Part 1 | Part 2
Combining performance and energy efficiency
Eco-design, reduced plastic packaging, energy efficiency, recycling… environmental impact has become a central criterion in the design of supercomputers. At a time when demand for high-performance computing continues to grow, the challenge is now to reconcile computing power, energy efficiency and responsible resource management.
The Green500 ranking provides a precise measure of energy performance, based on the ratio between delivered computing power and electricity consumption. In the June 2026 edition, Bull-designed supercomputers reaffirmed their leadership by securing the top three positions in the ranking for the second consecutive edition.
These results reflect a broader approach: managing environmental footprint is an integral part of Bull’s DNA. This commitment translates into concrete actions at every stage of the lifecycle, from design and manufacturing to the optimisation of operational processes. The objective is to deliver computing infrastructures that are both high-performing and responsible, supporting critical applications such as nuclear simulation, weather forecasting and fundamental research in healthcare.
Energy efficiency: faster computing with lower consumption
Optimising energy efficiency first relies on reducing the overall consumption of computing systems. In this respect, edge computing is a key lever: by processing data as close as possible to its source, it limits the volumes that need to be transferred and ensures that only already-processed data is transmitted, thereby reducing both computing and energy requirements. At the same time, continuous improvements in algorithms play a crucial role by decreasing the number of required operations while accelerating execution times.
System cooling, another major contributor to energy consumption, can also be optimised through less energy-intensive technologies such as Direct Liquid Cooling (DLC) Direct Liquid Cooling (DLC) is a technology that cools servers by circulating liquid directly to the hottest components, enabling heat to be captured and dissipated far more efficiently than with conventional air cooling. using warm glycol water. By capturing heat as close as possible to the components, this approach makes it possible to achieve a Power Usage Effectiveness (PUE) Power Usage Effectiveness (PUE) is a standard metric for assessing data center energy efficiency. It compares the total energy consumed by the facility to the energy used by its IT equipment. A PUE value closer to 1 indicates a highly efficient infrastructure, where most of the energy consumed is used directly for computing. close to neutrality, meaning as near as possible to 1.0, or even below when heat is recovered and reused in heating systems.
Bull began developing liquid cooling technologies as early as 2007, with the launch of its first Direct Liquid Cooling (DLC) system in 2012. Now fully mastered, this approach forms a cornerstone of Bull’s energy strategy. The latest generation BullSequana XH3500 systems operate entirely using our DLC technology, ensuring efficient, stable and energy-saving cooling, perfectly suited to the highest computing densities.
Bull’s Direct Liquid Cooling operates using warm glycol water in a closed loop, with no water loss. Each compute blade requires only the equivalent of a glass of water, while enabling heat recovery efficiency of up to 100%. This approach helps reduce energy consumption and carbon emissions by around 10%, while significantly limiting water usage compared with more resource-intensive cooling solutions.
This thermal recovery efficiency also makes it possible to directly reuse heat for building heating. At Bull’s industrial site in Angers, waste heat will be fed into the district heating network, covering the equivalent of 800 households per year by 2027.
Improving energy consumption by aligning power with usage
At the scale of a supercomputer, energy consumption does not depend solely on the installed hardware; it is primarily driven by how resources are used by workloads. Without a detailed understanding of application behaviour, data flows and the interactions between infrastructure and usage, a significant share of energy consumption can be inefficient or even unnecessary. Managing energy consumption therefore requires going beyond overall metrics to understand where, when and why energy is being used.
This is precisely the approach behind BullSequana ARGOS. By providing continuous, granular visibility into how resources are actually used, the solution turns data into a decision-making asset. It highlights gaps between resources consumed and value delivered, helps identify inefficient usage, and provides an objective basis to balance performance, time-to-results and energy consumption. By aligning resource usage more closely with the real needs of workloads, BullSequana ARGOS enables a more efficient and intelligent use of supercomputing resources, without disrupting existing applications or expected performance.
Designed to support a wide range of stakeholders (operators, scientific teams and decision-makers) the software fosters continuous improvements in energy efficiency, contributing to more controlled, sustainable and results-driven computing.
Optimising code: a key lever to reduce compute time
While the energy performance of a supercomputer partly depends on its infrastructure, it is equally shaped by how applications make use of available resources. Code optimisation is therefore a key lever to run scientific applications efficiently, making the best possible use of computing resources while reducing execution time and energy consumption.
This is precisely the focus of the initiatives led by Bull’s CEPP (Centre of Excellence for Performance and Programming), a team of experts dedicated to code optimisation, porting and diagnostics across all computing platforms.
Bull’s CEPP supports organisations in improving their applications, optimising simulations, adopting new hardware and accelerating time-to-results. In collaboration with leading institutions and partners, it contributes to a more efficient and resource-conscious use of advanced HPC systems, enhancing performance while reducing their energy footprint.
Code optimisation delivers tangible and measurable results. CEPP teams at Bull have notably contributed to a major international collaboration around the large-scale plasma simulation application WarpX, which was awarded the Gordon Bell Prize 2022 by the Association for Computing Machinery (ACM) The Gordon Bell Prize is a prestigious award presented by the Association for Computing Machinery to recognize outstanding achievements in supercomputing and high-performance computing applications. , recognising outstanding achievements in high-performance computing.
This optimisation work on a particularly demanding physical simulation code enabled exceptional performance on the first exascale supercomputer. The resulting acceleration reduced overall execution time by a factor of four, significantly shortening simulation durations and lowering associated energy consumption.
This achievement represents the first exascale demonstration of a Mesh-Refined Particle-In-Cell (MR-PIC) implementation MR-PIC combines particle-based simulation with locally refined meshes to achieve higher accuracy where needed without significantly increasing computational effort. , combining particles and adaptive mesh refinement to optimise both performance and accuracy.
Overall, this work demonstrates that high performance and scientific excellence can go hand in hand, making it possible to maximise the use of computing resources while reducing the energy footprint of simulations on exascale architectures.
While the optimisation levers presented demonstrate that it is possible to significantly reduce impact during the operational phase, the environmental performance of a supercomputer can only be fully assessed across its entire lifecycle.
From the supply chain and material choices through to end-of-life management, the supply chain plays a decisive role in this equation. It is precisely this broader, and often less visible, yet equally structuring dimension that we explore in the next part of this article.
