The Alice Recoque supercomputer represents a significant strategic investment by the European High Performance Computing Joint Undertaking (EuroHPC JU) and the Jules Verne consortium, designated as France”s first and Europe”s second system to achieve the Exascale performance milestone. Explicitly engineered as a purpose-built AI Factory, Alice Recoque is fundamentally designed to accelerate the convergence of High-Performance Computing (HPC) simulations with Artificial Intelligence (AI) workloads, positioning Europe to address critical societal and industrial challenges while establishing technological sovereignty in advanced computing infrastructure.
What”s happening: Alice Recoque, built by Eviden (an Atos Group brand) and powered by Advanced Micro Devices (AMD), guarantees a sustained performance exceeding 1 Exaflop/s (one billion billion calculations per second) on the High-Performance Linpack (HPL) benchmark, achieving France”s first Exascale milestone and Europe”s second following the JUPITER system in Germany. The system”s technological differentiator lies in the adoption of the next-generation AMD Instinct MI430X GPU, which features a leadership capacity of 432 GB of HBM4 memory and 19.6 TB/s of bandwidth per chip—substantial memory capacity that serves as the cornerstone of its AI mission, enabling efficient training of next-generation European generative AI and foundational models. Financially supported by a total cost of €554 million over five years of operation, the project strongly emphasizes energy efficiency, targeting operation at less than 15 MW of electrical power.
Why it matters: Alice Recoque represents a strategic imperative for European technological sovereignty, addressing the critical need for indigenous AI model development in an era where control over foundational digital assets influences information flow, economic competitiveness, and defense capabilities. The system”s specialization in memory-intensive AI workloads—enabled by the MI430X”s 432 GB HBM4 capacity, significantly surpassing the 192 GB offered by previous architectures—positions Europe to reduce dependence on externally trained, often proprietary models sourced from the United States or Asia, thereby securing true digital sovereignty. The integration of European components, such as Eviden”s BullSequana XH3500 platform and the Bull BXI v3 interconnect, reinforces the strategic goal of enhancing European technological sovereignty by diversifying the EuroHPC fleet and controlling critical system integration layers, mitigating geopolitical supply chain risks.
When and where: The Alice Recoque project is an internationally collaborative effort financed under the EuroHPC JU framework, drawing its budget from the Digital Europe Programme (DEP), with the overall cost projected at €554 million over five years of operation. The system”s operation is managed by the Jules Verne consortium, led by France through the Grand équipement national de calcul intensif (GENCI) and the Commissariat à l”énergie atomique et aux énergies alternatives (CEA), with key collaborative partners including the Netherlands (represented by SURF) and Greece (represented by GRNET). The physical infrastructure will be installed and operated at the CEA”s Very Large Computing Center (TGCC) in France, a site that already hosts other major Eviden-built systems. The system is named in honor of Alice Recoque (1929–), a French computer scientist and AI pioneer who specialized in computing architectures, including massively parallel machines, and later became a strategic leader in AI for the Bull group.
Who and how: The deployment of two distinct Exascale systems—JUPITER, utilizing a blend of NVIDIA accelerators and the European SiPearl ARM processor, and Alice Recoque, relying on AMD EPYC CPUs and AMD Instinct MI430X GPUs—is a deliberate and critical component of the European Union”s technological resilience strategy. This dual-architecture approach minimizes the risk associated with reliance on a single vendor or geopolitical supply chain, promoting diversification across major accelerator ecosystems (AMD ROCm and NVIDIA CUDA). The system”s architecture embraces a heterogeneous design, strategically combining traditional HPC simulation workloads with massive-scale data analysis and AI, with the acceleration partition anchored by next-generation AMD EPYC CPUs codenamed “Venice” and compute partitions unified by Eviden”s high-speed, low-latency Bull BXI v3 interconnect providing endpoint connection speeds of 400 Gb/s for CPUs and 800 Gb/s for GPUs.
