Huawei AI Chips Outperform Nvidia Processors

Huawei Technologies has significantly enhanced its AI capabilities with the introduction of its advanced data center architecture, CloudMatrix 384. This architecture enables the company’s Ascend chips to outperform Nvidia’s H800 graphics processing units (GPUs) when running DeepSeek’s R1 artificial intelligence (AI) model, as detailed in a recent technical paper.

Collaboratively authored by researchers from Huawei and the Chinese AI infrastructure start-up SiliconFlow, the paper characterizes CloudMatrix 384 as a specialized “AI supernode,” explicitly engineered to manage extensive AI workloads effectively. Huawei anticipates that this innovative infrastructure will “reshape the foundation of AI infrastructure,” offering a compelling solution for modern enterprises.

CloudMatrix 384 is composed of 384 Ascend 910C neural processing units (NPUs) and 192 Kunpeng server central processing units, interconnected through a unified bus that provides ultra-high bandwidth and low latency. Such architectural design enhances the overall efficiency of AI tasks, demonstrating advantages in processing power.

The advanced large language model (LLM) solution, referred to as CloudMatrix-Infer, showcases the prowess of this infrastructure. According to the paper, it has surpassed the performance levels of several renowned global systems in executing DeepSeek’s 671-billion-parameter R1 reasoning model.

This development reflects Huawei’s strategic response to US sanctions, illustrating the company’s efforts to overcome technological restrictions imposed by Washington. By pushing the boundaries of AI system performance, Huawei is positioning itself as a key player in the competitive AI market.

Data centers play a crucial role in supporting large-capacity servers and data-storage systems, equipped with multiple power sources and high-bandwidth internet connections. As enterprises increasingly adopt data centers to host or manage their AI projects, the significance of efficient AI architectures like CloudMatrix becomes even more pronounced.