A senior Chinese tech executive admits that using mainland-produced semiconductors could delay the development of artificial intelligence (AI) models by three months, but insisted that his company would continue to use these chips and push improved processes.
Liu Qingfeng, chairman and founder of Chinese voice recognition solution provider iFlytek, said that the extended AI model development process was due to the need for additional computing resources, compared with using Nvidia chips and their mature software ecosystem.
In a statement posted on the company’s official channel on WeChat, Liu said iFlytek was likely “the only Chinese AI model developer that insisted on using domestic chips”, including the Ascend 910B developed by Huawei Technologies, “to avoid the risk of losing access to imported chips”.
However, the company, based in Hefei, Anhui province, has been barred from purchasing advanced US chips since it was added to the US Entity List in October 2019.
According to Liu, when benchmarked against the A800 by Nvidia, the training efficiency of the Ascend 910B has increased to 73 per cent, up from 25 per cent at the end of last year.
