A Sovereign AI Stack for Portable European Cloud Services
| Reference number | |
| Coordinator | RISE Research Institutes of Sweden AB - RISE AB - Digitala System |
| Funding from Vinnova | SEK 1 997 349 |
| Project duration | November 2023 - May 2025 |
| Status | Completed |
| Venture | Advanced digitalization - Enabling technologies |
| Call | AI for advanced digitalization, 2 |
Important results from the project
The project reached its goals by demonstrating a dozen state-of-the-art LLM models and state-of-the-art AI frameworks running on at least two state-of-the-art AI accelerators -- all from a single stack. The project explored heterogeneity demonstrating the stack on both x86 and ARM, NVIDIA and AMD. The testbed ran over 10 000 GPU hours evaluating AI accelerators, AI frameworks, and AI stack to address performance gaps between hardware vendors to create a fully hardware-agnostic software stack.
Expected long term effects
The ability to deploy AI accelerators from varying vendors is important to secure sovereignty and robustness against supply chain disruptions. Over time this ability secures access to the most profitable and performant alternatives. As strategic investment in European chip production scale up, where AI chips can be expected to play a major role, the ability to deploy AI on emerging accelerators becomes both a competence in high-demand and a market advantage.
Approach and implementation
The project ran in four phases over 19 months. The first phase addressed procurement and deployment and the configuration of an initial testbed to determine the most promising upstream components. The second phase consisted of initial testing and deep-dives in efficient AI operations across the accelerators. The third phase of 10+ months hosted the bulk of AI software stack development, incorporating the lessons from the first two phases. The fourth phase consisted of benchmarking and reporting.