Key points are not available for this paper at this time.
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.
Kissami et al. (Fri,) studied this question.