Supercomputing changed in the late 90s. It used to be that the biggest and fastest computers were all completely custom jobs, with proprietary everything and exotic hardware running esoteric architectures. In the late-90s, someone figured out that you could achieve supercomputer performance on some tasks simply by wiring together a large cluster of off-the-shelf PCs. Ever since then, most of the world's fastest supercomputers have essentially been overgrown commodity servers, sometimes with custom interlinks or other peculiarities.
So it goes with the latest machine to top the Top500 list. For those unfamiliar, Top500 is a list of the top five-hundred fastest supercomputers in the world—or at least, those that submit their results to be ranked, anyway. The most-recent Top500 list just came out, and resting atop the pile is the brand-new Frontier system, owned and operated by the USA's Oak Ridge National Laboratory.
Frontier was built by HPE Cray, and makes use of an unbelievable number of third-generation AMD EPYC processors outfitted with AMD Instinct MI250X accelerators. On the standard HPL benchmark, it managed to crank out 1.1 EFLOPs, making it the very first machine in the world to surpass the exaFLOP barrier. For perspective, 1.1 EFLOPs is 1.1×10^18 floating point operations per second, or around the performance of 36,000 GeForce RTX 3080 cards.
Despite the gigantic size of the machine, its incredible performance means that it is actually not just the fastest, but also the most efficient: Frontier topped the Green500 list, too. Each cabinet of the machine delivers 62.68 GFLOPs of compute per watt consumed, although when you take into account all of the networking required for the full system its efficiency drops to "just" 52 GFLOPS/watt—still high enough to take the top spot.
HPE Cray's new AMD-powered machine knocked Japan's Fugaku system off the #1 spot. That machine is powered by ARM A64X processors, and it too has a peak theoretical performance over 1 EFLOP, but it hasn't been able to realize that performance on the Linpack benchmark that Top500 uses. It's held the top spot for two years with "just" 442 PFLOPs, or slightly under half of what Frontier is capable of.
To illustrate how far ahead Fugaku and Frontier are from the rest of the list, a very similar system to Frontier—also built by HPE Cray and using the same basic AMD components—slotted in freshly at position #3. That system is called LUMI, and it lives in Finland, where it produces 152 PFLOPs for that nation's IT Center for Science, known as CSC.
In fact, AMD proudly notes in its press release that it lays claim to 94 systems out of the Top500 list, which doesn't sound that impressive on the face of it, but that number has nearly doubled since last year. The company also gloats that it dominates the Green500 list, holding the top 4 positions, eight of the top ten, and 17 of the top 20.
A cynic might point out that that's simply because the AMD systems are also the newest systems, and that future machines using Intel, NVIDIA, IBM, or ARM hardware might similarly leapfrog the red team's equipment. That's potentially in the future, though, and not a guaranteed. As such, let's let AMD celebrate a bit, because they've earned it.