Axelera AI delivers 34 FPS edge AI inference using Ultralytics YOLO

Axelera AI is a European semiconductor company that builds powerful, energy-efficient AI chips for running computer vision at the edge. For instance, their Metis AI Processing Units (AIPUs) are used in industries like retail, security, and manufacturing.

To make it easier for customers to build and deploy AI applications, Axelera integrated Ultralytics YOLO models into their Voyager SDK, a software platform that streamlines model optimization, deployment, and acceleration on Metis AIPUs, making real-time Vision AI faster, simpler, and more scalable.

Link to this sectionRedefining Vision AI hardware acceleration#

Founded in 2021 and based in the Netherlands, Axelera AI set out to solve a fundamental problem: traditional AI hardware was built for the cloud, not the edge. To close this gap, the company developed the Metis AIPU.

It’s a high-performance, low-power chip designed specifically for edge workloads - AI tasks that run locally on devices where speed, privacy, and energy efficiency are critical.

It’s backed by their proprietary Digital In-Memory Compute (D-IMC) technology, which makes it possible for data to be processed directly where it’s stored and dramatically reduces latency and energy use. With a fully integrated software stack and a mission to democratize AI, Axelera is making high-performance AI more accessible.

Link to this sectionThe challenge of running models on edge AI hardware#

To deliver a seamless AI experience at the edge, Axelera AI aimed to provide more than high-performance hardware. Customers also required a simple and reliable way to run real-time computer vision solutions that didn’t involve complicated tools or time-consuming customization. Many existing models were too large, too slow, or not well-suited for resource-constrained environments.

Retail analytics, factory automation, and surveillance systems often rely on fast and accurate visual insights to support operations. However, traditional models and cloud-based solutions typically can’t meet the low latency, energy efficiency, and on-device processing requirements of these edge AI applications.

Axelera began looking for a model suite that was accurate, easy to implement, and able to run efficiently on its Metis AI Processing Units. The right solution would complement its hardware capabilities while simplifying development and speeding up deployment across a broad range of edge AI use cases.

Link to this sectionAccelerated AI inferences at the edge with Ultralytics YOLO#

Axelera AI integrated Ultralytics YOLO models into its Voyager SDK to make building and deploying edge AI applications faster, easier, and more scalable. The SDK includes a Model Zoo with ready-to-use YOLO models and automates the full pipeline, covering preprocessing, inference, and post-processing, optimized for Metis AI Processing Units.

Fig 1. Metis AI Processing Unit (AIPU).

This integration lets customers use pre-trained Ultralytics YOLO models out of the box or bring their own, with seamless acceleration on edge hardware. With support for parallel and cascaded model execution, Metis AIPUs enable advanced multi-model setups like pose estimation followed by segmentation. This is ideal for complex tasks in retail, security, and industrial automation.

The combination of real-time computer vision tasks supported by Ultralytics YOLO models and Axelera’s efficient hardware and software stack delivers exceptional performance per watt and per dollar. For customers, this means more accurate results, faster deployment, and a lower barrier to scaling Vision AI at the edge.

Link to this sectionWhy choose Ultralytics YOLO models?#

Axelera AI partnered with Ultralytics to integrate Ultralytics YOLO models into its platform for their exceptional balance of accuracy, speed, and ease of use. With support for multiple Ultralytics YOLO variants, Axelera AI customers can evaluate a wide range of workloads and performance needs on the Metis AIPU.

Through an Ultralytics Enterprise License, Axelera provides access to the full suite of YOLO models for evaluation and development. For commercial deployment, customers are required to obtain their own license directly from Ultralytics via the license form, ensuring compliance and supporting scalable innovation in Vision AI at the edge.

Link to this sectionLeveraging Ultralytics YOLO models at scale with Axelera AI’s Voyager SDK#

With Ultralytics YOLO models and Axelera AI’s Voyager SDK, users can deploy accurate, low-latency computer vision applications directly on Metis AI Processing Units. Also, having access to multiple YOLO variants allows customers to adjust performance based on the specific needs of their application.

For example, Axelera AI has seen customers test YOLO-powered solutions across a variety of domains like:

• Retail: Multi-model pipelines combining pose estimation, segmentation, and object detection can be used to support tasks like shrink reduction and inventory tracking. These deployments run efficiently on embedded platforms like the Raspberry Pi 5 paired with the Metis AIPU.
• Manufacturing: Customers can use parallel model execution to perform simultaneous tasks, such as defect detection, product classification, and pose estimation, on a single chip to improve throughput and reduce hardware costs.
• Surveillance: YOLO’s real-time object detection capabilities can be used to analyze 4K and 8K video streams at full resolution, facilitating high-accuracy safety monitoring and situational awareness in critical environments. This is a much better solution than traditional downscaling on high-resolution cameras.
• Healthcare: Optimized YOLO pipelines can support tumor identification, offering high inference speed and reliable accuracy for on-device medical imaging.

These use cases showcase how Axelera AI’s edge AI hardware and Ultralytics YOLO models empower high-performance, energy-efficient Vision AI innovations across industries.

Link to this sectionDriving the next wave of edge AI with Ultralytics YOLO#

As Axelera AI continues to focus on expanding access to high-performance edge AI, it’s bringing together powerful hardware and reliable vision models to help customers build smarter, faster applications.

With Ultralytics YOLO models available through the Voyager SDK and Metis AIPU hardware, users can easily develop and scale real-time computer vision solutions across industries. This collaboration supports a growing community of developers and businesses working to bring AI closer to where data is created, improving efficiency, responsiveness, and innovation at the edge.

Ready to accelerate your edge AI journey? Explore our GitHub repository to see how YOLO models are powering innovations like AI in retail and computer vision in logistics. Get hands-on with our Vision AI tools, learn about licensing options, and discover how you can unlock high-performance, energy-efficient computer vision at the edge.