Nvidia's planned launch of its new H20 chip for China has been delayed, according to Reuters. Rivals such as Huawei could benefit.
Nvidia has delayed the launch of its new H20 AI chip for China until the first quarter of 2023, according to a Reuters report. The H20, the most powerful of three chips Nvidia has developed for the Chinese market, was originally scheduled to be available on 16 November.
Nvidia recently generated around a fifth of its revenue in China. The delay could affect Nvidia's efforts to maintain its market share in China against local competitors such as Huawei. Huawei recently sold numerous AI chips to Baidu.
Problems with H20, other chips not affected
According to the report, server manufacturers had problems integrating the chip, leading to a delay in the launch. The H20 is now expected to be released in February or March next year.
In addition to the H20, Nvidia is planning two other chips, the L20 and the L2. According to Reuters, the L20 will be launched according to the original schedule. There is no current information on the status of the L2 chip.
The new chips are a response to tightened US export regulations that prevented Nvidia from shipping its more advanced A800 and H800 AI chips - which were already an adapted version of the top-of-the-range A100 and H100 models affected by an earlier export ban. Meanwhile, Nvidia's most powerful gaming graphics card, the RTX 4090, is reportedly being repurposed as an AI accelerator by some Chinese companies.
Intel, AMD and UK-based Graphcore, which recently pulled out of China, are also affected by the export restrictions.
Nvidia's next AI chip is the H200 GPU
Nvidia's new H200 GPU could help offset losses in its Chinese business. Unveiled at the SC23 Supercomputing Show, the H200 is expected to nearly double AI model inference thanks to faster HBM3e memory.
Systems and cloud instances with H200 are expected to be available from the second quarter of 2024, including HGX H200 systems and in various data center environments. From 2024, H200 will also be integrated into the GH200 supercomputers, including the JUPITER supercomputer in Germany.