Nvidia intends to introduce a specialized AI chip for the Chinese market, potentially as early as the first quarter of 2024. Reports from Chinese media indicate that Nvidia's Chinese customers are seeking alternatives because of the significantly reduced performance of the China-focused chip.
Chinese media outlets, such as Chinastarmarket, reported that Nvidia is collaborating with key server suppliers in China to test the HGX H20, customized for China to circumvent US export restrictions. Orders for this AI chip are anticipated to open no earlier than January 2024. At its last earnings call, Nvidia disclosed the AI chip plan compliance with export restrictions.
However, previous reports hinted at a significant reduction in the computing power of three reduced compliance AI chips, including the H20. For instance, the HGX H20 faces limitations in bandwidth and computational speed, which are estimated to have only about 20% of the computing power of the H100.
Chinese media reported that Nvidia plans to improve the HBM and NVLink interconnection configurations with the H20 to deal with scenarios with LLMs with fewer parameters.Nvidia might also enable customers to stack multiple AI chips designed for China to compensate for the lower computational power, potentially resulting in power consumption that is twice or even higher than that of the H100.
On the other hand, LLM developers in China are seeking alternatives to Nvidia's offerings and have shifted their focus to Huawei's Ascend AI chips. According to Reuters, Baidu has opted for Huawei's AI chips as a substitute for Nvidia's. Liu Qingfeng, the founder and chairman of Iflytek, remarked that Huawei's Ascend 910B matches the Nvidia A100 in performance.
However, the challenge lies in the supply capacity of the Ascend 910B, even if the chip can substitute for the upcoming Nvidia China-focused AI chip. Sources told Chinese media that the slowly growing production capacity of the Ascend 910B cannot meet the surging demands for AI chips in China.
