Frontier Supercomputer to Usher in Exascale Computing

In 2018, a brand new supercomputer referred to as Summit was put in at Oak Ridge Nationwide Laboratory, in Tennessee. Its theoretical peak capability was almost 200 petaflops—that’s 200 thousand trillion floating-point operations per second. On the time, it was probably the most highly effective supercomputer on the planet, beating out the earlier document holder, China’s Sunway TaihuLight, by a cushty margin, based on the well-known Top500 rating of supercomputers. (Summit is at the moment No. 2, a Japanese supercomputer referred to as Fugaku having since overtaken it.)

In simply 4 brief years, although, demand for supercomputing companies at Oak Ridge has outstripped even this colossal machine. “Summit is 4 to 5 occasions oversubscribed,” says Justin Whitt, who directs ORNL’s Management Computing Facility. “That limits the variety of analysis tasks that may use it.”

The apparent treatment is to get a sooner supercomputer. And that’s precisely what Oak Ridge is doing. The brand new supercomputer being assembled there’s referred to as Frontier. When full, it would have a peak theoretical capability in extra of 1.5 exaflops.

The outstanding factor about Frontier just isn’t that it is going to be greater than seven occasions as highly effective as Summit, beautiful as that determine is. The outstanding factor is that it’ll use solely twice the facility. That’s nonetheless loads of energy—Frontier is anticipated to attract 29 megawatts, sufficient to energy a city the scale of Cupertino, Calif. But it surely’s a manageable quantity, each when it comes to what the grid there can provide and what the electrical energy invoice will likely be.

“The effectivity comes from placing extra pc {hardware} in smaller and smaller areas,” says Whitt. “Every of those [computer] cupboards weighs as a lot as a full-sized pickup.” That’s as a result of they’re full of what ORNL’s spec sheet describes as “excessive density compute blades powered by HPC- and AI-optimized AMD EPYC processors and Radeon Intuition GPU accelerators purpose-built for the wants of exascale computing.”

Constructing a supercomputer of this capability is tough sufficient. However doing so throughout a pandemic has been particularly difficult. “Provide-chain points had been broad,” says Whitt, together with shortages of many issues that aren’t particular to constructing a high-performance supercomputer. “It might simply be sheet steel or screws.”

Provide points are certainly the rationale Frontier will turn out to be operational in 2022 forward of one other deliberate supercomputer, Aurora, which will likely be put in at Argonne Nationwide Laboratory, in Illinois. Aurora was to come back first, however its building has been delayed, as a result of Intel is having issue supplying the processors and GPUs wanted for this machine.

On the time of this writing, technicians at Oak Ridge had been assembling and testing elements of Frontier in hopes that the enormous machine will come collectively earlier than the top of 2021 and with the intention of constructing it absolutely operational and out there for customers in 2022. Will we then be capable to name it the world’s first exascale supercomputer?

That will depend on your definition. “[Japan’s Fugaku supercomputer] truly achieved 2 exaflops with a special benchmark,” says Jack Dongarra of the College of Tennessee, one of many specialists behind the Top500 checklist. These rankings, he explains, are primarily based on a benchmark that includes 64-bit floating-point calculations, the sort used to unravel three-dimensional partial differential equations as required for a lot of bodily simulations. “That’s the underside line of what supercomputers are getting used for,” says Dongarra. However he additionally factors out that supercomputers are more and more used to coach deep neural networks, the place 16-bit precision can suffice.

Will we be capable to name Frontier the world’s first exascale supercomputer? That will depend on your definition.

After which there’s Folding@Residence, a distributed-computing undertaking meant to simulate protein folding. “I might name {that a} specialised pc,” says Dongarra, one that may do its job as a result of the calculations concerned are “embarrassingly parallel.” That’s, separate computer systems can carry out the required calculations independently—or no less than largely so, with what little communication between them is required being conveyed over the Web. In March of 2020, the Folding@Residence undertaking proudly introduced on Twitter, “We’ve crossed the exaflop barrier!”

However in the event you persist with the standard benchmark, the one used for the Top500 rating, no supercomputer but qualifies as an exascale machine. Frontier would be the first. Or effectively, it’s on monitor to be the primary recognized exascale supercomputer, says Dongarra. He explains that earlier than the June 2021 Top500 rating got here out, a rumor emerged that China has no less than one, if not two, supercomputers already working at exascale.

Why would Chinese language engineers assemble such a machine with out telling anybody about it? On the time, Dongarra says, he thought that possibly they had been ready for the 100-year anniversary of the founding of the Chinese language communist occasion. However that date got here and went in July. He now speculates that Chinese language officers could also be anxious that making its existence public would exacerbate geopolitical rivalries and trigger the USA to limit the export of sure applied sciences to China.

Maybe that explains it. But it surely’s going to be more and more tough for Chinese language researchers to not let this cat, if it really exists, out of the bag. For the second, anyway, with solely rumors to go on, this exascale rival to Frontier is a Schrödinger’s cat—each right here and never right here on the identical time.

This text seems within the January 2022 print subject as “The Exascale Period Is Upon Us.”