Huawei is promoting its Supernode 384, a high-performance AI computing system using Ascend processors, as a competitor to Nvidia’s NVL72 for data centers.

The Supernode 384 serves as the basis of Huawei’s CloudMatrix 384 system, integrating 384 Ascend AI processors in 16 cabinets.

The CloudMatrix 384 has been implemented in Huawei data centers in Anhui, Inner Mongolia, and Guizhou provinces.

While Huawei has not specified the Ascend chip variants used, industry experts suggest it likely employs the Ascend 910C AI processor.

🔗 Source: South China Morning Post

🧠 Food for thought

1️⃣ Sanctions-driven innovation reshapes the AI chip landscape

Huawei’s Supernode 384 architecture represents a direct response to U.S. technology restrictions, showcasing how trade barriers can accelerate domestic innovation.

Despite being cut off from advanced semiconductor manufacturing, Huawei reported a 10% revenue increase in 2023 to 704.2 billion yuan ($97.4 billion) with net profits more than doubling to 87 billion yuan ($12 billion)1.

This architectural approach, emphasizing scale over individual chip performance, reflects Huawei’s strategic adaptation to manufacturing limitations by focusing on system-level engineering and networking capabilities2.

The CloudMatrix 384’s optical interconnection technology delivers approximately 300 petaflops of computing power, demonstrating how Chinese companies are developing alternative technological paths when denied access to Western components3.

Rather than attempting to match Nvidia’s per-chip efficiency, Huawei compensates through scale and communication bandwidth, leveraging China’s strengths in networking infrastructure while working around chip manufacturing constraints.

2️⃣ The diverging architectural approaches in AI infrastructure

Huawei’s Supernode 384 and Nvidia’s NVL72 represent fundamentally different approaches to solving AI computation bottlenecks, reflecting contrasting design philosophies.

While Nvidia focuses on maximizing per-chip performance with its GB200 Blackwell architecture, Huawei uses a more distributed approach with 384 Ascend 910C processors connected through an all-optical mesh network that enhances communication bandwidth by 15 times4.

The Chinese company’s peer-to-peer computing architecture specifically targets Mixture-of-Experts (MoE) models—a machine learning technique increasingly popular for large language models that requires significant inter-chip communication.

Recent Huawei developments