Nvidia and the German research center Jülich have the key data Jupiter announced: 23,752 GH200s make the supercomputer the fastest in Europe. There is a lot of performance not only for artificial intelligence (AI), the ExaFLOP brand also applies to HPC applications. And all this with only 18.2 megawatts.
Very fast and efficient
Climbing higher, faster and further without having to look at your electricity bill (and purchase costs) is quite easy. But if power consumption does not get completely out of control, this automatically limits the choice of hardware.
Germany’s fastest supercomputer, which is also the new No. 1 in Europe, can also be better sold politically by focusing on efficiency; After all, (industrial) electricity and its price are a constant problem in this country. And since the public sector massively supports the Jupiter project, so it should be: €250 million will come from the European supercomputing initiative EuroHPC JU and €250 million will come equally from the Federal Ministry of Education and Research (BMBF) and the Ministry of Culture and Science of the state of North Rhine-Westphalia (MKW NRW).
18.2 megawatts of power consumption with HPC performance of 1.0 ExaFPLOPS is in the league of the current #1 Frontier system with AMD hardware. According to the May 2023 Top500, Frontier has so far delivered 1.2 ExaFLOPS of performance with just over 20 megawatts. However, recently Frontier has been optimized further: for a perfect comparison with Jupiter we still have to wait for new data. Once again, Frontier and Jupiter make it clear one way or another: without a GPU, nothing works on supercomputers.
GH200 at its best
The GH200 superchip, also known as Grace Hopper, is used in the JUPITER supercomputer by tens of thousands, the abbreviation means “Pioneering joint venture in innovative and transformative exascale research“, installed, more precisely in 23,752 versions. The GPUs find their place in the so-called boost module.
To this end, Nvidia has created a new quad pack for the more efficient and space-saving use of four GH200s, which Eviden (part of the Atos Group) will then install in the Atos BullSequana XH3000. A quad pack marks one node, two nodes are installed on one blade; this has been the rule lately with these solutions. Of course, the blade is liquid cooled, which is no longer possible with such compact products that consume a lot of electrical energy.
The overall system will require the space of about four tennis courts and will use more than 260 km of high-performance cabling.
The cluster module will be equipped with SiPearl’s new Rhea processor designed and manufactured in Europe, a CPU with high memory bandwidth for complex workloads. The cluster and booster modules operate dynamically as a unified supercomputer using ParTec’s ParaStation Modulo modular operating system.
Installation from early 2024
Jupiter installation will begin in early 2024. With construction, scientific users will have the opportunity to prepare and test the system as part of the Early Access Program. This allows close collaboration between all parties involved to produce and configure the best possible version of the system for the scientific community, the partners explained in a press release.
Update November 14, 2023 9:52 am
ComputerBase received information about this article from Nvidia under NDA. The only requirement was the earliest possible publication date.