Photo credit: www.techradar.com
US bans Nvidia’s H20 just as Huawei reveals its next-gen Ascend 920 AI chipAscend 920 may fill China’s AI chip gap post restrictionsNvidia faces $5.5 billion hit as Huawei seizes opportunity
On April 9, 2025, the United States government announced a new wave of restrictions on AI chip exports to China, specifically targeting Nvidia’s H20 processor. This chip, designed to comply with previous export regulations, allowed Nvidia to maintain its operations in the Chinese market despite ongoing tensions.
In a timely move, Huawei unveiled its latest innovation, the Ascend 920, during a partner conference just a day after the US announcement. This next-generation AI chip is poised to enter mass production in the latter half of 2025.
As reported by DigiTimes Asia, the Ascend 920 will utilize SMIC’s 6nm manufacturing process, boasting capabilities of up to 900 TFLOPs of BF16 computation and 4000GB/s memory bandwidth, backed by HBM3 memory. It is also equipped with PCIe 5.0 compatibility and advanced interconnect technologies to support large-scale model training.
An effective alternative
Following the restriction of Nvidia’s H20 in China, market analysts suggest Huawei’s Ascend 920 could effectively fill the void left in the market. While benchmarks are not yet available, the chip’s specifications indicate potential as a viable substitute for Nvidia’s offering, particularly for Chinese tech giants such as Tencent and ByteDance who now find themselves seeking alternatives.
The current Ascend 910C is estimated to deliver around 60% of the inference performance of Nvidia’s H100. In comparison, the Ascend 920 is reported to enhance training efficiency by 30% to 40%, with design optimizations for Transformer and Mixture of Experts models.
Before the export limitations were imposed, Nvidia’s H20 was enjoying significant market growth in China, with sales reportedly increasing by 50% each quarter. However, the new licensing requirements introduced by the U.S. Department of Commerce now put those sales at risk, leading Nvidia to anticipate a business loss of approximately $5.5 billion.
At the same conference, Huawei also introduced its AI CloudMatrix 384 Supernode solution, a rack-scale platform touted as a ‘Nuclear-level product’ that reportedly outperforms Nvidia’s GB200. Despite its higher power consumption, this is not considered a significant drawback within the Chinese market context.
You might also like
Source
www.techradar.com