Axelera AI Acceleration
Coming soon — Q1 2026
Axelera support in ultralytics is in progress. The examples here show the planned UI/UX and will become runnable once the Axelera runtime package is released.
Ultralytics partners with Axelera AI to streamline high-performance, energy-efficient inference on Edge AI devices. This integration allows users to export and deploy Ultralytics YOLO models directly to the Metis® AIPU and Europa® platforms using the Voyager SDK.
Axelera AI provides dedicated hardware acceleration for computer vision and Generative AI at the edge. Their technology leverages a proprietary dataflow architecture and in-memory computing to deliver high throughput (up to 856 TOPS) within a low power envelope.
For Ultralytics users, this offers a scalable path to deploy object detection, pose estimation, and other YOLO tasks on devices ranging from embedded drones to edge servers.
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.
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 YOLO11 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.
Voyager SDK Integration
The Voyager SDK serves as the bridge between Ultralytics models and Axelera hardware. It handles the compilation, quantization, and runtime execution of neural networks.
Key features for Ultralytics users:
- Seamless Export: The SDK's compiler optimizes YOLO models for the Metis dataflow architecture.
- Quantization Engine: Automatically converts FP32 models to INT8 precision with minimal accuracy loss.
- Pipeline Builder: A YAML-based framework to chain multiple models (e.g., detection + pose estimation) without writing complex C++ code.
Installation & Setup
To use Axelera acceleration, you need the ultralytics package installed. Note that the Voyager SDK is a separate system-level installation required to interface with the hardware. Runtime wheels are expected in Q1 2026; the commands below reflect the intended setup flow.
# Install Ultralytics
pip install ultralytics
# Note: Download and install the Axelera Voyager SDK from the Axelera Developer Portal
# to enable the 'axelera' export format and runtime.
Exporting YOLO Models to Axelera
When the Axelera runtime package ships (target Q1 2026), you will export your trained YOLO models to the Axelera format using the standard Ultralytics export command. This process generates the artifacts required by the Voyager compiler.
Voyager SDK Required
The format='axelera' export requires the Axelera libraries to be available in your environment. Alternatively, you can export to ONNX and manually compile using the Voyager toolchain.
Export Examples
Convert a YOLO11 model for Metis deployment.
Export to Axelera Format
Future example — will work when runtime is released
This code block demonstrates the planned flow. It will require the upcoming Axelera runtime package (ETA Q1 2026) to execute successfully.
from ultralytics import YOLO
# Load a standard or custom trained YOLO11 model
model = YOLO("yolo11n.pt")
# Export to Axelera format
# int8=True enables quantization for the NPU
model.export(format="axelera", int8=True, imgsz=640)
# Export a model via CLI
yolo export model=yolo11n.pt format=axelera int8=True imgsz=640
For available arguments, refer to the Export Mode documentation.
Running Inference
Once exported, you will be able to load the Axelera-compiled model directly with the ultralytics API (similar to loading ONNX models). The example below shows the expected usage pattern for running inference and saving results after the runtime package ships.
Inference with Axelera Format
Future example — will work when runtime is released
This code block demonstrates the planned flow. It will require the upcoming Axelera runtime package (ETA Q1 2026) to execute successfully.
from ultralytics import YOLO
# Load the Axelera-compiled model (example path; same flow as ONNX)
model = YOLO("yolo11n_axelera.axmodel") # will work once Axelera runtime is released
# Run inference; you can pass a file, folder, glob, or list of sources
results = model("path/to/images", imgsz=640, save=True)
# Iterate over result objects to inspect or render detections
for r in results:
boxes = r.boxes # bounding boxes tensor + metadata
print(f"Detected {len(boxes)} objects")
# Save visuals per result (files saved alongside inputs)
r.save() # saves annotated image(s) to disk
# Or display interactively (desktop environments)
# r.show()
Inference Performance
The Metis AIPU is designed to maximize throughput while minimizing energy consumption. The benchmarks below illustrate the performance achievable with standard Ultralytics models.
| 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 (Sept 2025). 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.
FAQ
What YOLO versions are supported on Axelera?
The Voyager SDK and Ultralytics integration support the export of YOLOv8 and YOLO11 models.
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 quantization engine uses advanced calibration techniques to minimize accuracy drop. For most detection tasks, the performance gain significantly outweighs the negligible impact on mAP.
Where can I find the Voyager SDK?
The SDK, drivers, and compiler tools are available via the Axelera Developer Portal.
