ISC Germany’s long-awaited Jupiter supercomputer launched into the quantity 4 spot on the Top500 checklist of publicly ranked techniques, dethroning Italy’s HPC6 as Europe’s largest, baddest iron.
Developed as a part of the EuroHPC Joint Enterprise and built-in collaboration with Nvidia and French supercomputer builder Eviden (previously Atos), the system is designed to assist analysis throughout a number of fields, together with biophysics, mobile neuroscience, nuclear and elementary particle physics, astrophysics, and local weather science.
Boffins are additionally exploring the potential utility of machine studying algorithms like these utilized in diffusion fashions to generate pictures to advance areas like medical imaging and autonomous autos.
In its first look on the biannual rating, the Jülich Supercomputing Middle’s (JSC) flagship system managed 793 petaFLOPS of double precision (FP64) grunt within the time-honored Excessive-Efficiency Linpack (HPL) benchmark.
Nevertheless, the run is only a teaser for what’s to return. Very similar to the US Division of Power’s Aurora supercomputer in its first look on the Top500 in late 2023, Jupiter’s can also be a partial run. The system is broadly anticipated to be the European continent’s first true exascale supercomputer.
Constructed by Eviden for the EuroHPC Joint Enterprise and deployed on the Jülich Supercomputing Middle in Germany, Jupiter is anticipated to be Europe’s first exascale-capable supercomputer. – Click on to enlarge
Whereas Jupiter ranks as Europe’s new supercomputing beast, it nonetheless falls far wanting its American counterparts. The US DoE’s El Capitan, Frontier, and Aurora techniques stay uncontested as the one three exascale-capable supercomputers on the Top500 at 1.74, 1.35, and 1.01 exaFLOPS, respectively.
That is prone to change within the not-so-distant future as soon as Jülich can carry the complete pressure of the Jupiter system to bear on the semiannual benchmark. The machine will solely have to pack on one other 207 petaFLOPS to its subsequent HPL run to herald Europe’s entrance into the exascale period, and simply 220 extra petaFLOPS to fly above Aurora. That is assuming, in fact, the boffins at Argonne do not discover a technique to eke out a bit extra compute from the Intel-based monster.
Full of Nvidia GH200s
Nvidia insists that when the Jupiter system is full, it should be part of its American counterparts in exascale orbit. From our calculations, that must be properly inside its capabilities.
Based mostly on Eviden’s BullSequana XH3000 platform, the Jupiter Increase part will function roughly 6,000 compute nodes, every filled with 4 Nvidia Grace Hopper Superchips (GH200) and InfiniBand NDR200 networking.
Launched in 2023, the GH200 features a Grace CPU with 72 Arm Neoverse V2 cores and 480GB of high-speed LPDDR5x reminiscence meshed to a Hopper GPU with as much as 144GB of HBM3e reminiscence through Nvidia’s 900GB/s (450GB/s in every course) NVLink Chip-to-Chip (C2C) interconnect.
Every of those superchips boasts double precision efficiency (FP64) of 34 teraFLOPS vector or 67 teraFLOPS matrix for high-performance computing.
With about 24,000 superchips, Jupiter’s Booster part guarantees someplace between 872 petaFLOPS (vector) and 1.6 exaFLOPS (matrix) efficiency.
Nevertheless, Jupiter will not be restricted to conventional HPC functions. Every superchip is able to churning out roughly 2 petaFLOPS of dense FP8 compute (double that for those who can make the most of Nvidia’s {hardware} sparsity).
On paper, meaning Jupiter’s Increase ought to ship simply over half the AI efficiency of Lawrence Livermore Nationwide Laboratory’s (LLNL’s) El Capitan system at roughly 47 exaFLOPS of dense AI compute versus 87 exaFLOPS on the all AMD system.
And that is only for the Booster. Jupiter’s Common Cluster part is slated to make use of SiPearl’s Arm-based Rhea1 CPU, which options 80 Neoverse V1 cores and is supplied with 64GB of HBM.
With two Rhea1 processors per node and 1,300 nodes anticipated to fill the Common Cluster, the part is anticipated to contribute one other 5 petaFLOPS of compute to the general system.
Jupiter is not simply Europe’s strongest system — for its measurement, it is also among the many best of the exascale-class techniques, managing 60 gigaFLOPS per watt, placing it simply forward of El Cap at 58 gigaFLOPS per watt.
This should not come as an excessive amount of of a shock. As we have beforehand discussed, Nvidia’s Grace Hopper Superchips have confirmed to be remarkably environment friendly compute items with Jupiter’s test-bed system JEDI as soon as once more topping the Green500 rating of essentially the most environment friendly supercomputers.
Jupiter’s launch underscores Europe’s rising affect within the HPC area. As of in the present day, 5 of the ten strongest techniques within the Top500 are based mostly in Europe, a primary within the semiannual rating’s greater than 30-year historical past.
Alongside Germany’s Jupiter, Italy’s HPC6 and Leonardo techniques maintain the quantity six and 10 spots with 477 and 241 petaFLOPS, respectively. In the meantime, Switzerland’s Alps and Finland’s Lumi machines are available eighth and ninth place with 434 and 379 petaFLOPS.
And whereas it did not fairly make the highest 10, we will not ignore the UK’s Isambard-AI part 2 system, which got here in eleventh place and 216 petaFLOPS.
Other than the US’s three exascale giants, the one different American system to grace the top10 is Microsoft’s Eagle cluster at 561 petaFLOPS, whereas Riken’s Fugaku continues to carry sturdy as Japan’s solely contender within the Top500’s higher echelon at 442 petaFLOPS.
China, in the meantime, stays lacking in motion, having largely withdrawn from the Top500 over the previous few years. We all know the Center Kingdom already has a number of home-grown exascale supercomputers deployed, however it merely hasn’t launched Linpack outcomes. This continues to depart one thing of a blind spot on the time-honored rating. ®
Source link
