China Telecom, a telecom giant in the country has introduced two LLMs which were trained by using domestically produced chips. This breakthrough for the company has occurred during the increase in US export restrictions on high-end semiconductors, specifically limiting access to companies like Nvidia which produce advanced processors.
One of the model is named TeleChia2-115B. This model consists of 100 billion parameters. The second model remains unnamed but contains a big number of 1 trillion parameters.
These models were trained by tens of thousands of Chinese-made chips. The achievement highlights China’s increasing ability to independently work on AI technologies and advancements.
China Telecom demonstrates China’s efforts towards AI chip autonomy
The achievement highlights China’s increasing ability to independently work on AI technologies and advancements, at a time when the US imposed restrictions on the exportation of its chips to the Asian country.
These strict export rules have made it difficult for China to access high-end processors from companies like Nvidia and other foreign firms.
According to the company’s AI Institute, which shared a statement on WeChat, this development is a huge accomplishment demonstrating China’s capability to autonomously train large language models (LLMs) also signaling a new phase of innovation and self-reliance in the AI technology.
The company has also hyped its models describing their scale as “remarkable” boasting a trillion parameters which simply refers to the variables that assist a model in learning during its training. It therefore follows that, the more the parameters a model has, the more complicated and powerful the AI model becomes.
According to AINews, Chinese firms want to keep pace with global leaders in AI technology outside the country despite the export restrictions by the US.
The restrictions on Nvidia’s A100 and H100, the company’s latest and powerful AI chips have forced China to look elsewhere for alternatives. This has also seen Chinese firms creating their own processors to limit their reliance on US technologies. For instance, the TeleChat2-115B model has about 100 billion parameters, which makes it perform as well as mainstream platforms.
While China Telecom did not specify the company that supplied the domestically-designed chips used to train its models, Huawei’s Ascend chips play a key part in the country’s AI plans.
Local collaborations are helping firms like China Telecom serve the domestic market
Huawei, which has also faced US sanctions in recent years is also bolstering its efforts towards self-reliance in the AI industry. Not long ago, Huawei started testing its latest AI processor known as the Ascend 910C. Potential clients in China are already on the queue for the processor.
Big Chinese server firms, together with internet companies that have used Nvidia’s chips before are currently testing the performance of Ascend 910C, which is one of the few alternatives to Nvidia hardware.
Ascend processors are seen as one of China’s viable alternatives to Nvidia’s chips and will limit the country’s reliance on US.
Apart from Huawei, China Telecom is also collaborating with other local chipmakers like Cambrion, which is a startup that specializes in AI processors, reflecting China’s broader ambitions of producing homegrown AI solutions, further cautioning itself from the effects of the US export controls.
The country is gradually reducing its dependence on foreign hardware, particularly from Nvidia, whose GPUs are highly sought after.
The US restrictions have also led to a thriving black market for US AI hardware as they are smuggled into the country via different means.
It has also emerged that Chinese firms reportedly gain access to cutting-edge AI capabilities and restricted US Chips through Amazon cloud services or its competitors.
Reuters recently revealed they reviewed over 50 papers in the past year obtained from an open Chinese database. The documents showed that over 11 Chinese companies gained access to restricted US technologies or cloud amenities.
From these, four companies clearly indicated their involvement with Amazon, stating that they access the facilities through Chinese mediators and not directly from AWS.