SoylentNews

Arthur T Knackerbracket [soylentnews.org] writes:

--- --- --- --- Entire Story Below - Must Be Edited --- --- --- --- --- --- ---

Arthur T Knackerbracket has processed the following story [theregister.com]:

Developed as part of the EuroHPC Joint Undertaking and built-in collaboration with Nvidia and French supercomputer builder Eviden (formerly Atos), the system is designed to support research across a number of fields [fz-juelich.de], including biophysics, cellular neuroscience, nuclear and elementary particle physics, astrophysics, and climate science.

Boffins [scientists] are also exploring the potential application of machine learning algorithms like those used in diffusion models to generate images to advance areas like medical imaging and autonomous vehicles.

In its first appearance on the biannual ranking, the Jülich Supercomputing Center's (JSC) flagship system managed 793 petaFLOPS of double precision (FP64) grunt in the time-honored High-Performance Linpack (HPL) benchmark.

However, the run is just a teaser for what's to come. Much like the US Department of Energy's Aurora supercomputer in its first appearance on the Top500 in late 2023, Jupiter's is also a partial run. The system is widely expected to be the European continent's first true exascale supercomputer.

Built by Eviden for the EuroHPC Joint Undertaking and deployed at the Jülich Supercomputing Center in Germany, Jupiter is expected to be Europe's first exascale-capable supercomputer.

While Jupiter ranks as Europe's new supercomputing beast, it still falls far short of its American counterparts. The US DoE's El Capitan, Frontier, and Aurora systems remain uncontested as the only three exascale-capable [theregister.com] supercomputers on the Top500 at 1.74, 1.35, and 1.01 exaFLOPS, respectively.

That's likely to change in the not-so-distant future once Jülich can bring the full force of the Jupiter system to bear on the semiannual benchmark. The machine will only need to pack on another 207 petaFLOPS to its next HPL run to herald Europe's entrance into the exascale era, and just 220 more petaFLOPS to fly above Aurora. That's assuming, of course, the boffins at Argonne don't find a way to eke out a bit more compute from the Intel-based monster.

Nvidia insists that when the Jupiter system is complete, it will join its American counterparts in exascale orbit. From our calculations, that should be well within its capabilities.

Based on Eviden's BullSequana XH3000 platform, the Jupiter Boost section will feature roughly 6,000 compute nodes, each packed with four Nvidia Grace Hopper Superchips (GH200) and InfiniBand NDR200 networking.

Launched in 2023, the GH200 features [nvidia.com] a Grace CPU with 72 Arm Neoverse V2 cores and 480GB of high-speed LPDDR5x memory meshed to a Hopper GPU with up to 144GB of HBM3e memory via Nvidia's 900GB/s (450GB/s in each direction) NVLink Chip-to-Chip (C2C) interconnect.

Each of these superchips boasts double precision performance (FP64) of 34 teraFLOPS vector or 67 teraFLOPS matrix for high-performance computing.

With about 24,000 superchips, Jupiter's Booster section promises somewhere between 872 petaFLOPS (vector) and 1.6 exaFLOPS (matrix) performance.

However, Jupiter won't be limited to traditional HPC applications. Each superchip is capable of churning out roughly 2 petaFLOPS of dense FP8 compute (double that if you can take advantage of Nvidia's hardware sparsity).

On paper, that means Jupiter's Boost should deliver just over half the AI performance of Lawrence Livermore National Laboratory's (LLNL's) El Capitan system at approximately 47 exaFLOPS of dense AI compute versus 87 exaFLOPS on the all AMD system.

And that's just for the Booster. Jupiter's Universal Cluster section is slated to use SiPearl's Arm-based Rhea1 CPU, which features 80 Neoverse V1 cores and is equipped with 64GB of HBM.

With two Rhea1 processors per node and 1,300 nodes expected to fill the Universal Cluster, the section is expected to contribute another five petaFLOPS of compute to the overall system.

Jupiter isn't just Europe's most powerful system — for its size, it's also among the most efficient of the exascale-class systems, managing 60 gigaFLOPS per watt, putting it just ahead of El Cap at 58 gigaFLOPS per watt.

This shouldn't come as too much of a surprise. As we've previously discussed [theregister.com], Nvidia's Grace Hopper Superchips have proven to be remarkably efficient compute units with Jupiter's test-bed system JEDI once again topping the Green500 ranking of the most efficient supercomputers.

Jupiter's launch underscores Europe's growing influence in the HPC arena. As of today, five of the ten most powerful systems in the Top500 are based in Europe, a first in the semiannual ranking's more than 30-year history.

Alongside Germany's Jupiter, Italy's HPC6 and Leonardo systems hold the number six and 10 spots with 477 and 241 petaFLOPS, respectively. Meanwhile, Switzerland's Alps and Finland's Lumi machines come in eighth and ninth place with 434 and 379 petaFLOPS.

And while it didn't quite make the top 10, we can't ignore the United Kingdom's Isambard-AI phase 2 system, which came in 11th place and 216 petaFLOPS.

Aside from the US's three exascale giants, the only other American system to grace the top10 is Microsoft's Eagle cluster at 561 petaFLOPS, while Riken's Fugaku continues to hold strong as Japan's only contender in the Top500's upper echelon at 442 petaFLOPS.

China, meanwhile, remains missing in action, having largely withdrawn from the Top500 over the past few years. We know the Middle Kingdom already has [nextplatform.com] several home-grown exascale supercomputers deployed, but it simply hasn't released Linpack results. This continues to leave something of a blind spot on the time-honored ranking.

Original Submission