TOUBKAL: a high-performance supercomputer powering scientific research in Africa
Imad Kissami, Robert Basmadjian, Othmane Chakir, Mohamed Riduan Abid
Abstract
Abstract This paper presents the architecture, performance characterization, and early scientific impact of Toubkal, Africa’s most powerful supercomputer (TOP500 #356, Green500 #178 as of June 2025). Designed to support high-performance computing research across disciplines, Toubkal integrates CPU, high-memory, and GPU-accelerated nodes connected via a high-bandwidth InfiniBand fabric and powered in part by renewable energy sources. We benchmark Toubkal’s compute, memory, network, and storage subsystems using standard and reproducible tools (HPL, STREAM, IOR, OMB, HPL-MxP), reporting robust performance and energy-efficiency metrics. The CPU partition achieves 63% of its theoretical peak, while GPU-accelerated HPL-MxP benchmarks on H100 nodes exceed 209 TFLOP/s per GPU. We also introduce a supervised learning model that predicts GPU frequency ranges minimizing the energy-delay product (EDP) across diverse workloads. The model achieves over 92% accuracy, with average gains of 13% in EDP, 13% in energy consumption, and a 1.1% reduction in kernel time across all kernels. The paper also compares Toubkal to several US DoE systems, analyzes early user growth, and documents real-world adoption across domains including AI, bioinformatics, and material science. These results offer both a technical reference for reproducibility and a performance baseline for emerging HPC centers worldwide. Beyond presenting a system deployment, this work contributes an early blueprint for sustainable, regional supercomputing infrastructure aligned with global standards.