Cover Image for One of the largest mobile networks in the world will train its billion-parameter LLM on Huawei's AI chips, sidelining Nvidia and AMD.
Sat Oct 05 2024

One of the largest mobile networks in the world will train its billion-parameter LLM on Huawei's AI chips, sidelining Nvidia and AMD.

China Telecom turns to local chip manufacturers to circumvent the export restrictions imposed by the United States.

China Telecom, a state-owned enterprise, has announced the creation of two large language models (LLMs) that have been trained exclusively with chips produced in the country. According to a statement from the China Telecom AI Institute published on WeChat, its open-source model, TeleChat2-115B, has over 100 billion parameters, while the second model, whose name has not been disclosed, is estimated to have 1 trillion parameters. These models were developed using tens of thousands of locally manufactured chips.

The company's claim asserts that this advancement represents true self-sufficiency in training LLMs within China, an important challenge since the United States imposed strict export regulations that limit access to high-end graphics processing units, such as Nvidia's H100 and A100.

Regarding the suppliers of the chips, China Telecom has not disclosed specific information, but it is believed that Huawei has provided most, if not all, of the required devices. Huawei has positioned itself as a local alternative to Nvidia, and according to prior information from China Telecom, it is working on LLM technology using Ascend chips developed by the Shenzhen-based telecommunications company.

Recently, Huawei has begun sending samples of its new Ascend 910C processor to Chinese companies in the server and telecommunications sectors for testing. The company has also been reaching out to major Nvidia clients in China in hopes that they might consider switching at least part of their operations.

Despite the existence of a black market for high-end Nvidia GPUs in China, many companies, such as ByteDance and Alibaba, prefer to stay within the law by using lower-spec GPUs, like the Nvidia H20, to avoid legal risks and maintain access to Nvidia's support. These companies are viewing Huawei as a viable supplier for their AI needs. In fact, it was reported that ByteDance has placed an order for 100,000 Ascend processors.

In addition to collaborating with Huawei, China Telecom is also exploring the possibility of incorporating hardware from Cambricon, a new local AI chip startup, to further diversify its chip supply.