferencing for higher-resolution cameras
|
||
|
|
|
||
|
|
See the [model-zoo](https://github.com/axelera-ai-hub/voyager-sdk/blob/release/v1.5/docs/reference/model_zoo.md) for FPS benchmarks or [contact Axelera](https://axelera.ai/contact-us) for production support.
|
||
|
|
|
||
|
|
## Known Issues
|
||
|
|
|
||
|
|
!!! warning "Known Limitations"
|
||
|
|
|
||
|
|
- **PyTorch 2.9 compatibility**: The first `yolo export format=axelera` command may fail due to automatic PyTorch downgrade to 2.8. Run the command a second time to succeed.
|
||
|
|
|
||
|
|
- **M.2 power limitations**: Large or extra-large models may encounter runtime errors on M.2 accelerators due to power supply constraints.
|
||
|
|
|
||
|
|
- **First inference ImportError**: The first inference run may throw an `ImportError`. Subsequent runs work correctly.
|
||
|
|
|
||
|
|
For support, visit the [Axelera Community](https://community.axelera.ai/).
|
||
|
|
|
||
|
|
## FAQ
|
||
|
|
|
||
|
|
### What YOLO versions are supported on Axelera?
|
||
|
|
|
||
|
|
The Voyager SDK supports export of [YOLOv8](https://docs.ultralytics.com/models/yolov8/) and [YOLO26](https://docs.ultralytics.com/models/yolo26/) models.
|
||
|
|
|
||
|
|
### Can I deploy custom-trained models?
|
||
|
|
|
||
|
|
Yes. Any model trained using [Ultralytics Train Mode](https://docs.ultralytics.com/modes/train/) 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](https://www.ultralytics.com/glossary/object-detection) tasks, the performance gains (higher FPS, lower power) significantly outweigh the minimal impact on [mAP](https://docs.ultralytics.com/guides/yolo-performance-metrics/). 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](https://www.axelera.ai/).
|