2nd December 2023 – (Beijing) China has disclosed one of its high-performing supercomputers, built in the shadows of U.S. sanctions. The newly unveiled Sunway supercomputer, powered by an upgraded indigenous chip, asserts its prowess as the world’s second most powerful system, surpassed only by the Frontier supercomputer housed at the US Department of Energy’s Oak Ridge National Laboratory.
A Chinese computer scientist, who chose to remain anonymous, asserted that the new Sunway is not the most powerful supercomputer in China at present. However, its details, presented at the Supercomputing 2023 (SC23) conference in Denver, U.S., earlier this month, provide an intriguing glimpse into how China has managed to bypass US sanctions and construct its own state-of-the-art supercomputers.
The home-grown chips utilised in the new Sunway, while not as potent as their US counterparts, have been compensated for by the Chinese scientists who have amplified the number of cores in the machine. This strategy, however, comes at the cost of increased energy consumption, size, and operational expenses.
However, the outcome of these measures to circumvent U.S. sanctions is a rapid and formidable machine. The supercomputer exhibits a speed of a billion operations per second, measured as 5 exaflops under the “HPL mixed precision” benchmark, a technical standard for assessing the performance of supercomputers. This score falls just short of the Frontier’s performance of 9.95 exaflops. The new Sunway also boasts more than 41 million CPU cores, almost five times as many as the Frontier.
In terms of computing efficiency, this unanticipated contender from China has outperformed leading supercomputers, including the Frontier. It is capable of maintaining over 85 per cent of its peak performance in regular operation, ranking it the highest among all heterogeneous systems – a prevalent supercomputing architecture – and second among all systems.
Meanwhile, China’s most powerful supercomputer remains a closely guarded secret, and the development of other supercomputing chips is also underway, informed the anonymous Chinese scientist, who is affiliated with a leading mainland university.
Known also as Sunway OceanLight, this new Sunway iteration is a progression of the Sunway TaihuLight, developed by the National Research Centre of Parallel Computer Engineering and Technology in Wuxi, in the eastern province of Jiangsu.
TaihuLight once held the coveted position of the fastest supercomputer system on the Top500 list, a prestigious ranking of the world’s 500 speediest supercomputer systems. It retained the top spot from 2016 to 2018 but dropped to 11th place in the most recent ranking.
The new generation of Sunway has been made possible primarily due to its foundation on a more potent processor, the Sunway SW26010 Pro CPU, an upgraded version of the Sunway SW26010 used in the TaihuLight system.
This new processor offers advancements in clock speed, the instruction set, and memory bandwidth, culminating in a quadrupling of performance.
On reading recent news about the supercomputer, the Chinese scientist manifested no surprise, stating to the Post, “This processor is not new. It has been used in China’s supercomputing systems for the past two or three years, but was just made known to the public, especially the Western world.”
The SW26010-Pro, a home-grown chip that aimed to bolster China’s supercomputing capabilities and self-reliance, still falls short of its most advanced rivals, according to experts interviewed.
To counteract this, in the new Sunway system, Chinese scientists have adopted a strategy where, despite the individual chips’ inferiority, they employ a larger number of cores – the basic units of computation – to enhance the computer’s performance.
The SW26010-Pro chip, compared to its previous version, has seen an increase in core clusters from four to six, with each cluster containing 64 cores, resulting in a total of 384 cores. The new Sunway is equipped with more than 100,000 chips and over 41 million cores.
However, limitations such as the on-chip memory and network bandwidth of the processor necessitate careful optimisation of the system.