Joshua Resnick - Fotolia
Japan's Fugaku ARM-based supercomputer is the world's most powerful in the latest Top500 list, setting a world record of 442 petaflops. But the real competition -- the first exascale systems -- will arrive in 2021.
The Top500 list, a ranking of the world's fastest HPC systems, also saw a "flattening performance curve," said Jack Dongarra, one of the academics behind the twice-a-year ranking and director of the Innovative Computing Laboratory at the University of Tennessee. He said this was a result of Moore's Law slowing down as well as a slowdown in the replacement of older systems. Under Moore's Law, the number of transistors on a processor doubles every two years or so, but that advancement is now hitting technological barriers.
"There are limits on processor technology, but there is plenty of room to grow," Dongarra said. "It's not just about the processor; we are adapting in developing better algorithms and software systems to handle the situation."
The U.S. remains on track to deliver an exascale machine in 2021 out of the Oak Ridge National Laboratory, a Cray Inc. system named Frontier, using AMD chips. An exascale system can reach one quintillion calculations -- that's a one followed by 18 zeros -- per second. An exascale is 1,000 petaflops.
There's a real need for exascale computing power, especially in crunching the massive data sets created by climate science, said Bjorn Stevens, a professor at the Max Planck Institute for Meteorology in Hamburg, Germany.
During his opening keynote at the international high-performance computing conference SC20, Stevens stressed that these massive supercomputers have to be accessible to a wide range of people.
Exascale has to be accessible
Policymakers need to "work through the consequences of their actions, of their policies" to understand, for instance, "how changes in agriculture will affect food security in Africa, or how changes with warming will affect flooding in Northern Europe," Stevens said.
"This is what we need to do," Stevens said, imploring his virtual audience of HPC vendors and research scientists at SC20 to make exascale systems broadly usable.
China has three exascale projects underway and is due to deliver at least one of them next year. Europe is also rapidly developing "pre-exascale" systems, or system designs that can scale to exascale.
"It's truly a global competition," said Steve Conway, senior advisor of HPC market dynamics at Hyperion Research. For instance, by the end of 2021, Europe is due to have eight of these large systems.
One such system in Finland is a 550 petaflop Hewlett Packard Enterprise, a European High-Performance Computing Joint Undertaking, a European Union funded project. This $160 million system is due next year.
The x86 chips still dominate supercomputing and will likely continue to do so for the next five years, Conway said.
The ARM approach
But the ARM-based approach, such as the one used in Japan, is gaining interest. ARM chips are customizable, appealing to system developers who have specialized use cases, such as AI development, Conway said.
Japan's system is at the Riken Center for Computational Science in Japan and developed with Fujitsu Limited.
On the Top500 list, the second-ranked system was IBM Power Systems at nearly 149 petaflops using its Power9 CPUs and Nvidia Tesla GPUs. It is at the Oak Ridge National Lab in Tennessee.
Third place went to Sierra supercomputer, which also uses Power9 and Nvidia GPUs, at about 95 petaflops. It is at Lawrence Livermore National Laboratory in Livermore, Calif.
In fourth came China's Sunway TaihuLight by China's National Research Center of Parallel Computer Engineering & Technology, installed at the National Supercomputing Center in Wuxi. It uses China's homegrown technology, Sunway SW26010 processors. It is a 93-petaflop system.