China Telecom has reportedly used entirely Chinese-made components and technology to build a supercomputer for AI training.
According to Chinese news outlet ITHome, the machine, based in Wuhan at the Central Intelligent Computing Center, has the ability to train LLMs with trillions of parameters. If true, the machine would rival the Frontier supercomputer, the California-based machine that has officially been recognized as the world’s fastest supercomputer by the Top500 list.
A Google-translated version of the report said the machine also includes an “advanced independent liquid-cooled intelligent computing cluster” that circulates from the chip to the server, to the cabinet and provides a PUE of under 1.15 in all scenarios.
ITHome stated that China Telecom has built a number of multi-point intelligent computing centers across the country, with a total intelligent compute power of 5,000 petaflops.
Although China doesn’t take part in the Top500 competition, meaning reliable performance figures are hard to come by, in 2022 the Chinese exascale supercomputer OceanLight was shortlisted for the Gordon Bell Prize, placing it in direct competition with Frontier.
That same year, research from Dr. David Kahaner, director of the Asian Technology Information Program, found that the Chinese government was developing ten exascale systems, although a lack of information about the systems and their specifications makes it difficult to judge just how powerful these machines will be in reality.
However, since Kahaner presented his research, Chinese media outlets have reported on the unveiling of several more supercomputers.
At the start of December 2023, Chinese state news agency Xinhua said that the National Supercomputing Center in Guangzhou, China had debuted a new domestically-developed supercomputer, the Tianhe Xingyi. No specifications were provided but the news outlet said the system outperforms the 100 petaflop Tianhe-2.
In January 2024, Beijing reportedly launched a public AI computing platform, operated by state-backed Beijing Energy Holding (BEH). Scheduled to go online in the first quarter of 2024, BEH said the platform will provide 500 petaflops of computing power in the first phase and increase that to 1,500 petaflops in phase two.