Chinese AI Firm iFlyTek Trains LLMs on Huawei Platform, Defying US Restrictions
iFlyTek trains its large language models (LLMs) entirely on a home-grown computer platform created with Huawei, circumventing US constraints. Most of China's LLMs use computing platforms with US components, whereas iFlyTek prioritises self-developed infrastructure. iFlyTek's Xinghuo models are trained on Huawei's Ascend computing systems, which are becoming popular in China.
This action is a key step towards self-sufficiency in the AI business, as it seeks to circumvent constraints imposed by the United States.
While the majority of China's LLMs use computational platforms that include components from US businesses such as Nvidia, iFlyTek is committed to training its models on "self-developed, controllable" infrastructure. This move comes as Washington has severely restricted China's access to advanced processors and AI knowledge. Liu Qingfeng, chairman and president of iFlyTek, underlined the need of having self-developed skills to aid AI development.
During the presentation of iFlyTek's new Xinghuo model, Liu emphasised the impact of US prohibitions on technology that contribute to LLM development. These constraints have encouraged China to prioritise the creation of its own computing infrastructure for AI training. The US sanctions have had a particularly strong impact on the export of powerful Nvidia GPUs, which have played a critical role in the recent AI growth.
iFlyTek's Xinghuo models are trained on Feixing Yihao, a computing platform built on Huawei's Ascend computing solutions. Huawei's Ascend AI chips have acquired favour in China, with the Ascend 910B chip performing admirably when training LLMs. According to some studies, it can outperform Nvidia's A100 processor by 80% to 120%. Analysts feel that the Ascend solution is China's best attempt to establish its own AI infrastructure.
However, Huawei's hardware lacks the active developer community that Nvidia's Cuda platform provides. The software that supports AI compute processes on Nvidia GPUs has played a key role in its success. Without a comparable ecosystem, Huawei faces difficulties in catching up with Nvidia's dominance in the AI business.
Despite these challenges, iFlyTek's most recent model, Xinghuo 4, has capabilities that rival OpenAI's GPT-4 Turbo, one of the most powerful models available. Xinghuo 4 has already been used in a variety of corporate settings, including healthcare, where it has aided doctors and patients in diagnosis and treatment.
iFlyTek trains its large language models (LLMs) exclusively on a home-grown computing platform developed with Huawei, defying US restrictions.
Most of China's LLMs rely on computing platforms with US components, but iFlyTek prioritises self-developed infrastructure.
iFlyTek's Xinghuo models are trained on Huawei's Ascend computing solutions, which have gained popularity in China.
Source: SCMP