What are the Huawei AI server series

Huawei's leading AI server models include the Atlas 950 and 960 SuperPoD servers, powered by Ascend 950/950PR AI chips, designed for ultra-large-scale AI workloads and LLM inference. Atlas 950 and 960 SuperPoD ServersHuawei's Atlas 950 and 960 SuperPoD servers are next-generation rack-scale AI clusters aimed at competing with NVIDIA's high-end AI systems. The Atlas 950 SuperPoD features 8,192 Ascend 950 AI chips, delivering 8 EFLOPS FP8 and 16 EFLOPS FP16 performance, with a total interconnect bandwidth of 16.

3 PB/s. These servers can be combined into the Atlas 950 SuperCluster, scaling up to 524,288 Ascend 950 chips, supporting between 500,000 and 1 million processors, making them among the largest AI compute clusters globally. The Atlas 960 is a further iteration designed for extreme scalability and high-performance AI workloads.

CloudMatrix384 ArchitectureThe CloudMatrix384 is Huawei's next-generation AI datacenter architecture, integrating 384 Ascend 910 NPUs and 192 Kunpeng CPUs with an ultra-high-bandwidth Unified Bus (UB) network. This architecture supports all-to-all communication, dynamic resource pooling, and optimized performance for large-scale mixture-of-experts (MoE) models. CloudMatrix384 also incorporates CloudMatrix-Infer, a solution for LLM serving with peer-to-peer scaling, expert parallelism, and INT8 quantization for efficient inference.

Ascend A

I ChipsHuawei's AI servers are powered by the Ascend 950 and 950PR chips. The Ascend 950PR delivers 1. 56 PFLOPS FP4 performance, with 112 GB HiBL 1.

0 memory and 1. 4 TB/s bandwidth, optimized for large language model inference. These chips use a monolithic die design, achieving high compute density and efficiency, and are compatible with CUDA-like software stacks to facilitate migration from other platforms.

The Ascend 950 series provides the backbone for Huawei's SuperPoD and CloudMatrix servers, enabling multi-petaflop AI training and inference.

AI Ecosystem and Software Integration

Huawei A

I servers integrate with ModelEngine and Ascend AI Cloud services, offering zero-code model deployment, fine-grained resource scheduling, and high utilization of xPU resources. The ecosystem supports trillion-parameter model training, multi-modal data processing, and knowledge memory platforms for ultra-scale inference, ensuring high throughput and low latency for enterprise AI applications. SummaryHuawei's AI server lineup, including Atlas 950/960 SuperPoD, CloudMatrix384, and Ascend 950/950PR chips, is designed for large-scale AI training and inference, particularly for LLMs and multi-petaflop workloads.

These servers combine massive compute power, high-bandwidth interconnects, and advanced software integration, positioning Huawei as a major competitor in the global AI infrastructure market.