The CloudMatrix 384 features a full optical mesh network with 6,912 x 800Gbps transceivers, designed to optimize internal bandwidth and reduce latency across the cluster. Spanning 16 racks—12 for computing and four for networking—the system consumes 559kW of power. Huawei claims it delivers over 300 petaFLOPS of compute performance using BF16 precision, surpassing comparable systems in throughput, with triple the memory capacity and double the bandwidth.
A Huawei spokesperson stated: “CloudMatrix 384 represents a significant step in building advanced AI infrastructure, offering unmatched performance to meet the growing demands of AI workloads in China.”
The system’s energy efficiency is lower than some alternatives, but Huawei emphasizes that competitive domestic electricity costs and energy reliability address these concerns for Chinese data centers. Priced at up to $8.2 million per unit, the CloudMatrix 384 has already been deployed in over 10 locations, including Huawei’s Wuhu data center and facilities for public and financial sector partners.
Integrated with Huawei’s proprietary software, including MindSpore and CloudMatrix-Infer, the system is tailored for the Chinese market and does not support the CUDA ecosystem widely used internationally. This focus limits its global adoption but aligns with Huawei’s strategy to prioritize localized AI infrastructure development.
The CloudMatrix 384 reflects trends in high-bandwidth optical networking and dense silicon interconnects, similar to advancements in global markets for AI training and inference workloads. By leveraging its Ascend NPUs and custom software, Huawei aims to meet the increasing demand for AI computing power within China.
The launch underscores Huawei’s commitment to advancing AI infrastructure, supporting the development of scalable and high-performance systems for domestic clients. The CloudMatrix 384 positions the company as a key player in meeting the technological needs of China’s AI-driven industries, contributing to the nation’s digital innovation goals.
