Huawei's chip architecture for AI datacenters. CloudMatrix 384 (CM384) employs 384 Ascend 910C chip modules, each containing two processors that connect to four HBM memory banks. CM384 is China's alternative to NVIDIA's NVL72 (see
NVLink).
CM384 achieves 300 petaFLOPS compared to 180 for NVL72 but uses four times as much power. It also uses optical connections between modules, and the optical transceivers are the cause of the greater power consumtion. However, over the years, China has expanded its power grid with nuclear, solar and hydroelectric generation and electricity is less expensive than the U.S.
Replacing NVIDIA
Although less efficient, the CM384 system provides greater performance than NVIDIA H20 GPU clusters. The H20 is a scaled down version of NVIDIA's H100 GPU and is the most advanced AI chip allowed for sale to China. See
Ascend chip.