European AI Factories increasingly need to support both specialized AI accelerators and traditional GPU resources. OpenNebula now offers this flexibility by making Axelera’s Metis AI Processing Units available as on-demand infrastructure devices for supporting AI inference.

Using virtualization with PCI passthrough, OpenNebula enables operators to securely launch AI inference tasks on Axelera hardware, while allowing these European accelerators to coexist with NVIDIA GPUs within the same compute nodes.

Managing Accelerated Hardware Resources

OpenNebula provides HPC and cloud administrators with full visibility into the available accelerator resources across their AI infrastructure. 

Individual devices in each node can then be automatically or manually assigned to virtual machines using PCI passthrough. This allows administrators to allocate dedicated acceleration resources to specific AI workloads while maintaining clear visibility into resource consumption.

Secure Multi-Tenant AI Infrastructure

OpenNebula enables administrators and users to interact with the AI infrastructure through clearly separated roles. Administrators can inspect available accelerator devices and prepare virtual machine templates with PCI passthrough already configured, while users simply instantiate preconfigured templates to deploy AI-enabled virtual machines. Once the virtual machine has been provisioned, the assigned Axelera accelerator becomes immediately available inside the guest operating system.

This workflow allows AI accelerator resources to be consumed on-demand under a self-service provisioning model without exposing the underlying infrastructure, maintaining secure isolation between tenants and simplifying day-to-day operations.

End-to-End Demonstration

The accompanying screencast displays the complete workflow, from inspecting available accelerator resources in OpenNebula to deploying a virtual machine and validating an AI inference workload on an Axelera Metis accelerator. In the demonstration, an object detection model precompiled specifically for the Axelera platform processes a high-resolution traffic video, illustrating how European AI accelerators can efficiently execute inference workloads within a virtualized AI Factory environment.

Enabling Sovereign AI Infrastructure

OpenNebula enables organizations to integrate Axelera AI accelerators into Sovereign AI Factories through virtualization, PCI passthrough, and multi-tenant infrastructure management. The result is a flexible open source platform for securely deploying AI inference workloads on European accelerated hardware. 

Interested in learning more? Explore the AI Factory landing page to discover how OpenNebula enables sovereign AI infrastructure, or browse our technical screencast collection to see additional AI Factory deployment scenarios in action.