On March 21, during the Huawei China Partner Conference 2026, Huawei made a significant announcement by introducing the AI training and inference acceleration card, Atlas 350. This cutting-edge card is powered by the Ascend 950PR processor, marking a substantial leap forward in computing technology. The chip boasts remarkable enhancements in low-precision data formats, vector computing capabilities, interconnection bandwidth, and its proprietary high-bandwidth memory (HBM).

Notably, the computing prowess of a single Atlas 350 card surpasses that of NVIDIA's H20 by an impressive 2.87 times. As the sole inference product in China capable of supporting FP4 low-precision computing, it stands out with an HBM capacity of 112GB. This translates to a 60% boost in multimodal generation speed and a fourfold improvement in memory access efficiency for small operators, thanks to optimized memory access granularity.

In a collaborative effort, seven key partners, including Kunlun and Higon, simultaneously launched complete machine products based on the Atlas 350. This strategic move aims to co-create industry-tailored inference solutions, further solidifying Huawei's position at the forefront of computing innovation.