Tencent has made a major upgrade to its HPC network, known as Xingmai, enhancing AI capabilities by up to 60% in network communications and up to 20% in LLM training.
The upgrade, reported by the South China Morning Post, reflects a broader effort by Chinese tech giants to become more technologically independent at a time when strict U.S. export controls limit their access to advanced processors such as Nvidia’s H100.
According to the report, the Xingmai 2.0 network will greatly improve the efficiency of communication between computing clusters. Previously, the communication time between clusters was too long, which led to the GPU’s capabilities being underutilized. By upgrading its network, Tencent has not only improved the communication process, but also reduced costs, which is beneficial for the company.
In-house Development LLM
The upgrade comes as Chinese companies seek creative ways to reduce their reliance on foreign technology and get around U.S. export bans. In a previous example, Huawei improved the performance of its AI chips by adding vector units to each core and increasing clock speeds to compensate for a reduction in active AI cores.
Unlike its US competitors, which focused on increasing spending and acquiring cutting-edge semiconductors, Tencent achieved performance improvements by optimizing its existing facilities: its upgraded network now supports more than 100,000 GPUs in a single cluster, doubling its original capacity and reducing problem-resolution time from days to minutes.
The Shenzhen-based company’s AI advancements aren’t limited to improved infrastructure: According to the South China Morning Post, Tencent is aggressively promoting LLM for its homegrown enterprise applications, and even offering services to help other companies develop their own AI models.
China’s AI industry is currently locked in a price war, with giants like Tencent, Bytedance, Baidu and Alibaba slashing costs to sell cheaper than their Western competitors. Tencent recently followed similar moves by its competitors, making the lite version of Hunyuan LLM free and slashing the price of the standard version.
