Rumor has it that the Huawei Ascend 920 will be released in the second half of the year to replace the Nvidia H20 domestically produced

Zhitongcaijing · 04/21 02:57

The Zhitong Finance App learned that according to some sources, Huawei has developed the successor to the Ascend 910CAI chip, the Shengteng 920, and plans to officially launch it in the second half of 2025. According to the latest market news, Huawei has scheduled to begin mass production of this next-generation AI semiconductor product in the second half of this year. In response, several industry experts said that the Ascend 920 chip is expected to fill the gap in the Nvidia H20 chip in the Chinese market due to the latest US restrictions.

On the 15th local time, the US Department of Commerce announced that it has issued new export license requirements for NVIDIA H20, AMDMI308 and equivalent chips exported to China. Under US export controls, manufacturers must obtain a license before they can export AI chips to China. According to reports, China is a key market for H20 chips. However, after this ban, the export restrictions on H20 chips will be in effect indefinitely. At this time, the advent of the Shengteng 920 is expected not only to break Nvidia's monopoly position in the field of AI chips, but also to provide stronger computing power support for the development of China's AI industry and promote the application and innovation of Chinese AI technology in more fields.

According to SevenTech, the specifications and performance of the Huawei Ascend 920AI chip have the following characteristics:

Process technology: The Shengteng 920 chip will be built based on SMIC's 6nm (N+3 node) process technology.

Computing power and memory bandwidth: With the HBM3 memory module, it can provide 900 Tflops of BF16 computing power and 4000 GB/s of memory bandwidth.

Architecture and training efficiency: It continues the design architecture of the Ascend 910C. It is said that the training efficiency is 30%-40% higher than the 910C, and the performance is expected to surpass the Nvidia H20.

Interface support: This new chip supports PCIe 5.0 and next-generation high-throughput interconnect protocols, which can optimize resource scheduling and cross-node collaboration.

Enhanced features: The tensor manipulation accelerator has been further enhanced and optimized for Transformer and MoE models to better meet the training needs of larger and more complex models.