China’s state-run telecommunications company, China Telecom, has successfully trained two large language models (LLMs) to operate exclusively on domestically manufactured processing units.
China’s milestone in achieving chip self-sufficiency showcases substantial advancements in the realm of artificial intelligence?
The AI Institute at the corporation announced that its open-source model, TeleChat2-115B, along with another unidentified model, has been trained on hundreds of thousands of Chinese-language chips. As tensions escalate between the US and China, the latter’s access to advanced chip technology, such as Nvidia’s pioneering AI processors, has become increasingly restricted, posing significant implications for global trade.
According to the AI Institute, this milestone showcases China’s complete autonomy in developing its own large language models, rivaling capabilities like those of OpenAI’s ChatGPT. As they see it, this milestone marks the beginning of a fresh era in the country’s quest for autonomous AI innovation.
China Telecom announced that its unnamed AI model boasts a staggering one trillion parameters. AI models utilize certain building blocks called parameters to learn during training; the more parameters a model has, the wiser and more powerful it becomes. With a staggering count of over 100 billion parameters, TeleChat2-115B stands out as its opposing counterpart.
Chinese language companies are working tirelessly to stay competitive in AI, especially considering Washington’s restrictive exports. Although China Telecom did not disclose the exact supplier of the chips, reports suggest that the company had previously mentioned Huawei’s Ascend chips in relation to its large language model development.
Huawei, operating under the constraints of US sanctions, has started testing its latest AI processor among select Chinese customers. As corporations seek alternatives to Nvidia’s premium hardware, Huawei appears poised to capitalize on the gap left behind.
According to sources, Huawei has reportedly sent trial versions of its Ascend 910C processor to China’s major server manufacturers for evaluation purposes. The chip is now being made accessible to Chinese-language web companies, many of which have historically relied on Nvidia’s technology.
Native AI firms are increasingly looking at Huawei’s Ascend chips and training tools as viable alternatives to Nvidia’s offerings. China’s State Cloud service, provided jointly by Huawei and China Telecom, now features processors from Cambricon, a pioneering Chinese startup, onboard.
China’s collaborations will likely help minimize its dependence on foreign semiconductors, such as Nvidia’s GPUs, a long-standing primary choice for training AI models.
Due to restrictive US export regulations governing access to high-end Nvidia GPUs like the A100 and H100, several Chinese AI firms have turned to illicit marketplaces to obtain these powerful processors. Despite this, some opt for lower-performance alternatives to ensure continued access to Nvidia’s official support and services.
