CloudMatrix 384
-
Re-architecting for Advantage: Huawei’s AI Stack
Huawei’s CloudMatrix 384, powered by Ascend 910C processors and the MindSpore framework, challenges Nvidia’s dominance in AI acceleration. Adopting Huawei’s ecosystem requires significant adaptation, including transitioning from PyTorch/TensorFlow to MindSpore and utilizing the CANN software stack. ModelArts, Huawei’s AI platform, supports the entire development lifecycle. While lacking the maturity of Nvidia’s ecosystem, Huawei aims to offer a viable alternative, reducing reliance on US-based technology. Transitioning requires personnel training and code re-architecting, but Huawei provides resources to facilitate the process.
-
Huawei Ascend 384 Supernode Debuts, Outperforming NVIDIA and AMD’s Previous Generation
The 2025 WAIC features Huawei’s debut of the Atlas 900 A3 SuperPoD, built on the Ascend 384 Super Node. This super-node utilizes advanced bus technology for high-bandwidth, low-latency interconnection between 384 NPUs, addressing communication bottlenecks in large AI clusters. Huawei’s CloudMatrix 384 (CM384) AI cluster, built around Ascend chips, delivers 300 PFLOPs of dense BF16 compute power, reportedly surpassing NVIDIA’s GB200 NVL72. Analysts suggest Huawei’s scaled solution surpasses current market offerings from NVIDIA and AMD.
