Event language
UI language
As AI becomes a core requirement for modern desktop and edge systems, operating systems are evolving from passive application hosts into active providers of secure, low-latency, and resource-aware AI services. Drawing on practical experience with the Debian/Ubuntu AI subsystem and insights from upstream ecosystem developments, this talk presents an architectural blueprint for a truly native AI subsystem on Linux.<br><br>The session will introduce <strong>a complete AI subsystem technology stack</strong>, including:<br><ul><li><strong>Southbound Layer</strong>: Unified abstraction and scheduling for heterogeneous compute units (CPU/GPU/NPU), model lifecycle management, and a decoupled runtime execution framework.</li><li><strong>Core Runtime Layer</strong>: Capability governance, resource arbitration, and global AI capability discovery mechanisms.</li><li><strong>Northbound Layer</strong>: Standardized APIs and developer toolkits that expose AI capabilities—such as speech recognition, OCR, vector retrieval, and multimodal inference—to applications.</li><li><strong>Security & Privacy Layer</strong>: Local-first execution, trusted model execution boundaries, and support for collaborative cloud-edge computation.</li></ul><br><strong>Attendees will learn:</strong><br><ul><li>Practical architecture and component layout for building a Linux-native AI subsystem from the ground up</li><li>Strategies for model management, heterogeneous compute scheduling, and hybrid local-cloud inference</li><li>Hands-on lessons in performance optimization, extensibility design, privacy protection, and long-term maintainability</li></ul><br><strong>Target Audience</strong><br><ul><li>Operating system engineers</li><li>Desktop and edge platform developers</li><li>AI runtime and framework contributors</li><li>Open-source community members interested in integrating native AI capabilities into Linux distributions</li></ul>
