A Chinese supercomputer that uses domestic chips and computing architecture has been unveiled in Guangzhou but its operator refused to disclose the machine’s performance parameters.
The supercomputer, Tianhe Xinyi, has recently commenced operation, the National Supercomputing Center (NSC) in Guangzhou said at the 2023 Supercomputing Innovation Applications Conference on December 6.
Chen Zhiguang, deputy director of the NSC in Guangzhou, said Tianhe Xinyi’s general computing power is about five times of that of Tianhe-2A, which had been the world’s fastest supercomputer during the six half-year reviews done by the TOP500 List between 2013 and 2015.
He did not provide Tianhe Xinyi’s detailed information, such as its processors’ speed and interconnect and memory capacity.
Tianhe-2A has its maximum achieved performance (Rmax) listed at 61.4 quadrillion (1015) floating-point operations (petaflops) per second and theoretical peak performance (Rpeak) at 100.7 petaflops.
Xinhua reported that Tianhe Xinyi achieved a “fold increase” in speed and memory capacity when compared with Tianhe-2A.
Some Chinese commentators said the “fold increase” mentioned by state media should refer to a five-to-ten-fold increase, not just a double. They said the NSC in Guangzhou did not disclose the parameters of Tianhe Xinyi as it wants to keep the profile low and avoid triggering more sanctions from the United States.
“The prototype of Tianhe-3, an exascale supercomputer unveiled by the NSC in Tianjin in June 2018, has already commenced operations in 2019,” a Chinese commentator says in an article published on Monday. “Why is the new supercomputer in Guangzhou not named Tianhe-4?”
“If it is called Tianhe-4, people will immediately think that it is an exascale supercomputer. And such a move will make the US wary,” he suggests, adding that China does not want to face more US sanctions just for seeking higher ranking on the global supercomputer list.”
Exascale supercomputer
An exascale supercomputer refers to a computer that runs at a speed of more than one exaflop or 1,000 petaflops per second.
The world’s first exascale supercomputer is Frontier (OLCF-5), which was launched by Hewlett Packard Enterprise in June 2022. It is still the fastest supercomputer in the world with an Rpeak of 1.69 exaflops and an Rmax of 1.19 exaflops per second.
The OLCF-5 used 9,472 AMD EPYC central processing units (CPUs) and 37,888AMD Radeon Instinct MI250X graphic processing units (GPUs), which include 8.7 million processor cores.
It is about 16 times faster than Tianhe-2A, which is now the No. 14 fastest supercomputer in the world. Tianhe-2 used 48,000 Intel Xeon Phi and 32,000 Intel Xeon chips, which include 3.1 million processor cores.
In early 2015, the US government banned Intel, Nvidia and AMD from shipping high-end chips for use in Chinese supercomputers, accusing China of using its high-performance computers for research into nuclear explosive activities.
“Since the US started restricting the exports of its processors, the development of China’s supercomputers has faced a huge challenge,” a Heilongjiang-based technology columnist using the pen name of “Uncle Wan” says in an article.
“In recent years, the US has also tried to use technological blockages to suppress China’s development. It is crucial that we can continue to boost the performance of our supercomputers,” he says. “Otherwise, our development across many other sectors, such as science, defense, medical, astronomy, geography, aerospace and aviation, will be obstructed.”
More applications
In fact, China already had two exascale supercomputers: Sunway TaihuLight at the NSC in Wuxi and Tianhe-3 at the NSC in Tianjin. But the duo did not seek ranking on the Top500 List.
They were only known to have passed the one-exaflop mark in 2021 after they submitted data to judges for the Gordon Bell Prize.
One more prototype exascale system is reportedly based in the NSC in Shenzhen. Public information showed that this supercomputer will be completed in 2025.
In the press conference on December 6, the NSC in Guangzhou highlighted the use of local chips and applications in Tianhe Xinyi, instead of its speed.
“Tianhe Xinyi is application-centric and built using key technologies such as domestic advanced computing architecture, high-performance multi-core processors, high-speed interconnection networks, and large-scale memory capacity,” said Lu Yukon, director of the NSC in Guangzhou.
“It will provide users in the Guangdong-Hong Kong-Macao Greater Bay Area a strong platform and high-end computing power to achieve technological breakthroughs and industrial upgrading,” she added.
“The success of Tianhe-2A is an important sign that China’s supercomputing system has entered the world’s leading ranks,” said Liao Xiangke, chief designer of Tianhe-2A. “However, the development of local supercomputing applications has remained a challenge for China.”
Liao said the NSC in Guangzhou will launch more applications for users in different sectors, hoping to build an ecosystem.
Over the past decade, the number of Tianhe-2A’s applications has grown from about 300 to more than 300,000 at present, according to the NSC in Guangzhou.
Read: China pushes supercomputing goal despite US curbs
Follow Jeff Pao on Twitter at @jeffpao3