AI conference spotlights Chinese GPU advances
Chinese GPU developers are grappling with manufacturing hurdles due to US export restrictions and a lack of sufficient demand, which leaves a considerable amount of computational power idle.
At the recent World Artificial Intelligence Conference in Shanghai, Chinese GPU developers seized the opportunity to showcase their products in Nvidia’s absence. Prominent companies such as Iluvatar Corex, Moore Threads, Enflame Technology, Sophgo, and Huawei’s Ascend were at the forefront, highlighting their advancements despite significant challenges in manufacturing and software ecosystems.
Enflame Technology emphasised the shift from foreign-dominated computing clusters to a mix of Chinese and foreign GPUs. The company, along with AI solutions firm Infinigence, is promoting compute resources that utilise a variety of chips from both Nvidia and Chinese manufacturers. However, US export restrictions have prevented Nvidia from selling its most advanced chips in China, and several Chinese firms, including Huawei, are struggling with manufacturing hurdles due to being blacklisted by the US.
Huawei’s booth was a major attraction, showcasing its Ascend 910B chips, which train numerous large language models in China. Meanwhile, Enflame presented its Cloudblazer T20 and T21 AI-training chips, benefiting from not being on the US trade blacklist, which allows it access to global foundries like TSMC.
Despite these efforts, Chinese GPUs still need to catch up with their global counterparts regarding performance. Nvidia remains a dominant player, with tailored chips for the Chinese market continuing to be popular. Nvidia is expected to deliver over 1 million H20 GPUs in China this year, generating $12 billion in sales. However, experts highlight that China’s in-house technology still needs to meet its substantial domestic AI demand.