Huawei demonstrated its CloudMatrix rack-scale cluster at the WAIC conference in Shanghai World Expo Center for the first time. The Atlas 900 A3 Superpod contains 384 Ascend 910C chips arranged through all-to-all topology configuration. This system delivers 300 PetaFLOPS of BF16 computing power, which doubles the performance of NVIDIA's GB200 NVL72. The Chinese manufacturer addresses architectural limitations by deploying five times more Ascend processors than competing NVIDIA solutions. NVIDIA CEO Jensen Huang acknowledges that Huawei successfully competes with Grace Blackwell systems.
The CloudMatrix 384 consumes 3.9 times more power than GB200 NVL72, resulting in poor performance-per-watt metrics. Single units cost approximately $8 million, which represents three times the price of NVIDIA's comparable configuration. Huawei prioritizes domestic manufacturing capabilities over cost-effectiveness when competing against Western alternatives. The company advances its position within the artificial intelligence accelerator market through the Ascend product lineup.
The CloudMatrix 384 consumes 3.9 times more power than GB200 NVL72, resulting in poor performance-per-watt metrics. Single units cost approximately $8 million, which represents three times the price of NVIDIA's comparable configuration. Huawei prioritizes domestic manufacturing capabilities over cost-effectiveness when competing against Western alternatives. The company advances its position within the artificial intelligence accelerator market through the Ascend product lineup.