This comprehensive strategic analysis examines Alice Recoque”s architectural specifications and core technology (AMD MI430X GPUs, HBM4 memory, Bull BXI interconnect), analyzes the strategic context and geopolitical positioning within the global Exascale competition, evaluates functional capabilities driving scientific breakthroughs and the AI Factory mission, assesses competitive positioning against U.S. systems (Frontier, El Capitan) and European peers (JUPITER), and delivers strategic conclusions regarding the system”s impact on European technological sovereignty, research capabilities, and industrial competitiveness.
Strategic Context and Geopolitical Positioning
The Exascale capability, defined as achieving 10^18 Floating Point Operations per second (FLOPS), has become the requisite baseline for global scientific and technological leadership, with Alice Recoque securing France”s position at this leading edge and marking Europe”s complete official entry into the Exascale era.
The Exascale Imperative: Global Competition and the Need for 10^18 FLOPS
Exascale capability represents a major generational performance leap for the European research community. The capacity to surpass one Exaflop/s for scientific simulation is equivalent to achieving more calculations than humanity could accomplish through four years of relentless mental computation. This extreme performance will mark a fiftyfold increase in computing capacity compared to the previous system, while remarkably only multiplying the electrical power requirement by five. This ratio underscores the critical advancements in power efficiency inherent in the newest generation of heterogeneous architectures.
The strategic investment in Alice Recoque is viewed as laying the foundation for future computing paradigms. While the primary mission addresses current Exascale requirements, the development plan explicitly mentions that the system will pave the path toward post-Exascale services. This forward-looking approach involves federating high-performance computing, artificial intelligence, and nascent quantum technologies to accelerate innovation and strengthen technological competitiveness. The implication is that the design is intended to serve not just as a computational workhorse, but as a critical infrastructure piece capable of incubating the hybrid classical/quantum architectures anticipated for the late 2020s and beyond.
Governance and Consortium: The Jules Verne Partnership
The Alice Recoque project is an internationally collaborative effort financed under the EuroHPC JU framework, drawing its budget from the Digital Europe Programme (DEP). The overall cost is projected at €554 million over five years of operation. The system”s operation is managed by the Jules Verne consortium, which is led by France through the Grand équipement national de calcul intensif (GENCI) and the Commissariat à l”énergie atomique et aux énergies alternatives (CEA). Key collaborative partners include the Netherlands, represented by SURF, and Greece, represented by GRNET.
The collaborative nature of the Jules Verne consortium is essential to the EuroHPC model. By co-funding the infrastructure, participating nations ensure the financial viability of these massive projects, while guaranteeing their national researchers, academics, and industries access to a portion of the system”s computational capacity. This distributed investment maximizes the political and scientific return on investment across the EU, thereby enhancing the competitive position of European researchers in a global context.
Dual Identity: France”s First Exascale and Europe”s Second System
Alice Recoque holds the distinction of being France”s first Exascale supercomputer. Within the broader EuroHPC fleet, it is designated as Europe”s second Exascale supercomputer, following the JUPITER system located in Germany.
The deployment of two distinct Exascale systems—JUPITER, utilizing a blend of NVIDIA accelerators and the European SiPearl ARM processor, and Alice Recoque, relying on AMD EPYC CPUs and AMD Instinct MI430X GPUs—is a deliberate and critical component of the European Union”s technological resilience strategy. This dual-architecture approach minimizes the risk associated with reliance on a single vendor or geopolitical supply chain, promoting diversification across major accelerator ecosystems (AMD ROCm and NVIDIA CUDA). Furthermore, it ensures that the EuroHPC fleet is equipped to optimally handle the diverse spectrum of scientific applications, from traditional FP64-heavy HPC simulations to the memory-intensive, low-precision demands of cutting-edge artificial intelligence, fostering a more robust and competitive internal high-end compute market.
Architecture and Core Specifications: Eviden”s BullSequana XH3500 Platform
The Alice Recoque system is built by Eviden upon its flagship BullSequana XH3500 platform, embracing a heterogeneous design that strategically combines traditional HPC simulation workloads with massive-scale data analysis and AI.
