What are Huawei s AI servers

Huawei's AI servers are high-performance computing systems designed for large-scale AI workloads, powered by Huawei's Ascend AI chips and organized into SuperPoD and SuperCluster architectures.

Overview

Huawei AI servers are specialized computing platforms built to handle massive AI workloads, including machine learning training, inference, and enterprise AI applications. They leverage Huawei's in-house Ascend AI chips and Kunpeng ARM-based processors, combined with proprietary interconnect technologies, to deliver extremely high computational performance and scalability .

Key Products

1. CloudMatrix 384

• Powered by 384 Ascend 910C chips, delivering 300 PetaFLOPS of BF16 compute performance .
• Designed as a domestic alternative to Nvidia's GB200 NVL72, providing China with self-sufficient AI computing resources .
• Uses a “supernode” architecture with high-speed interconnects to enable efficient communication between chips .
• Focused on raw computing power rather than energy efficiency, consuming nearly four times the power of comparable Nvidia systems . 2. Atlas 950 and 960 SuperPoDs
• Atlas 950 features 8,192 Ascend 950 AI chips, achieving 8 EFLOPS FP8 and 16 EFLOPS FP16 performance with 16.3 PB/s interconnect bandwidth .
• Atlas 960 scales even further, with up to 15,488 Ascend NPUs per SuperPoD .
• These servers are designed to be modular, allowing multiple SuperPoDs to form SuperClusters with 500,000 to over 1 million AI chips, making them among the largest AI compute clusters globally . 3. TaiShan 950 SuperPoD
• A general-purpose compute pod capable of handling enterprise workloads, including AI inference, databases, and application servers .
• Powered by Kunpeng 950 processors with up to 192 cores and 384 threads, supporting AI-native workloads and integration with Huawei's software ecosystem .

Technology and Ecosystem

• Ascend NPUs: Huawei's proprietary AI chips optimized for deep learning and AI workloads.
• Kunpeng CPUs: ARM-based processors for server and AI infrastructure, reducing reliance on x86 architecture .
• UnifiedBus 2.0: Huawei's high-speed interconnect protocol, enabling efficient communication across thousands of NPUs in SuperPoDs and SuperClusters .
• Software Stack: Huawei provides an AI software ecosystem including MindSpore, CANN, and openEuler, offering a full-stack alternative to Nvidia's CUDA .

Applications

Huawei AI servers are used for:

• Large-scale AI model training and inference
• Enterprise AI adoption in sectors like finance, smart cities, and energy
• High-performance computing tasks requiring massive parallel processing
• Building AI infrastructure that is independent of foreign hardware restrictions

Summary

Huawei's AI servers represent a vertically integrated approach to AI infrastructure, combining proprietary chips, high-speed interconnects, and a full software stack. From the CloudMatrix 384 to the Atlas 950/960 SuperPoDs and SuperClusters, these systems are designed to deliver unprecedented AI compute power, scalability, and independence from Western hardware, positioning Huawei as a major player in global AI infrastructure .