China designs extremely powerful computer vision chip

According to Xataka , the next round of US sanctions against China will make it impossible for Nvidia to sell its A800 and H800 AI chips in China. Even the most powerful GeForce RTX 4090 GPU for the consumer market will be affected. This will not only affect Nvidia now but also in the future when they may lose many Chinese customers after the ban is lifted, as Nvidia CEO Jensen Huang commented: "If China cannot buy GPUs for AI from the US, they will simply make them themselves."

Chip schematic diagram from Tsinghua University

In the short term, Nvidia’s decline in China is particularly beneficial to Huawei, which already has Ascend AI chips to try to fill the gap in the market that Nvidia will leave in about 10 days. In addition, MetaX, Alibaba, Biren Technology, Moore Threads, Innosilicon, Zhaoxin, Iluvatar CoreX, DenglinAI and Vast AI Tech are developing advanced chips for AI applications.

Recently, some reputable media outlets in Asia reported that some Chinese companies specializing in graphics and AI hardware such as MetaX or Moore Threads have recruited engineers from AMD and Nvidia to serve the tasks they are aiming for.

Things got even more interesting after a paper published in the journal Nature by a group of Chinese scientists from Tsinghua University. In the paper, the researchers detailed the architecture of a powerful and extremely fast computer vision chip, at least in theory. Graphics power aside, the scientists said the chip performed 3.7 times faster at computer vision tasks than Nvidia’s A100 chip when both were faced with the same intensive task of classifying images.

Tsinghua University is one of the most prestigious scientific institutions in China, so the content of the article is completely reasonable. The most interesting thing about this chip is that its theoretical performance is higher than Nvidia's A100 GPU.

The chips are manufactured using 180nm CMOS integration technology, a very mature lithography technique that is much less complex than the most advanced technologies currently being produced. It is difficult to determine how much their performance would improve if they were manufactured using much more advanced lithography techniques. Of course, these integrated circuits are not yet mass-produced, so they still have to overcome the challenge of leaving the lab and entering mass production.

Source link: https://thanhnien.vn/trung-quoc-thiet-ke-chip-thi-giac-may-tinh-cuc-manh-185231107111707111.htm