The Heterogeneous Compute Strategy: Convergence of HPC and AI
The architecture embraces a heterogeneous design, strategically combining traditional HPC simulation workloads with massive-scale data analysis and AI. The system”s purpose-built designation as an AI Factory dictates a structure optimized not just for raw floating-point calculation, but for the efficient orchestration of data throughout the entire computing lifecycle.
For a system dedicated to foundation model training—which involves distributing massive datasets and parameter updates across thousands of accelerators—the data infrastructure must sustain extreme input/output (I/O) performance. The architectural convergence mandate necessitates a fluid, low-latency data pipeline connecting the AMD CPU/GPU nodes to the storage subsystem. This optimization is crucial, as the performance of the AI Factory hinges on ensuring that the accelerators, the most expensive components, are never idled while waiting for data.
Accelerator Technology Deep Dive: The AMD Instinct MI430X GPU
The primary engine driving Alice Recoque”s Exascale performance and AI capabilities is the AMD Instinct MI430X GPU, a part of the newest MI400 Series accelerators specifically engineered for sovereign AI and scientific computing.
Compute Power and High-Bandwidth Memory (HBM4 Capacity): The most defining characteristic of the MI430X is its leadership memory configuration. Each GPU integrates an exceptional 432 GB of HBM4 memory, paired with an unprecedented memory bandwidth of 19.6 TB/s per chip. The massive 432 GB HBM4 capacity represents a profound architectural shift, providing an enormous advantage in the domain of large-scale AI. Modern foundation models, including generative AI and large language models (LLMs), frequently contain hundreds of billions or even trillions of parameters. Efficient training of these models necessitates that the parameters reside entirely within the fast, high-bandwidth memory accessible by the GPU. By dramatically increasing the HBM capacity—significantly surpassing the 192 GB offered by previous architectures such as the MI300X—Alice Recoque is uniquely capable of hosting and training the next generation of memory-intensive European AI models natively, directly supporting the objective of sovereign AI development.
Low-Precision Leadership: Native Support for FP4 and FP8 Data Types: While the system is guaranteed to exceed 1 Exaflop/s in double-precision (FP64) HPL performance, its effective throughput for AI is much higher due to native support for advanced, low-precision AI data types. The AMD Instinct MI430X supports FP4 and FP8 data formats, providing leadership AI FLOPs performance. In AI training and inference, lower precision formats are sufficient and significantly increase arithmetic density. The use of FP8, for instance, typically yields a theoretical performance increase of 4x or greater compared to FP16/BF16. Consequently, for deep learning workloads optimized for these low-precision formats, the system”s effective peak performance for AI training is projected to be in the multi-Exaflop range, positioning it directly against global supercomputing systems specialized in massive data processing and deep learning.
Host Processor and Interconnect Fabric
The acceleration partition of the system is anchored by next-generation AMD EPYC CPUs, codenamed “Venice,” which manage general-purpose computation and coordinate the massively parallel GPU cluster.
The compute partitions are unified by Eviden”s internal networking solution, the high-speed, low-latency Bull BXI v3 interconnect. This specialized fabric provides endpoint connection speeds of 400 Gb/s for CPUs and 800 Gb/s for GPUs. The selection of the European-developed BXI interconnect is a significant architectural decision that reinforces technological sovereignty. This strategy ensures European control over the essential inter-node communication layer, which is crucial for managing synchronous, large-scale distributed training across thousands of accelerators and mitigating potential geopolitical risks associated with relying entirely on external proprietary networking stacks. The high bandwidth of 800 Gb/s on the GPU links is necessary to ensure the MI430X”s computational capability is not hindered by communication bottlenecks during the most intensive parallel workloads.
Energy Efficiency and Sustainability Profile
Europe has placed energy efficiency at the forefront of its Exascale deployment strategy, adhering to aggressive green computing goals. Alice Recoque is committed to achieving its sustained Exascale performance with less than 15 MW of electrical power.
