Huawei has begun shipping its CloudMatrix 384 AI chip cluster to Chinese clients, aiming to fill the void left by US export controls on Nvidia's advanced AI chips. The CloudMatrix 384 system uses 384 Ascend 910C AI processors interconnected in a full-mesh optical network across 16 racks. This configuration delivers 300 PFLOPs of BF16 compute performance, surpassing Nvidia's GB200 NVL72.
While the CloudMatrix 384 offers substantial computing power, it consumes significantly more energy, approximately 559 kW compared to Nvidia's 145 kW. Huawei's strategy involves leveraging a large number of processors to achieve high performance, compensating for limitations in accessing cutting-edge chip manufacturing technologies. Despite the higher power consumption, Chinese firms may find Huawei's solution a viable alternative, given the restrictions on Nvidia and the relatively lower electricity costs in China. Huawei plans to ship over 800,000 units of the 910B and 910C processors in 2025.