Huawei AI Supercomputing Server

Huawei's AI supercomputing servers, including the Atlas 950 and 960 SuperPoD, deliver world-leading AI performance with massive scalability and advanced interconnect architecture.

Overview

Huawei has developed a series of AI supercomputing solutions designed for large-scale AI model training and high-concurrency inference. Key products include the Atlas 950 SuperPoD, Atlas 960 SuperPoD, CloudMatrix 384, and the TaiShan 950 SuperPoD. These systems are engineered to handle trillion-parameter models and multi-trillion token datasets, addressing the surging compute demands of generative and agentic AI applications.

Atlas 950 and 960 SuperPoDThe Atlas 950 SuperPoD features a "cluster + SuperPoD" architecture powered by Huawei's Ascend 950 AI chips. Each cabinet integrates 64 NPUs, and the system can scale up to 8,192 NPUs per SuperPoD, delivering 8 exaFLOPS FP8 and 16 exaFLOPS FP16 performance with a total interconnect bandwidth of 16. 3 PB/s.

The Atlas 960 SuperPoD further expands this scale, incorporating 15,488 Ascend 960 chips across 220 cabinets, achieving 30 EFLOPS FP8 and 60 EFLOPS FP4, with 4,460 TB of memory and 34 PB/s interconnect bandwidth. These clusters are designed to operate as a single logical machine using Huawei's UnifiedBus interconnect, enabling thousands of compute nodes to function cohesively with microsecond-level latency.

Cloud

Matrix 384The CloudMatrix 384 is a large-scale AI cluster built with 384 Ascend 910C AI processors, designed for foundation model training. It provides a domestic alternative to NVIDIA GPUs, offering high scalability and performance for China's AI infrastructure needs. This system emphasizes both compute power and interconnect efficiency, supporting large-scale AI workloads in cloud environments.

Tai

Shan SeriesHuawei's TaiShan 950 SuperPoD and other TaiShan servers (200 and 500 series) target general-purpose and enterprise computing.

The Tai

Shan 950 delivers hundred-nanosecond latency, terabyte-level bandwidth, and memory pooling, enabling efficient cross-node data transfer and reducing costly data migration. These systems are particularly suited for replacing legacy mainframes and mid-range servers in sectors like finance, supporting distributed databases such as GaussDB. Architecture and InnovationHuawei's SuperPoD architecture allows thousands of AI chips to operate as a single logical computer, effectively “thinking” as one system.

The Unified

Bus protocol addresses long-range communication reliability and bandwidth-latency challenges, enabling high-speed interconnects across hundreds of cabinets. This architecture supports both AI-specific workloads and general-purpose computing, providing flexibility for diverse enterprise applications.

Competitive Positioning

Huawei positions these servers as alternatives to NVIDIA's high-end AI systems, such as the Vera Rubin NVL72.

While NVIDI

A maintains higher compute density per rack, Huawei's systems offer greater total computing power and scalability, with clusters capable of supporting 500,000 to 1 million AI chips in combined SuperClusters. This makes Huawei a significant player in the global AI infrastructure market, particularly for large-scale AI model training and enterprise deployment.

Applications

These A

I supercomputing servers are designed for: Large-scale AI model training (trillion-parameter models)High-concurrency inference for agentic AIEnterprise computing and mainframe replacementCloud and carrier AI infrastructure Huawei's open-source and collaborative approach ensures developers can leverage these systems for AI innovation and industrial transformation. In summary, Huawei's AI supercomputing servers combine massive compute power, advanced interconnects, and scalable architecture, positioning them as a competitive alternative in the global AI infrastructure landscape.