Eviden”s architectural design contributes significantly to this efficiency, promising up to 50 percent better energy efficiency per GPU compared to previous generations, while requiring 25 percent less racks and components than other Exascale systems. The cooling system relies extensively on highly efficient methods, specifically implementing direct warm-water liquid cooling for 100% of the unified rack components, supplemented by chilled-door technologies for scalar racks. If the system meets its sub-15 MW target, it will establish a high bar for global energy efficiency, potentially positioning it as a world leader in the Green500 ranking and validating Europe”s commitment to combining scientific excellence with environmental sustainability.
| Metric | Specification | Significance |
|---|---|---|
| Exascale Status | France”s 1st, Europe”s 2nd | Critical national and EU milestone |
| Rmax (HPL Performance) | > 1.0 Exaflop/s | Sustained Exascale capability |
| Platform | Eviden BullSequana XH3500 | European hardware platform integration |
| Accelerator GPU | AMD Instinct MI430X (MI400 Series) | Next-gen AI and scientific compute engine |
| HBM4 Memory per GPU | 432 GB | Leadership capacity for large foundation models |
| HBM Bandwidth per GPU | 19.6 TB/s | Maximizes data throughput efficiency |
| AI Data Type Support | FP4, FP8 | Optimization for cutting-edge deep learning |
| Interconnect Fabric | Eviden Bull BXI v3 | European sovereign networking solution |
| Power Consumption Target | < 15 MW | Adherence to European Green Computing goals |
Functional Capabilities: Driving Scientific Breakthroughs and the AI Factory Mission
Alice Recoque is designed to tackle Europe”s most pressing societal and scientific challenges by mastering the integration of numerical simulation, massive data analysis, and artificial intelligence.
High-Performance Computing (HPC) Simulations
The system”s double-precision Exascale performance is dedicated to several critical areas of traditional scientific simulation:
Climate Modeling and Environmental Forecasting: Alice Recoque will enhance the development of high-precision models necessary for mitigating the effects of climate change. The Exascale capacity allows for simulations at resolutions previously unattainable, improving the accuracy of complex global models and enabling the refined prediction of localized extreme weather events.
Advanced Materials Science and Sustainable Energy: The computing power will accelerate innovation in new materials discovery, such as those crucial for advanced battery chemistries or structural components. It will also support large-scale energy research programs, including fusion energy development.
Processing Observational and Experimental Data: Beyond traditional simulations, the system is designed to handle the immense data challenges posed by modern scientific instruments, including astronomical telescopes, Earth observation satellites, and vast networks of IoT devices. Its capability to process, infer, and leverage this data at incredibly high speeds and maximum energy efficiency is paramount for contemporary research.
The AI Factory: Enabling Next-Generation European Models
The designation of Alice Recoque as an AI Factory highlights its primary specialized role: facilitating the training and deployment of large, cutting-edge AI models. The system is specifically engineered to underpin the training of Europe”s next generation of generative AI, multimodal, and foundational models.
This specialization is critical for establishing European technological autonomy. In the current geopolitical environment, control over foundational digital assets, particularly the massive AI models that influence information flow, economic competitiveness, and defense, is a strategic imperative. By dedicating the resources and the unique memory configuration of the MI430X to indigenous model development, Europe can reduce its dependence on externally trained, often proprietary, models sourced from the United States or Asia, thereby securing true digital sovereignty.
Personalized Medicine and Digital Twin Technology
A key application area leveraging the system”s converged HPC-AI architecture is advanced biomedical research. Alice Recoque will support the creation of digital twins of the human body for personalized medicine.
This application requires the convergence of complex numerical simulations—modeling biological processes such as fluid dynamics or cellular interactions—with the analysis of massive, individualized patient data sets (AI/data analytics). The system”s exceptional memory performance and combined processing power are essential for running these multi-physics simulations and AI inference models simultaneously, enabling faster diagnostics, personalized treatment plans, and breakthroughs in areas such as drug discovery.
Competitive Assessment: Alice Recoque in the Global and European Arena
Alice Recoque”s performance commitment of exceeding 1.0 Exaflop/s places it firmly within the top tier of global supercomputing systems, achieving competitive parity with the fastest machines in the world.
