Europe's first exascale supercomputer is now up and running, using 24,000 Nvidia GH200 Superchips to perform more than one quintillion operations per second with nearly 1,000,000 terabytes of storage
Europe's fastest supercomputer is now operational, and boy howdy does it have an impressive specs sheet. The JUPITER project (which stands for Joint Undertaking Pioneer for Innovative and Transformative Exascale Research, apparently) had its official ribbon-cutting ceremony earlier this week at the Forschungszentrum Jülich campus in Germany, attended by the great and the good of European scientific researchers—and the odd politician, too.
JUPITER itself is built around 50 container modules over a 2,300 square meter site, and contains a simply astonishing amount of hardware. The Booster module contains roughly 6,000 nodes, over which 24,000 Nvidia GH200 Grace Hopper Superchips are interconnected via the company's Quantum-2 InfiniBand networking system, complete with a slew of Arm-based CPU cores, 288 per node.
The Cluster module, interlinked with Booster, is made up of more than 1300 nodes containing a wealth of Rhea1 processors, another Arm instruction set chip and designed by SiPearl. Each cluster node has two of these chips, with each containing 80 Arm Neoverse Zeus cores.
JUPITER is expected to be capable of exceeding 90 exaflops worth of AI performance and 1 quintillion FP64 operations per second running at full pelt, which is... well, rather speedy. Enough to make it the fourth fastest supercomputer in the world, in fact.
You're keeping up with me, yes? Good, I'll keep going. According to an Nvidia blog, 51,000 network connections let JUPITER transmit three times more data than all global data traffic at any given moment. It's also got nearly 1 exabyte of storage, equating to around 1,000,000 terabytes, which gives it plenty of room to install at least one copy of Call of Duty: Black Ops 6.
Ah, but JUPITER is destined for much greater purposes than playing silly old video games. It will instead be used for research purposes, focusing on climate science, generative AI models, neuroscience for drug discovery, mapping the human brain, and quantum simulation, among others.
As a given example, the Max Planck Institute for Meteorology is said to be using JUPITER to simulate climate predictions and extreme weather events. I'm imagining the results are going to come back as something like "we're really screwed", just like every other climate science prediction I've seen over the past two decades—but with all this computer power at its disposal, at least we've all got a good chance of knowing exactly how screwed we are in more accurate ways than ever before.
Which will be nice, won't it? Anyway, it seems like some real good might come from this fantastic machine, as beyond simulating climate events, it's also going to be simulating protein assembly in cells in an effort to combat HIV, study (and potentially discover new) particle interactions, and "develop spatio-temporal compression and diffusion architectures that enable the creation of high-quality, accessible video models to advance applications from medical imaging to autonomous driving."
All of which sounds like good stuff. So, put away the tired old Crysis gags, folks. This is an Nvidia-powered machine that might be doing some serious good, and one that I reckon might be worth setting aside the power requirements for. Godspeed, JUPITER. Perhaps you might be the eventual saviour of us all.