What It Actually Takes to Run Code on 200M€ Supercomputer

2026-04-16 · Source: Towards Data Science · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure, Data Science & Analytics · Depth: Intermediate, quick

Summary

MareNostrum V, one of the world's fifteen most powerful supercomputers, located at the Polytechnic University of Catalonia, offers a unique High-Performance Computing (HPC) environment. This 200M€ machine, a joint investment from EuroHPC, Spain, Portugal, and Turkey, features a General Purpose Partition with 6,408 Intel Sapphire Rapids nodes (45.9 PFlops) and an Accelerated Partition with 1,120 NVIDIA H100 SXM GPU nodes (260 PFlops). Its architecture utilizes an InfiniBand NDR200 fat-tree topology to ensure non-blocking, low-latency communication across 8,000 nodes. MareNostrum V also integrates quantum infrastructure, including a digital gate-based system and a MareNostrum-Ona quantum annealer, for hybrid classical-quantum computing. Access is managed via SLURM, a workload manager that queues and executes jobs, enforcing strict resource limits and an airgapped network for compute nodes.

Key takeaway

For data scientists or ML engineers considering large-scale simulations or model training, understanding HPC architecture and operational rules is crucial. You must adapt to an airgapped environment, strict job scheduling via SLURM, and resource quotas. Plan your workflows to pre-compile code and manage data transfer, recognizing that communication overhead and Amdahl's Law will dictate optimal parallelism, rather than simply maximizing core count.

Key insights

Supercomputers like MareNostrum V are distributed systems requiring specialized architectural and operational understanding.

Principles

• The network is the computer in distributed systems.
• Amdahl's Law limits parallel speedup due to serial fractions.
• HPC environments enforce strict resource management and security.

Method

Interact with supercomputers via SSH to login nodes, use SLURM to define and submit job scripts with resource directives like `--nodes`, `--ntasks`, and `--time`, and manage data via `scp` or `rsync`.

In practice

• Use `module` system for pre-installed libraries.
• Chain SLURM jobs with dependencies for automated pipelines.
• Monitor jobs with `squeue` or `tail -f` on log files.

Topics

• MareNostrum V
• Supercomputer Architecture
• HPC Job Scheduling
• SLURM Workload Manager
• Quantum Computing Integration

Best for: Data Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Machine Learning Model Development on HPC Clusters with SLURM: From Basics to Production — HPC clusters with SLURM provide scalable resource management for production-grade machine learning model development.
• Running AI Workloads on Rack-Scale Supercomputers: From Hardware to Topology-Aware Scheduling — NVIDIA Mission Control and related tools bridge hardware topology with scheduler abstractions for optimal AI/HPC workload placement.
• CodegenBench: Can LLMs Write Efficient Code Across Architectures? — LLMs generate efficient code for common architectures but struggle with specialized HPC platforms due to data scarcity.
• A Guide to Understanding GPUs and Maximizing GPU Utilization — GPU bottlenecks often stem from inefficient data pipelines, not compute, requiring optimized data transfer and processing.
• Here’s how you can secure access to the UK’s most powerful supercomputer — Access to powerful supercomputing is critical for developing advanced AI models and maintaining competitive advantage.
• HPC Coding Agent - Part 3: MCP Tool for Profiling — An AI agent can autonomously profile and optimize GPU-accelerated HPC applications using a custom MCP server for tool integration.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards Data Science.