Jupiter Booster: Almost the first exascale supercomputer in Europe
The Jülich Jupiter Booster ranks fourth in the Top500 list of the world's fastest supercomputers, making it the fastest supercomputer in Europe.
(Image: Forschungszentrum Jülich)
The fastest European supercomputer currently comes once again from Germany, more precisely from Forschungszentrum Jülich. Jupiter replaces the Italian HPC6 as the fastest system in Europe and the US supercomputer Eagle as the fastest supercomputer with Nvidia technology. However, because Jupiter Booster does not yet calculate at full power, it just misses the exaflops mark. The supercomputer is due to go into full operation in July 2025.
A partial installation has already moved up to fourth place in the 65th Top500 list behind El Capitan, Frontier and Aurora, whose ranks and performance have not changed. The current expansion stage of the Jupiter Booster – achieves 793.4 petaflops, making it almost 18 times faster than its predecessor Juwels (44.12 petaflops).
As is usual with supercomputers, these measurements are carried out using the high-performance Linpack benchmark in double-precision floating-point arithmetic (floating point 64-bit, FP64) and consist mainly of matrix multiplications.
The accelerator part Booster is still supported by a classic supercomputer part Jupiter Cluster. As early as November 2024, the Jupiter Exascale Transition Instrument (JETI), a smaller sub-installation of Jupiter Booster, beat its predecessor Jewels by a factor of 2. It not only previously ranked 18th in the Top500 supercomputers, but also sixth in the Green 500, the most efficient supercomputers.
Videos by heise
Massive hardware
The system comes from Eviden and Atos from the BullSequana XH3000 series and is based on Nvidia's Grace Hopper Superchips GH200, which combine a 72-core ARM processor with an H100 GPU. Each GPU uses 96 GByte HBM3 RAM. In total, the Jupiter Booster will occupy 125 racks and have around 24,000 GH200s in 6000 nodes. This means that an enormous 2.3 petabytes of super-fast HBM3 memory are available. Cooling is provided by hot water.
The Jupiter Boost calculates even faster when training AI models because simpler data formats such as FP8 are used there. The FZ Jülich is projecting more than 70 AI exaflops when the supercomputer is ready soon. Nvidia is certain that it will have over 90 AI exaflops. However, the system achieves the enormous AI performance not only through the low FP8 accuracy, but also through the calculation of sparsely populated matrices. With fully populated matrices, the throughput is halved again.
According to plans, Jupiter will also be equipped with 21 petabytes of fast mass storage, using 40 IBM Elastic Storage Server 3500 with NVMe SSDs. Together, they should provide data transfer rates of up to 4 TByte/s for reading and 3 TByte/s for writing.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier ein externer Preisvergleich (heise Preisvergleich) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (heise Preisvergleich) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
(csp)
