MareNostrum 5, located in Spain, holds the prestigious title of the fourteenth most powerful supercomputer globally, as ranked on the TOP500 list<\/a>. This isn\u2019t just a supercomputer; it is a “living” entity that continually evolves. Recently, it received a significant boost in funding for a critical upgrade focused on artificial intelligence (AI) model training and inference.<\/p>\n The Barcelona Supercomputing Center (BSC) announced an investment of approximately 129 million euros<\/strong> to enhance MareNostrum 5. This funding, secured through a collaboration with Fsas Technologies (Fujitsu) and Telef\u00f3nica, will be co-funded 50% by the European High Performance Computing Joint Undertaking (EuroHPC JU) and the other 50% by Spain, Portugal, and Turkey. The upgrade is set to be implemented in the first half of 2026<\/strong>, marking a significant step in the supercomputer’s evolution. Detailed documentation can be found on the European Commission’s website<\/a>.<\/p>\n The upcoming update for MareNostrum 5 isn\u2019t just a reiteration of previous capabilities but is solely focused on artificial intelligence and machine learning workloads. A key goal is to enhance European digital sovereignty, thus reducing technological dependencies and preventing unintentional technology transfers out of the EU.<\/p>\n MareNostrum 5’s expanded architecture will introduce specialized compute partitions dedicated to tasks such as large language model (LLM) training and inference. Currently, this supercomputer operates with five distinct partitions, providing a staggering 175.30 PFLOPS<\/strong> of Rmax power (FP64, Linkpack) or 215.40 PFLOPS<\/strong> according to EuroHPC. Its total power capacity reaches approximately 20 MW<\/strong>.<\/p>\n Aiming for 30 FP4 exaflops<\/strong> specifically for AI tasks is ambitious, especially when considering that the current power dedicated to traditional HPC is just 0.17 FP64 exaflops<\/strong>. These new FP4 exaflops are tailored for training and large-scale inference and shouldn’t be directly compared with their FP64 counterparts.<\/p>\n While the technical documentation lacks specific details on the hardware configuration, it does mention a GH72 partition<\/strong>, which bears similarities to the specifications of NVIDIA\u2019s DGX GB200 NVL72 supercomputer. Minimum requirements, including 180 GB of HBM3e memory or higher<\/strong>, suggest potential compatibility with multiple high-performance GPU models, including NVIDIA and AMD options.<\/p>\n In conjunction with these developments, BSC has also announced progress in semiconductor design. The successful validation of the Cinco Ranch TC1 chip<\/strong>, developed by the Barcelona Zettascale Lab, represents a stride towards technological independence for the EU. Utilizing RISC-V architecture and Intel’s 3 nm technology, this chip operates at 1.25 GHz<\/strong> and marks a notable achievement in semiconductor design.<\/p>\n The MareNostrum 5 upgrade is a vital component of EuroHPC\u2019s broader mission to advance European supercomputing capabilities. Currently, Europe boasts 12 supercomputers<\/strong>, including notable systems like Jupiter in Germany and the incoming Alice Recoque in France, which are expected to be among the first exascale systems in Europe.<\/p>\n In conclusion, MareNostrum 5 is not just a supercomputer; it is a symbol of Europe\u2019s ambition to lead in supercomputing and AI technology, ensuring its independence and fostering innovation within the continent.<\/p>\nA Major Investment for Innovation<\/h3>\n
Adapting for a New Era: Focus on AI<\/h3>\n
The Shift from HPC to AI<\/h3>\n
Targeting 30 Exaflops for AI<\/h4>\n
Hardware Innovations: NVIDIA on the Horizon<\/h3>\n
A Step Towards Semiconductor Independence<\/h3>\n
Strengthening European Supercomputing<\/h3>\n