From JUPITER To Pre-Zettascale Supercomputing

Advanced Pre-Zettascale Infrastructure // Co-Designing Modular Supercomputers // Highly Scalable, Optimized for AI // High-Capacity Storage // Energy-Efficiency // Internationally
competitive // JAZMIN – JUPITER – JURECA – HoreKa

Jülich Supercomputing Centre (JSC) at Forschungszentrum Jülich and the Scientific Computing Center (SCC) at Karlsruhe Institute of Technology (KIT) operate internationally competitive supercomputing, high-capacity storage and cloud infrastructures for the national and European research communities and industry. The supercomputers are highly scalable and optimized for artificial intelligence – and among the most energy-efficient systems of their scales.

The Jülich Advanced pre-Zettascale Modular INfrastructure – JAZMIN – at JSC provides exascale HPC and data storage and pre-zettascale AI-capabilities for users from science and industry across Germany and Europe. JAZMIN’s modular design enables seamless expansion and integration following emerging technologies, while bundling expertise in infrastructure, HPC/AI and data systems, network technologies and user support. As part of a federated European HPC, AI and data ecosystem, JAZMIN advances HPC simulation, AI and data analytics across all programs of Helmholtz Information to address scientific challenges and major societal issues.

JAZMIN’s HPC and AI capacities comprise the modular supercomputers JURECA-DC (Jülich Research on Exascale Cluster Architectures – Data Centric) for both local and national users, JUWELS (Jülich Wizard for European Leadership Science) for German scientists and their partners, JUPITER (Joint Undertaking Pioneer for Innovative and Transformative Exascale Research) for German and European users. 

JURECA‑DC was co‑funded by the European Commission through a joint procurement process with three partner centers. JURECA‑DC is designed to support mid-scale, data‑intensive applications, and also incorporates an evaluation platform for novel computing paradigms and JURECA-HWAI for AI research. JUWELS, funded by BMFTR and the Ministry of Culture and Science of the German state of North Rhine-Westphalia, was developed as a blueprint for exascale systems with research and industry partners. It was the first leadership system to integrate CPU‑centric resources (JUWELS Cluster) with NVIDIA Ampere A100 GPUs (JUWELS Booster) in a single coherent architecture. JUWELS Booster, among the world’s first large‑scale GPU installations, ranked seventh on the TOP500 list and remains Germany’s second‑fastest machine in 2025, ranking just behind JUPITER. JUWELS will be decommissioned in 2026, with the JUWELS Booster potentially being repurposed for industrial use to meet rising GPU demand for AI applications. JUPITER comprises three modules. JUPITER Booster is Europe’s first exascale system, with nearly 24,000 NVIDIA GH200 CPU-GPU superchips delivering 1 ExaFLOP/s HPL performance, ranking fourth on the TOP500 list in November 2025, and 80 ExaFLOP/s at 8-bit AI. In 2025 it was the most energy-efficient Exascale supercomputer on the Green500 list. JUPITER Cluster is designed for workloads that are less suited to GPUs. It will feature the SiPearl Rhea1 processor, co‑designed through the European Processor Initiative (EPI). JARVIS (JUPITER Advanced Research Vehicle for Inference Services), the third module, is being acquired within the JUPITER AI Factory (JAIF) to provide cloud‑native resources for sovereign AI inference applications.

JAZMIN features a layered storage system beyond 1 ExaByte based on IBM Spectrum Scale, combining high-speed flash (NVMe), high-capacity hard disk drives (HDDs) and tape. Backed by an advanced high-bandwidth network, it supports the massive data demands of complex simulations and AI training.

JSC also offers a cloud computing platform, JSC Cloud, for users with a demand for more complex scientific workloads, such as hosting web-platforms or large-scale data services residing in the HPC file systems. One large cloud resource consumer is the AI inference service Blablador, which provides 24/7 cloud APIs based on GPU resources of JSC Cloud. To support the rising demands, the cloud environment will be scaled out with JARVIS in the future. JSC Cloud comprises OpenStack compute, network, and volume services as well as an integration with the DATA file system also available on the HPC systems. Additional OpenStack services as well as Kubernetes platforms are available.

When it was commissioned in mid-2021, HoreKa (Hochleistungsrechner Karlsruhe) at KIT-SCC was among the 15 most powerful computers in Europe and ranked 13th in the Green500. It provides a computing power of more than 17 PetaFLOP/s. The system is available to scientists from all over Germany. With HoreKa, researchers in the areas of Earth System Sciences, the Materials Sciences, Energy and Mobility research in Engineering and Particle Physics are able to gain a more detailed understanding of highly complex natural and technical processes. It is an innovative hybrid system with almost 60.000 Intel Xeon "Ice Lake" Scalable Processor cores and 668 NVIDIA Ampere A100 GPUs. A non-blocking NVIDIA Mellanox InfiniBand HDR network is used for communication between the nodes. Two IBM Spectrum Scale parallel file systems offer a total storage capacity of more than 16 PetaByte. A central aspect in the design of the system has been the enormous amounts of data generated by scientific research projects. Depending on the application, several hundred terabytes of data can be generated by a single simulation. To keep up with the growing amounts of data, the computing nodes, the InfiniBand network, and the parallel file systems of HoreKa each provide up to four times the throughput of its predecessor system, ForHLR.

In the next years, the co-designing of future modular supercomputers suitable for highly scalable applications and artificial intelligence workloads towards a pre-zettascale system will be a key activity. The JURECA successor (codename JURECA-NEXT) will serve as a live innovation testbed for assessing pre-zettascale capabilities within a production environment for local HPC and AI users. The experience gained with JUPITER, Europe’s first Exascale supercomputer and the most energy-efficient system of its scale, will feed into the design of a pre-zettascale system (codename JUPITER-NEXT). This will be a frontier-scale system, engineered to drive the next generation of HPC simulations and, in particular, AI and AI-infused workloads. It will enable large-scale scientific foundational models, supporting frontier research at unprecedented scale and energy efficiency.

Contact

Benedikt von St. Vieth

Jülich Supercomputing Centre
Forschungszentrum Jülich GmbH

Achim Streit

Scientific Computing Center
Karlsruhe Institute of Technology