Copyright © Unsplash / Christian Lue
Project profile
Turkey, Norway, Portugal, Germany
01/04/2022 - 31/03/2024
Motivation and Goals
- Identify optimisation potentials: characterise applications & data by analytical and machine-learning performance and energy models
- Increase performance via advanced node-level static and dynamic code optimisations
- Increase performance via advanced system-level, topology-aware code optimisations
- Create a “digital twin” for supercomputers and evaluate what-if scenarios
- Demonstrate SparCity results with four challenging real-life applications
- Deliver a robust, well-supported and documented SW development framework
Technical Approach
- Inspection
- Feature extraction
- ML-base recommendation systems
- Performance & energy models
- Node-level optimisations
- Mixed-precision support
- Data locality optimisation
- Data and computation reordering
- System-level optimisations
- Memory constrained partitioning
- Optimising ML loads on emerging hardware
- Topology mapping
- Communication optimisation
- Digital Supercomputer Twin
- Automated HW topology discovery
- Micro-benchmarks
- Metric collection, performance evaluation, ML
- Demonstrators
- Four real-world applications
- A complete framework
- Software tools and open data repositories
Copyright © SparCity
Status / Highlights
- 11 high-impact publications and 14 presentations in relevant
meetings, conferences and workshops - MLOpt 2023 Workshop, co-located with HiPEAC 2023
Project coordinator
