US to rev fastest supercomputer with powerful test system

The United States has opened up a small but powerful test system for researchers to get a taste of what is expected to become the country's fastest supercomputer and one of the world's fastest. And the tiny testbed, named "Crusher" is already faster than previous top US supercomputers.

The US Department of Energy's Oak Ridge National Laboratory said on Monday that four well-established research projects are using the test system. Crusher is essentially a miniature version of Frontier, the country's first exascale supercomputer expected to come online in the second half of 2022, which means an already late system (Frontier was supposed to appear in 2021) is even later than expected (appearing after the expected ISC 2022 incarnation of the Top 500 list).

Frontier is a much-hyped system that is expected to provide at least 1.5 exaflops of theoretical peak performance, with some sources claiming two exaflops based on unpublished conference material. Peak performance is the metric by which top performance is gauged in supercomputing.

It will undoubtedly become one of the world's fastest supercomputers, easily surpassing Japan's 442-petaflop Fugaku, which is in the current No 1 spot. But as our sister site The Next Platform reported last year, China quietly beat the US to the exascale punch with not one but two systems that reached a peak performance of 1.3 exaflops.

• Look at me. Look at me. I'm the El Capitan now: Cray to build US govt's $600m cray-cray exascale nuke app super
• Hey, US taxpayers. Filed your taxes? Good, good. $500m of it is going on an Intel-Cray exascale boffinry supercomputer
• AMD tries to spoil Nvidia's week by teasing high-end accelerators, Epyc chips with 3D L3 cache, and more
• We need a 20MW 20,000-GPU-strong machine-learning supercomputer to build EU's planned digital twin of Earth
• US Air Force boots up not one but two AMD-powered supercomputers after five years of Intel Haswell CPUs

Even still, the Crusher test system itself is very fast compared to modern supercomputer standards, according to the lab, especially when you consider it is a 1.5-cabinet version of the 100-plus-cabinet Frontier.

Crusher consists of 192 Hewlett Packard Enterprise Cray server nodes connected by HPE's Slingshot interconnect, with each node containing one third-generation AMD EPYC CPU and four AMD Instinct MI250x GPUs.

To underline how impressive it is, the lab's Oak Ridge Leadership Computing Facility said Crusher is faster than the lab's previous Titan supercomputer while only being 1/100^th the size. The fact that Crusher, which takes up only 44 square feet of floor space, is faster than the 4,352-square-foot Titan is significant, considering that the now-decommissioned supercomputer was the world's fastest when it debuted in 2012.  

Oak Ridge said research teams have already seen promising results from Crusher, which is helping them prepare for when Frontier opens to all science users on Jan. 1, 2023.

For example, one team said Crusher provided a "roughly 15-fold speedup" for the Cholla astrophysical hydrodynamics code compared to the lab's current fastest supercomputer, the IBM Power-based plus GPU accelerated Summit machine, which is the fastest supercomputer in the US and has the number two spot on the Top 500 list of the world's fastest supercomputers.

Evan Schneider, assistant professor at the University of Pittsburgh and principal investigator for the Cholla code, said about 3x of that performance boost for Cholla is coming from hardware while software improvements are multiplying the impact of the hardware by 5x.

Another team, which operates the labs' NuCCOR nuclear physics code, said one AMD Instinct MI250X GPU was up to 8x faster than one of Summit's Nvidia V100 GPUs for the program, which is used to calculate the properties of atomic nuclei and their reactions.

"Crusher is the latest in a long line of test and development systems we have deployed for early users of OLCF platforms and is easily the most powerful of these we have ever provided," said Bronson Messer, director of science for Oak Ridge's supercomputer center. "The results these code teams are realizing on the machine are very encouraging as we look toward the dawn of the exascale era with Frontier."

The US is preparing another exascale supercomputer, the long-delayed Aurora, at the Department of Energy's Argonne National Laboratory, and it could be faster than Frontier when it debuts in late 2022.

Intel, which is providing the cluster's CPU and GPU innards, said last fall that Aurora will provide more than 2 exaflops, though The Next Platform suspects that may be for peak performance. ®