Huawei Unveils NVIDIA GB200-Equivalent CloudMatrix AI Cluster, Highlighting China’s Advanced AI Hardware Innovation

Huawei is making significant advancements in the artificial intelligence (AI) sector. The company has recently unveiled its CloudMatrix rack-scale cluster, marking a major milestone in its AI capabilities.

Huawei’s AI Cluster Offers Double FP16 Performance Compared to GB200, but Comes at a High Cost

In the competitive landscape of Chinese AI firms, Huawei stands out as a strong contender against NVIDIA. The company not only excels in performance but also ensures a steady availability of its hardware offerings. While Huawei has gained recognition for its “Ascend”AI accelerator line, the recent venture into rack-scale solutions is making waves in the industry. Such advancements have reportedly raised concerns even with NVIDIA’s leadership.

Huawei tech showcase with digital displays and attendees at a bustling event exhibition.

The unveiling of the CloudMatrix 384 at the WAIC conference, held at the Shanghai World Expo Center, is noteworthy. This event showcased the CloudMatrix cluster for the first time, rebranded as the Atlas 900 A3 Superpod. The technical specifications reveal that the CloudMatrix 384 (CM384) AI cluster incorporates 384 Ascend 910C chips configured in an “all-to-all topology.”This design addresses previous architectural constraints by integrating five times the number of Ascend chips compared to NVIDIA’s GB200 series.

The performance metrics of the CloudMatrix cluster are impressive, boasting a remarkable 300 PetaFLOPS of BF16 computing. This is nearly double the capability offered by NVIDIA’s GB200 NVL72. However, this exceptional performance comes with a significant caveat: the CloudMatrix 384 is projected to consume 3.9 times the power of the GB200 NVL72, resulting in less-than-optimal performance-per-watt figures across various AI applications.

AI servers on display at a tech expo, highlighting data center advancements and technology innovations.

Pricing for the CloudMatrix 384 AI cluster is positioned at approximately $8 million. This price point is nearly threefold that of NVIDIA’s GB200 NVL72 configuration. Huawei’s motivations behind this pricing strategy appear centered on showcasing a high-performance product that leverages in-house resources, aiming to directly compete with Western alternatives. This claim has been substantiated by NVIDIA’s CEO, Jensen Huang, who acknowledges that Huawei is effectively competing with Grace Blackwell systems.

Source & Images