No Effect from US Sanctions...? Chinese Startup Successfully Upgrades Data Center

GPU Startup Moore Thread
Founded by Former Nvidia China Business Manager

Despite the US semiconductor export controls against China, a startup has succeeded in upgrading its data center, South China Morning Post (SCMP) reported on the 5th.

According to the report, Chinese graphics processing unit (GPU) startup Moore Threads announced on the 3rd that it had "upgraded its artificial intelligence (AI) data center solution KUAE (誇娥)" and "increased the number of GPUs that can be connected to a single computing cluster to 10,000." KUAE is a data center server composed of eight MTT S4000 GPUs, a large language model (LLM) GPU independently developed by Moore Threads.

[Image source=Moore Thread homepage capture]

Moore Threads, founded in 2020 by Zhang Qianzhong, former general manager of Nvidia China, has emerged as a leader in the GPU sector in China by consecutively launching GPU chips and gaming graphics cards. However, after being blacklisted by the US in October last year, it faced difficulties in finding wafer foundry (semiconductor contract manufacturing) partners, and access to US electronic design automation software was also restricted, ultimately leading to layoffs.

Despite the impact of these US export controls, Moore Threads succeeded in launching new products and opening a computing center. Recently, it signed contracts related to building computing cluster projects with three major Chinese state-owned telecom companies, including China Mobile and China Unicom. Local financial media Jiemian News also reported that Moore Threads separately completed new funding of up to 2.5 billion yuan (approximately 470 billion KRW).

SCMP evaluated, "(Moore Threads') recent upgrade is the latest example showing that despite the US restricting exports of advanced AI chips to China over national security concerns, Chinese companies are striving to catch up with US AI capabilities."

Moore Threads explained on its website that "GPUs can be interconnected using a technology called 'MTLink' developed by Moore Threads, and connecting multiple GPUs can form a more powerful system." However, in terms of product performance, it is still considered to lag behind Nvidia. According to specifications released by Nvidia, the computing performance of the Nvidia A100 GPU is 312 teraflops, three times that of Moore Threads' MTT S4000.

Meanwhile, AI chip leader Nvidia is expected to ship more than one million AI chips 'H20' to the Chinese market this year despite US semiconductor export controls against China, generating sales of $12 billion (approximately 16.58 trillion KRW). The H20 is a product Nvidia created by reducing the computing power to one-fifth of its main AI chip 'H100' to circumvent US export restrictions to China. The New American Security Center estimates that the volume of banned chips flowing into China by evading US surveillance could reach 12,500 units annually.