The AI chip specially developed by Nvidia for Chinese manufacturers may be delayed and will be released in the first quarter of next year at the earliest. According to previously exposed information, Nvidia's three AI chips are not "improved versions" but "shrunk versions", which are HGXH20, L20PCle and L2PCle.
Sources said that due to problems in the production of server manufacturers, the release time of H20 will be postponed to the first quarter of next year, while the remaining L20 and L2 are not yet clear.
It is reported that the H20 and other three AI chips are all modified from the H100, but the HGXH20 used for AI model training has limitations in bandwidth, computing speed, etc. In theory, the overall computing power is about 80% lower than the NVIDIA H100 GPU chip.
Nvidia said at this week’s earnings briefing that they are indeed developing specific AI chips for Chinese manufacturers, but they still have to pass U.S. regulations.