China may already have two exascale supercomputers
China may have been the first to break the exascale barrier with two systems. Anonymous sources report this to The Next Platform. The alleged supercomputers are based on processors and accelerators developed in China.
According to the prominent Top500 list of supercomputers, Japan’s Fugaku is currently the most powerful supercomputer in the world, with a computing power of up to 0.537 exaflops. However, sources from The Next Platform report that China ran Linpack on a Sunway Oceanlite supercomputer in March. Peak performance of 1.3 exaflops was achieved. The “sustained” performance, according to the sources, amounted to 1.05 exaflops, with a power consumption of 35 megawatts. This would make China the first to have an exascale supercomputer. The country appears not to have submitted its exascale systems to the Top500 list due to political tensions.
The Sunway Oceanlite supercomputer is located at the National Supercomputing Center in Wuxi, according to The Next Platform, and uses RISC-like processors and accelerators developed in China. Compared to the previous Sunway TaihuLight from 2016, the chips in the new variant are made on smaller nodes, which should enable twice the performance per socket. The number of sockets is doubled at the same time. In total, Oceanlite must have 42 million cores and the system would be able to perform quantum simulations, among other things.
Sources report that the Wuxi team has developed a Tensor simulator for arbitrary quantum circuits, in collaboration with Chinese universities. That would “reduce the simulation sampling time of Google Sycamore to 304 seconds, compared to the previously claimed 10,000 years.” More concrete details about this are not known, but The Next Platform reports that China will share a paper with official details about that system in mid-November.
The current Chinese Sunway TaihuLight supercomputer. Source: Top500
China has also been running an exascale run on a second system, called Tianhe-3, according to sources. It would have achieved “near identical” peak performance of 1.3 exaflops, as well as providing enough computing power for sustained exascale performance.
The Next Platform does not yet have figures on the power consumption of Tianhe-3, but the machine is based on FeiTeng processors from Phytium, according to the website. They are based on Arm cores with a matrix accelerator. China reportedly developed these processors after it faced trade sanctions from the US, which prevented China from purchasing Xeon Phi processors from Intel.
Tianhe-3 and its Phytium CPUs were benchmarked in March. A month later, the company, along with Sunway, was added to the US list of Chinese companies subject to the trade ban. As a result, Phytium was no longer able to have chips produced at TSMC, DataCenterDynamic notes.
So China may have been the first to break the exascale barrier. Frontier will also be online in the US soon. That supercomputer is based on AMD processors and GPUs and is expected to reach peak performance of 1.5 exaflops, with 1.3 exaflops of sustained processing power, slightly higher than the Chinese systems. Intel’s Aurora supercomputer was also planned for this year, but has been delayed until 2022. The manufacturer announced this week that Aurora should reach 2 exaflops in computing power.