US regains lead in supercomputer race

The United States has regained the coveted crown of computing speed with a powerful new supercomputer in Tennessee, a milestone for technology that plays an important role in science, medicine and other fields.

Frontier, the name of the massive machine at Oak Ridge National Laboratory, was declared on Monday to be the first to demonstrate the performance of 1 quintillion operations per second – a billion billion calculations – on a set of standard tests used. by researchers to classify supercomputers. The US Department of Energy pledged US$1.8 billion several years ago to build three systems with this “ex-scale” performance, as scientists call it.

But the crown has a caveat. Some experts believe that Frontier was defeated in the exascale race by China’s two systems. The operators of these systems have not submitted test results for evaluation by the scientists who direct the so-called Top500 ranking. Experts said they suspected tensions between the United States and China could be the reason the Chinese did not submit test results.

“There are rumors that China has something,” said Jack Dongarra, professor emeritus of computer science at the University of Tennessee who helps spearhead the Top500 initiative. “There’s nothing official.”

Supercomputers have been a hotspot in international competition. The room-sized machines were built to crack codes and design weapons, but they now also play important roles in vaccine development, testing car designs and modeling climate change.

The field has been dominated by American technology for decades, but China has become a dominant force. A system called Sunway TaihuLight was ranked the fastest in the world from 2016 to 2018. China accounted for 173 systems in the most recent Top500 list, compared to 126 machines in the United States.

Japan has been a smaller but still potent competitor. A system called Fugaku in Kobe took first place in June 2020, replacing an IBM system in Oak Ridge.

Frontier returns that maximum position to the lab. The system, built by Hewlett Packard Enterprise using two types of chips from Advanced Micro Devices, was twice as fast as the Fugaku in tests used by the Top500 organization.

“It’s a proud moment for our country,” said Thomas Zacharia, director of Oak Ridge, in an online statement from an industry event in Germany. “It reminds us that we can still go after something that is bigger than ourselves.”

Construction of the system, made up of 74 cabinets weighing 3,600 kilograms each, has been hampered by the pandemic and problems obtaining components in the supply chain crisis, Zacharia said. But he predicted that Frontier would quickly make a big impact on studying Covid and helping the transition to cleaner energy sources, for example.

Chinese researchers used to participate in the classification process. But the country has adopted a more low-key profile in publicizing its supercomputers, as the United States has taken a number of steps to slow China’s technological advances — including making it harder for some Chinese companies to acquire the foreign chips that can be used to make supercomputers.

But China has been making significant progress in designing its own microprocessors, a key to advances in supercomputers. David Kahaner, an authority in the field who directs the Asia Information Technology Program, reported details last year of two exascale-class supercomputers that he says use Chinese chip technology.

One is a successor to the earlier Sunway machine, called OceanLight, according to a presentation Kahaner shared at a technical conference. The other machine, Tianhe-3, succeeds a system called Tianhe-1A that in 2010 became the first Chinese machine to take the top spot on the Top500 list.

More evidence that China has broken the exascale barrier emerged in November, when a group of 14 Chinese researchers won a prestigious award from the Association of Computing Machines, the Gordon Bell Prize, for simulating a quantum computing circuit in the new Sunway system running. at exascale speeds. The calculation work, estimated to take 10,000 years on Oak Ridge’s previous fastest supercomputer, took 304 seconds on the Chinese system, the researchers reported in a white paper.

“They kind of leaked that they had machines running at exascale levels,” said Steve Conway, an analyst at Hyperion Research. “There’s been a lot of speculation that they didn’t want to attract more US sanctions.”

Conway and other experts said they believed the chips in the new Chinese machines were made in Taiwan, which is true of Frontier’s flagship chips. China remains far behind in advanced chip manufacturing capacity, he said.

The Oak Ridge machine, in addition to helping scientists, could help vendors popularize some new products. Hewlett Packard Enterprise, which in 2019 bought supercomputer pioneer Cray, contributed networking technology called SlingShot, which had a significant impact on Frontier’s performance, Zacharia said.

And AMD contributed not just microprocessors, but a type of graphics processing chip that was sold primarily for supercomputers by a rival, Nvidia.

The same two AMD chips were selected for an exascale system called El Capitan, which is scheduled to be installed in 2023 at Lawrence Livermore National Laboratory in California.

A third exascale machine, from Argonne National Laboratory in Illinois, using three types of Intel chips, was originally scheduled for delivery in 2021. But manufacturing issues at Intel have delayed the system, which is expected later this year.

Translated by Luiz Roberto M. Gonçalves