Alibaba Group Holding, and China Telecom, have launched a new AI data center in Shaoguan city located in the Guangdong province, South China. What makes this data center unique is that instead of using foreign GPUs, it is relying entirely on domestically manufactured chips.
This reflects a broader shift in China’s AI strategy as it is refusing to let restrictions on access to advanced semiconductors from companies such as NVIDIA be an impediment to the growth of its digital infrastructure sector. Rather than matching the most advanced foreign chips individually, Chinese companies are focusing on building larger clusters with domestic hardware and using network design to improve overall performance.
The 10,000 card computing cluster is the first deployment of Zhenwu chips at this scale in the Greater Bay Area while the system uses a networking architecture with 4-microsecond latency, allowing 10,000 chips to function as a single system. According to a report from CNBC, this increases training and inference efficiency by 30 percent, while single-chip throughput is nearly 10 times higher.
The new cluster showed China’s advanced computing power was “moving from high-end performance breakthroughs to large-scale industrial implementation”, Alibaba Cloud said in its statement on Tuesday, as reported by South China Morning Post.
China Telecom will own and operate the facility, which is designed to support AI training and inference workloads for models.
The Shaoguan project follows the activation in late March of China’s first 10,000-card AI cluster based on Huawei Technologies’s Ascend 910C chips in Shenzhen. The Shenzhen installation has 11,000 petaflops of computing capacity and has been linked with an existing 3,000-petaflop cluster that went online last year.
China’s AI sector is moving from “hardware replacement” to “software collaboration” as government and city administration systems are deploying domestic computing infrastructure more quickly than other sectors because of security requirements.
The new cluster is already being used in healthcare and manufacturing which enables small and medium-sized businesses to rent computing power through China Telecom’s platform on an hourly basis or by chip usage.
Alibaba and China Telecom stated that the Shaoguan site could eventually expand to 100,000 chips. China has also made large-scale AI infrastructure a national priority in its latest five-year plan, while local governments including Shenzhen and Shanghai are developing their own domestic computing networks.