So China can resource internally all the computing power it needs to pursue AI development. Ten clients have adopted the CloudMatrix 384 AI server, according to the FT.
Nvidia’s GB200 NVL72 integrates 36 Grace CPUs and 72 Blackwell GPUs delivering up to 130 TB/s of bandwidth within a single rack.
The CloudMatrix 384 cluster contains 384 Ascend 910C chips – 5x as many ICs as the Nvidia GB200 NVL72 – delivering 300 PetaFLOPS of BF16 computing.
At $8 million, the price of the Huawei cluster is 3x the price of the Nvidia GB200 NVL72, while power consumption is almost 4x more, which shows that Huawei has developed the cluster to provide an all-China resource for AI development rather than a product for world markets.
However, the Ascend 910C’s successor chip, the Ascend 910D, is said to be sampling this month and to be focussed on power efficiency.
