• TWEAK
• Software

Chinese AI company trains LLMs at Huawei

Translate from : Kinesisk AI-firma træner LLMs på Huawei

Chinese artificial intelligence (AI) company iFlyTek claims to have the only large language models (LLMs) trained entirely on a homegrown computing platform, co-developed with Huawei Technologies, marking another step in the industry's quest for independence in the face of US restrictions.

While most of China's LLMs are supported by computing platforms that include American components such as graphics processors (GPUs) from Nvidia, iFlyTek is determined to train its LLMs on "self-developed, controllable" infrastructure, at a time when Washington has largely cut off the country's access to the most advanced chips and AI know-how, according to Liu Qingfeng, the company's chairman and president.

"How far we can go depends on whether we have self-developed, controllable basic capabilities to support [AI development]," Liu said Thursday at a ceremony where iFlyTek presented its updated Xinghuo model. In his keynote speech, Liu pointed out that US restrictions on technologies that contribute to the development of LLMs, the basis of AI products such as OpenAI's ChatGPT, underscore the need for China to develop its own computing infrastructure for AI training.

US sanctions have limited exports to China of certain technologies, most notably the advanced Nvidia GPUs that have helped drive the AI boom in recent years. Liu told the audience that the company's Xinghuo models were trained on a computing platform called Feixing Yihao, which is based on Huawei's Ascend computing solutions.

Huawei's Ascend AI chips have quickly become popular in China. The performance of the Ascend 910B chipset has been found in certain tests to deliver between 80 and 120 percent of Nvidia's A100 when training LLMs, said Wang Tao, Chief Operating Officer of the Jiangsu Kunpeng Ecosystem Innovation Centre, on the sidelines of the Nanjing World Semiconductor Conference this month.

Analysts have said the Ascend solution is China's best attempt yet to develop homegrown AI infrastructure. There is also increasing support from state-backed companies as they place orders for Ascend-powered projects, said Li Yangwei, a Beijing-based technical consultant in intelligent computing.

Huawei's hardware still lacks the vibrant developer community of Nvidia's Cuda platform, the software that supports AI computing processes on the company's GPUs, according to Li. "The chance of Huawei catching up with Nvidia is minimal if they don't solve this problem," Li said.

At the unveiling of the Xinghuo 4, iFlyTek's Liu praised that the new model's capabilities can match those of the GPT-4 Turbo, one of the most advanced models from OpenAI launched more than eight months ago. Xinghuo 4 has already been used in various business scenarios, including healthcare, where it has helped doctors and patients with diagnosis and treatment, according to a presentation at the event.