China Telecom unveils Chinese-made supercomputer for AI Training in Wuhan
China Telecom has reportedly unveiled a supercomputer dedicated to AI training, utilizing entirely Chinese-made components and technology. The machine, located in Wuhan at the Central Intelligent Computing Center, is said to have the capability to train Large Language Models (LLMs) with trillions of parameters, potentially rivaling the Frontier supercomputer based in California, officially recognized as the world’s fastest supercomputer by the Top 500 list.
According to reports from Chinese news outlet ITHome, the supercomputer incorporates an “advanced independent liquid-cooled intelligent computing cluster,” circulating from the chip to the server and cabinet, providing a Power Usage Effectiveness (PUE) of under 1.15 in all scenarios.
China Telecom has reportedly established multiple multi-point intelligent computing centers across the country, boasting a total intelligent compute power of 5,000 petaflops. While China does not participate in the Top 500 competition, making reliable performance figures challenging to obtain, it has been actively developing high-performance computing infrastructure.
In 2022, the Chinese exascale supercomputer OceanLight was shortlisted for the Gordon Bell Prize, competing directly with Frontier. Dr. David Kahaner’s research indicated that the Chinese government was working on ten exascale systems, although specific details about these systems were limited.
Recent reports also highlight the unveiling of additional supercomputers in China, such as the Tianhe Xingyi at the National Supercomputing Center in Guangzhou and a public AI computing platform in Beijing, operated by state-backed Beijing Energy Holding. The latter is expected to provide 500 petaflops of computing power in its initial phase, increasing to 1,500 petaflops in the second phase, scheduled to go online in the first quarter of 2024.
China has been in the news recently for its digital transformation milestones. In November, China launched the world’s fastest 1.2 TBPS optical backbone, made progress with its commercial underwater data center development and reportedly spent up to $4o billion in chip making equipment in 2023.