Comparative Benchmark: U.S. Systems (Frontier and El Capitan)
Globally, the Exascale landscape is led by systems primarily located in the United States, utilizing highly advanced heterogeneous AMD architectures. As of late 2025, the leading systems include:
- El Capitan (USA): Retains the top spot on the TOP500 list, achieving a verified 1.809 ExaFLOP/s on the HPL benchmark
- Frontier (USA): The first confirmed Exascale machine, with a measured Rmax of 1.353 ExaFLOP/s
- Alice Recoque (EU): Expected to surpass 1.0 Exaflop/s HPL
Although its FP64 theoretical peak performance may not immediately challenge the absolute leadership of systems like El Capitan, the strategic specialization in AI gives it a functional advantage in memory-intensive workloads. The adoption of the newer MI430X architecture provides a technological leap, notably the 432 GB HBM4 capacity, which significantly surpasses the memory density available on previous AMD Instinct architectures used in earlier Exascale systems like Frontier. This architectural decision confirms that the system is optimized specifically for the growing complexity of memory-bound AI models, prioritizing capability over raw FP64 arithmetic speed.
The European Dual Strategy: Alice Recoque vs. JUPITER
Alice Recoque operates as the second pillar in Europe”s strategic Exascale infrastructure, complementing the first European Exascale system, JUPITER, hosted in Germany.
The two systems exhibit deliberate architectural contrast:
- JUPITER employs Eviden”s BullSequana XH3000 platform, with its Booster Module utilizing NVIDIA technology and a Cluster Module based on the European SiPearl Rhea1 ARM processor
- Alice Recoque is based on the newer BullSequana XH3500, utilizing AMD EPYC “Venice” CPUs and MI430X GPUs, unified by the European Bull BXI v3 interconnect
This contrast establishes a crucial resilience and diversification mechanism for the European research ecosystem. By investing in systems based on two major, yet distinct, technology stacks (AMD/Eviden BXI and NVIDIA/SiPearl/InfiniBand), EuroHPC mitigates vendor lock-in risks and ensures that European scientists have access to platforms optimized for different application characteristics and software environments (ROCm vs. CUDA). This robust, diversified infrastructure is essential for long-term technological stability and independence.
Energy Efficiency Rankings (Green500 Context)
Energy efficiency is a primary metric for next-generation systems. Alice Recoque”s commitment to achieving Exascale performance below 15 MW is highly ambitious. This compares favorably to early Exascale systems like Frontier, which operates at approximately 21 MW. The promised efficiency gains—including up to 50% better energy efficiency per GPU—underscore a deliberate engineering effort to reduce the Total Cost of Ownership (TCO) and meet stringent EU environmental mandates. Success in meeting this efficiency target will grant Europe a powerful position of soft power in global supercomputing, influencing future international infrastructure standards by demonstrating that scientific leadership and massive computational scale can be achieved sustainably.
| System Name | Location/Consortium | Rmax HPL (ExaFLOP/s) | Primary Accelerator | Key Architectural Feature | Power (MW Est.) |
|---|---|---|---|---|---|
| El Capitan | USA (LLNL/NNSA) | 1.809 | AMD Instinct (MI300 Series) | National Security Focus, FP64 optimized | N/A |
| Frontier | USA (ORNL) | 1.353 | AMD Instinct MI250X | First E-scale, HPE Cray EX platform | 21 MW |
| JUPITER | Germany (EuroHPC) | > 1.0 | NVIDIA (Booster), SiPearl (Cluster) | Dual-module, European CPU/GPU diversification | N/A |
| Alice Recoque | France (EuroHPC/Jules Verne) | > 1.0 | AMD Instinct MI430X | Leadership HBM4 (432 GB/GPU), European BXI Interconnect | < 15 MW |
Strategic Impact on European Technological Sovereignty and Research
A primary motivation for the Alice Recoque investment is the achievement of technological sovereignty in advanced computing, with European autonomy significantly advanced through control over the system”s integration layers and proprietary components.
