At the Huawei China Partner Conference 2026, Huawei proudly introduced its latest AI training and inference accelerator card, the Atlas 350, which is driven by the cutting-edge Ascend 950PR processor. In comparison to its forerunner, the Ascend 950PR has undergone substantial enhancements in several key areas, including low-precision data format handling, vector computing capabilities, interconnect bandwidth, and its proprietary HBM technology. Remarkably, the single-card computing power of the Ascend 950PR is 2.87 times greater than that of NVIDIA's H20, positioning it as the sole inference product in China capable of supporting FP4 low-precision computing. Additionally, it accelerates multimodal generation speed by 60% and boosts memory access efficiency for small operators by a factor of four.
