Here, \u201cFLOPS\u201d (or \u201cFLOP\/s\u201d) doesn\u2019t refer to \u201cblunders\u201d – but rather to floating-point operations per second<\/em>. \u201cExa\u201d is the SI prefix for 10\u00b9\u2078 – i.e., one trillion on the long scale, or one quintillion on the short scale, as used in the U.S. or Russia.<\/p>\n The Jupiter supercomputer, officially inaugurated in early September 2025 at the J\u00fclich Research Centre<\/a> between Cologne and the Belgian border, is expected – according to Heise – to ultimately deliver approximately 1.4 exaFLOPS for double-precision floating-point calculations (FP64), powered by nearly 24,000 NVIDIA GH200 boards. That far exceeds its initial benchmark result for the Top500 list. Even so, Jupiter already secured fourth place on that list with a preliminary score of 793 petaFLOPS. For AI applications – which often use lower-precision FP8 arithmetic – Jupiter is projected to reach over 70 exaFLOPS.<\/p>\n With 24,000 NVIDIA GPUs, Jupiter is ready for AI, quantum physics, and data-intensive research. Image source: Pexels\/panumas nikhomkhai.<\/p>\n<\/div>\n \u201cThe Jupiter supercomputer strengthens Germany\u2019s digital sovereignty,\u201d says Dr. Ralf Wintergerst, President of Bitkom, adding in a press release from the German digital association: \u201cWith Jupiter, Germany has entered the global elite of high-performance computing – and thereby improved conditions for developing artificial intelligence. Crucially, access to Jupiter must be made as straightforward and unbureaucratic as possible – for both startups and established companies. This would give AI development in Germany a real boost and help attract top international talent.\u201d<\/p>\n As t3n reports<\/a>, China has invested massively in supercomputer development and construction for roughly two decades. In 2013, the Tianhe-2 – with 33 petaFLOPS – dethroned the previous U.S.-built No. 1 system, outperforming it by a factor of two. It held the top spot until 2016, when it was overtaken by China\u2019s Sunway TaihuLight – only to lose first place again in 2018 to the U.S. system Summit at Oak Ridge National Laboratory in Tennessee. Summit, after an upgrade, reached 148,600 teraFLOPS – but was itself dethroned in May 2020 by Japan\u2019s Fugaku supercomputer at the RIKEN Center for Computational Science in Kobe.<\/p>\n Europe\u2019s top-ranked system in 2024 was the world\u2019s fifth-fastest HPC6, delivering 477.90 \/ 606.97 petaFLOPS. As of today, El Capitan<\/a> – also built on an HPE Cray platform – holds the global lead with 1.742 and 2.74638 exaFLOPS, respectively.<\/p>\n As t3n further notes, however, the Top500 list is steadily losing relevance – and since 2024, China has stopped reporting any of its supercomputers to the list\u2019s organizers. Beyond that, a growing number of AI-dedicated systems – especially from the U.S. and China – are entering the exaFLOPS realm. That\u2019s because AI training typically requires far less precision than standard 64-bit arithmetic.<\/p>\n As noted earlier, Jupiter achieved 793 petaFLOPS during its trial operation starting in June 2025 – landing it squarely at No. 4 on the Top500. Benedikt von St. Vieth, head of the High-Performance Computing, Cloud and Data Systems and Services department at the J\u00fclich Supercomputing Centre (JSC), remains confident: \u201cWe expect to hit one exaFLOP before the end of this year.\u201d<\/p>\n Jupiter is built modularly – and has been since May 2025. Yet during the June benchmark run, some racks remained switched off to protect the system. At that time, only 22,000 of the now fully deployed 24,000 GPUs (graphics processing units) were active. \u201cWe observed the system running too hot – so we had to optimize cooling,\u201d von St. Vieth told t3n.<\/p>\n In 2026, Jupiter Booster – a dedicated expansion module – will add at least one more unit. The Jupiter Cluster will be equipped with especially powerful processors designed for highly \u201cdata-intensive applications.\u201d However, these processors – developed in Europe by Si-Pearl and manufactured by TSMC – have not yet been delivered.<\/p>\n \n\u201cReality is far more complex,\u201d says Herten in t3n, adding: \u201cIt depends on the model architecture, how much training data is used, and what level of precision is required.\u201d\n<\/p><\/blockquote>\n Jupiter is primarily designed to run all kinds of simulation calculations. Its application spectrum spans cosmological models in astrophysics, simulations of biological systems, particle physics computations – and even AI tools for better analysis of brain data and neural functions.<\/p>\n AI capabilities are naturally central to Jupiter, emphasizes Andreas Herten, head of the Accelerating Devices Lab at JSC. The Jupiter Booster, developed by Eviden, according to the Helmholtz Institute<\/a>, integrates around 24,000 NVIDIA GH200 Grace Hopper Superchips. Altogether, the module comprises roughly 6,000 compute nodes – each equipped with four Grace Hopper Superchips and correspondingly powerful processors featuring AI accelerators. Thus, the module houses approximately 24,000 GPUs – and just as many CPUs.<\/p>\n This computational capacity is sufficient to train a large language model (LLM) with 100 billion parameters in about one week. \u201cThat\u2019s obviously a highly simplified estimate,\u201d says Herten in t3n, adding: \u201cIt depends on the model architecture, the volume of training data, and the desired precision.\u201d In theory, even significantly larger models could be trained – but that would take considerably longer. A planned AI model named Jarvis – scheduled for early 2026 – will serve primarily to train AI applications under the European AI Factory initiative.<\/p>\n <\/p>\n <\/p>\n Header Image Source: Forschungszentrum J\u00fclich \/ Sascha Kreklau<\/em><\/p>\n\u201cJupiter Strengthens Germany\u2019s Digital Sovereignty\u201d<\/h2>\n
![\"\"](\"https:\/\/www.cloudmagazin.com\/wp-content\/uploads\/2025\/09\/pexels-cookiecutter-17489152-250x166.jpg\")
<\/a><\/p>\nAlready Ranked #4 in Trial Operation<\/h2>\n
Also Suited for Training Large Language Models<\/h2>\n
Continue reading on cloudmagazin.com<\/h3>\n
\n