China Telecom, a state-owned carrier, has made a significant stride in chip independence by developing two large language models (LLMs) trained entirely on domestically produced processors. This achievement showcases China’s advancement in AI technology and its ability to train LLMs without relying on foreign chips.
The company’s AI Institute unveiled that its open-source model, TeleChat2-115B, as well as another undisclosed model, were both trained on tens of thousands of Chinese chips. This development is crucial as the US continues to impose restrictions on China’s access to cutting-edge chips like Nvidia’s latest AI processors.
With one trillion parameters, the unnamed model revealed by China Telecom highlights the country’s self-sufficiency in training powerful AI models. Parameters are essential variables that contribute to the model’s learning process during training. In comparison, TeleChat2-115B boasts over 100 billion parameters, further emphasizing China’s dedication to AI technology innovation and independence.
Amidst export restrictions imposed by Washington, Chinese companies are striving to remain competitive in the AI sector. While the source of the chips used by China Telecom was not specified, reports suggest that Huawei’s Ascend chips may have played a role in LLM development. Despite being under US sanctions, Huawei is exploring new AI processors, offering potential alternatives to Nvidia’s high-end hardware.
Huawei has distributed samples of its Ascend 910C processor to major Chinese server companies for testing, positioning itself as a viable option for local AI firms seeking alternatives to Nvidia’s GPUs. Additionally, China Telecom is collaborating with Cambricon, a Chinese startup, to provide processors through its State Cloud service. These partnerships aim to reduce China’s dependence on foreign chips and promote the use of domestic alternatives for training AI models.
As US export regulations limit access to advanced Nvidia GPUs, Chinese AI firms are exploring various options to ensure continuous support and services. While some turn to black markets for high-powered GPUs, others opt for lower-powered alternatives to maintain official connections with Nvidia.
By embracing collaborations with local chip providers and startups, China is taking steps towards minimizing reliance on foreign technology in the AI sector. This shift signifies a new era of innovation and independence in China’s pursuit of AI advancements.
[Original article source: South China Morning Post]
