The US Department of Energy has opened the bidding process for construction of the country’s second and third exascale computers, machines that offer 50–100 times the performance of the US’s fastest current supercomputer. Awarded contracts could be worth as much as $1.8 billion, according to a 9 April announcement by Energy secretary Rick Perry.
Exascale computers will be capable of performing at least 1018 floating-point calculations per second, or 1000 petaflops. China currently operates the world’s top two computers, rated at 93 and 34 petaflops by the independent Top 500 list. The US’s most powerful system, Oak Ridge National Laboratory’s (ORNL’s) Titan, is ranked fifth, at 18 petaflops.
ORNL and Lawrence Livermore National Laboratory (LLNL) will house the newly announced systems, which will open in 2022 and 2023, respectively. The two new supercomputers will follow Aurora, an exascale system already in development by DOE that will be hosted at Argonne National Laboratory (ANL) beginning in 2021.
Each machine, including another potential one at ANL that could be awarded under the contract, will cost an estimated $400 million–$600 million, according to DOE. They are being procured under an unusual partnership between ORNL, ANL, and LLNL, with funding provided by DOE’s Office of Science and its National Nuclear Security Administration.
China appears on track to field its first exascale system in 2020. But Barbara Helland, associate director of advanced scientific computing research in the Office of Science, says DOE has been developing software for the exascale since 2009 and will have 15–20 scientific applications ready to go as soon as its machines are finished. “We’re going to make good use of those machines once they are there,” she says.
Historically, teams led by Cray and Intel have taken turns building most of the leading-edge supercomputers for the national laboratories. More recently, the labs have turned to industry collaborations, including Intel with Cray and Nvidia with IBM.
Helland says the ORNL machine must differ in design from Aurora’s Cray–Intel architecture. That requirement reflects the broad range of research that’s performed on DOE’s nondefense supercomputers, including using machine learning to develop new cancer drugs (see Physics Today, January 2018, page 27) and quantum Monte Carlo methods to design materials. “We like architectural diversity, and a lot of times what that means is the memory or the nodes will be configured differently to enhance certain types of calculations,” Helland says. “We know one size doesn’t fit all.”
The LLNL exascale system, which will be used primarily for nuclear weapons simulations, could be a clone of one of the two civilian ones or may use a completely different architecture, Helland says.
One of the major challenges in building exascale computers is finding ways to reduce their energy needs. Designers are improving the ratio of power consumption to performance by bringing memory closer to processors and increasing the use of more efficient graphics processing units. Compared with an earlier projection that Aurora would draw up to 200 MW of power, for example, it’s now expected to consume just 45–60 MW.