Exascale Computing on 8000 GPUs

Exascale Computing

8000 GPUs on LUMI supercomputer with fractal decomposition for Bayesian optimization

Pushing the boundaries of optimization at scale with successful deployment on 8,000 GPUs of the LUMI supercomputer, demonstrating true exascale readiness.

1. Overview

Our Bayesian optimization framework has been successfully deployed on LUMI (Finland), one of Europe’s most powerful supercomputers, achieving unprecedented scale for uncertainty quantification and optimization workflows.

2. Key Achievement

8,000 GPU Scale

LUMI Supercomputer (Finland)
Fractal domain decomposition
Bayesian optimization workflows

3. Technical Details

3.1. Computing Infrastructure

  • System: LUMI (Large Unified Modern Infrastructure)

  • Location: CSC Data Center, Kajaani, Finland

  • Resources Utilized: 8,000 AMD MI250X GPUs

  • Peak Performance: Multi-petaflop range for optimization campaigns

3.2. Algorithmic Innovation

  • Method: Fractal domain decomposition for parallel Bayesian optimization

  • Scalability: Near-linear weak scaling to thousands of GPUs

  • Efficiency: Load balancing across heterogeneous search spaces

  • Integration: Coupled with WP1 discretizations and WP3 solvers

4. Fractal Decomposition Strategy

The fractal approach enables:

  1. Hierarchical Partitioning: Recursive subdivision of parameter space

  2. Adaptive Refinement: Focus computational effort on promising regions

  3. Asynchronous Execution: Minimize idle time across GPU nodes

  4. Memory Efficiency: Distributed model storage and evaluation

5. Impact on Scientific Computing

This achievement demonstrates:

  • Exascale Readiness: Methods proven at realistic exascale dimensions

  • Uncertainty Quantification: UQ now feasible for complex, expensive models

  • Design Optimization: Thousands of design parameters simultaneously explored

  • EuroHPC Integration: Workflows validated on production European infrastructure

6. Application Domains

Successfully applied to:

  • Material Science: High-dimensional parameter searches for material properties

  • Climate Modeling: Ensemble-based uncertainty quantification

  • Engineering Design: Multi-objective optimization of complex systems

  • Drug Discovery: Molecular configuration space exploration

7. Related Work Packages

8. Software & Resources

  • Open-source Bayesian optimization libraries

  • LUMI deployment recipes and tutorials

  • Integration with EuroHPC software stack

  • Performance benchmarking dashboards

← Previous: Speedups | Back to Results | Next: Neural Operators →