This article is more than 1 year old

US fears China may have ten exascale systems by 2025

China refuses to share benchmarks, US sharpens focus on developing optimized software

The US is racing to catch up with China in supercomputing performance amid fears that the country may widen its lead in exascale computers over the next decade, according to reports.

The Frontier supercomputer at Oak Ridge National Laboratory is expected to be the first exascale system in the US once it is fully operational, but China already has two exascale systems up and running since last year, as reported on our sister site The Next Platform.

This lead may widen as the US has three exascale systems in the pipeline, while China aims to have up to 10 operational systems by 2025, says a report in the Financial times.

What is at stake here is not just national pride, because supercomputers provide the raw calculating power required for simulation of nuclear physics used for atomic weapons, but also for discovery of new materials and to drive engineering breakthroughs.

According to the FT, at this rate China will beat the US to the next big breakthrough by fielding a larger number of exascale machines that will put it in a position to "seize the high ground of computing for years to come."

Exascale systems are supercomputers capable of operating at one exaFLOPS or greater, which means being able to calculate 1018 floating point operations per second.

The US Department of Energy has been running its Exascale Computing Project for several years now, but it has been subject to delays. The first planned exascale system was to be called Aurora based on Intel's now defunct Xeon Phi processors. This has since been resurrected and will instead be built around the Sapphire Rapids chips.

The Frontier supercomputer at Oak Ridge is now the US front-runner. Based on a custom version of AMD's Epyc CPUs and Instinct GPUs, this was installed towards the end of last year and is expected to become operational in the second half of this year.

However, US fears about being left behind by China may be exaggerated. The FT quotes Dr David Kahaner, founder of analyst outfit the Asian Technology Information Program (ATIP), as saying that half of the $3.2 billion the US Department of Energy has so far plowed into the exascale project has been spent on development of software to take advantage of exascale architectures, meaning that the US may have better optimized capabilities here.

Meanwhile, doubts have been expressed about China's two exascale systems as the Middle Kingdom has declined to make public any standard benchmark figures that would demonstrate their true performance level, as discussed by The Next Platform.

Steve Conway, senior advisor at HPC analyst outfit Hyperion Research, also pointed out that getting a computer to run at some notional benchmark speed is different from doing real-world work with it.

"You can build a supercomputer that can run tests, but if it can run actual applications at that speed, that's a different matter," he told us.

Conway said that in some respects, China is level with the US, EU, Japan, and UK in some areas of supercomputing, but noted that they do not have any domestic chip fabrication facilities more advanced than 14nm, while rivals are at 7nm and 5nm.

"But it is still very impressive progress they have made in such a short time," he added.

The two systems, the Sunway Exascale system at the National Supercomputing Center in Wuxi and the Tianhe-3 at the National University of Defense Technology (NUDT) in Hunan province, are both based on homegrown technology, and are not relying on processors from the US or elsewhere.

In the FT report, Kahaner calls for greater collaboration between China and the US, saying that the US should consider loosening its sanctions against China's leading national supercomputing center at Wuxi in the hope of getting a better look at China's exascale system, a move that seems unlikely given the level of mistrust between the two nations. ®

More about


Send us news

Other stories you might like