Japan’s Riken institute has ripped through prior records on computing speed, with its brand-new Fugaku supercomputer performing 2.8 times more calculations per second in a biannual speed ranking than the previous record holder, the IBM system at Oak Ridge National Laboratory.
Summit, the Oak Ridge computer, is now second place. Another IBM system at Lawrence Livermore National Laboratory as well as two Chinese supercomputers also slid one place in the rankings, run by the Top500 project, thanks to Fugaku.
Per the New York Times, the total bill for the Riken institute’s six-year plan for Fugaku came in at around $1 billion. It is based on the ARM architecture more commonly associated with cell phones and other mobile devices, the first time an ARM-derived system has hit the number one slot and a notable divergence from prior supercomputers that have mostly relied on chips derived from Intel or AMD designs. According to Anandtech, Fugaku has 7.3 million cores, consumes 28 megawatts of power, and was able to perform at 415 petaflops, the unit of measurement for a quadrillion floating point operations per second. (Fugaku has a maximum theoretical maximum performance of nearly 514 petaflops.)
While Fugaku dominated in raw computing speed, it also took the top three slots in tests designed to rank its capacity applications for industrial, artificial intelligence, and big data analytics, according to Kyodo News. That’s another record, Riken told the network, as no prior system has held all four spots at once. It is expected to be operational by April 2021 though it is already assisting some medical research on the novel coronavirus pandemic, Kyodo News reported.
Fugaku “is the culmination of almost 10 years of investment and work,” Arm senior vice president Christopher Bergey told the Times. “It’s a pretty exciting time.”
Fugaku may not rule for long, however. The Department of Energy is building another supercomputer named Frontier with Cray Inc., which the agency says will be able to process 1.5 exaflops per second. As of May 2019 that would have made it as powerful as the next 160 fastest supercomputers combined (a number that may no longer be accurate, as Top500 says Fugaku was responsible for the majority of a roughly 35 percent increase in global supercomputing capacity in the last six months on its own). Frontier is scheduled to go online in 2021. The DOE is also building another supercomputer, Aurora, in partnership with Intel, which is slated to beat Frontier into service and will be the first exascale unit.
China has three exascale projects of its own. University of Tennessee electrical engineering and computer science professor Jack Dongarra told IEEE Spectrum he didn’t expect either the U.S. or Chinese exascale projects to actually go online in 2021, but that Chinese rivals may give the competition a run for its money.
“China is very aggressive in high-performance computing,” Dongarra told the magazine. “Back in 2001, the Top 500 list had no Chinese machines. Today they’re dominant.”