Reinforcing Autonomy: Controlled Supply Chain and European Technology Integration
While the system relies on state-of-the-art processors from international partners (AMD), European autonomy is significantly advanced through control over the system”s integration layers and proprietary components. Sovereignty is materialized through the integration of the Eviden BullSequana XH3500 platform and the crucial, European-developed Bull BXI v3 interconnect. Control over these networking fabrics and the system administration framework (such as CEA”s Ocean suite) ensures that Europe is not dependent on external entities for the core functionalities of inter-node communication, system monitoring, energy optimization (via Eviden”s Argos software), and overall operational control. This comprehensive control over the integration stack mitigates geopolitical supply chain risks and ensures the long-term operational integrity and flexibility of the asset.
Economic Development: Empowering European Startups and SMEs
Beyond fundamental scientific research, Alice Recoque is intended to serve as a catalyst for economic growth and industrial competitiveness. By acting as a backbone for Europe”s network of AI Factories, the system provides European startups and Small and Medium-sized Enterprises (SMEs) with the immense computing power required to train models and develop products that can compete globally.
This effort is a critical step toward the industrialization of AI, effectively moving high-end computational resources from purely academic domains into the commercial sector. By democratizing access to Exascale resources for industrial AI development, the system directly supports the transfer of scientific breakthroughs into tangible economic products and services, securing a competitive advantage for European industry in advanced AI services and technological infrastructure.
Access Policy and Allocation for Researchers
The EuroHPC JU ensures that Alice Recoque operates as a resource for the entire European scientific and industrial community. The Joint Undertaking manages a substantial share of the access time—ranging from 35% up to 50% of the system”s total capacity.
Access is granted via open calls for expression of interest and is allocated to users residing or located in an EU Member State or a country associated with the Digital Europe Programme or Horizon Europe. The allocation process is managed through an international, merit-based peer-review process. This standardized and transparent framework is vital for ensuring the €554 million investment maximizes its public and scientific return. This transparent allocation process assures the global scientific community that access is based solely on the quality and scientific merit of the research proposal, thereby attracting world-class research talent to European infrastructure.
Conclusion: Strategic Implications and Future Trajectory
Alice Recoque represents a profound investment in France”s and Europe”s digital future, successfully achieving the Exascale milestone while concurrently establishing a new model for advanced computation that balances scientific excellence, technological sovereignty, and environmental sustainability.
Specialized Exascale Capability: Memory-Intensive AI Leadership
Alice Recoque is strategically specialized for memory-intensive AI. The core decision to utilize the AMD Instinct MI430X GPU, providing an industry-leading 432 GB of HBM4 memory, confirms a strong prioritization of large-scale foundation model training over maximal FP64 peak performance. This specialization is the most critical technical attribute, enabling the system”s success as a European AI Factory and addressing the immediate geopolitical need for sovereign AI model development.
Architectural Resilience: Dual-Architecture Strategy
By deploying Alice Recoque (AMD/Eviden BXI) as Europe”s second Exascale system alongside JUPITER (NVIDIA/SiPearl), the EuroHPC JU has implemented a resilient dual-architecture strategy. This approach diversifies technological dependency, mitigates risks associated with any single vendor or software stack, and ensures optimal resource matching for a broad range of scientific applications.
Sovereignty through Integration: European Control of Critical Layers
While core processors are internationally sourced, the system achieves technological sovereignty through the deployment of European intellectual property in the integration layers—the Eviden BullSequana platform and the Bull BXI v3 interconnect. Controlling these critical system elements secures long-term operational autonomy and supports the growth of indigenous European HPC expertise.
Global Leadership in Sustainability: Green Computing Excellence
The aggressive operational target of less than 15 MW and the commitment to warm-water cooling technology position Alice Recoque to potentially set global benchmarks in the Green500 ranking, leveraging technological superiority to align with European environmental policy goals.
Alice Recoque is fundamentally a declaration of European intent: to secure technological autonomy, foster industrial innovation, and maintain scientific leadership through the resilient, specialized, and sustainable deployment of next-generation computing infrastructure. The system”s forward-looking design, capable of federating HPC, AI, and nascent quantum technologies, ensures that this €554 million investment will remain relevant as computing research progresses toward the quantum era, positioning Europe at the forefront of the next generation of computational paradigms.
