With the new UnifiedBus connection architecture introduced within the scope of HUAWEI CONNECT 2026, Huawei brings together processor, memory, storage and network components more tightly in large-scale artificial intelligence infrastructures. The company uses this infrastructure up to one million NPU It aims to support SuperCluster systems that can grow. UnifiedBus aims to reduce bottlenecks caused by data transfer, especially in the training and inference processes of large models, and constitutes the core connection technology of Huawei’s SuperPoD strategy.
With UnifiedBus, Huawei gathers the tasks undertaken by different connection protocols today under a single protocol and common memory semantics. According to the data announced by the company, the technology increases the connection bandwidth from hundreds of GB per second. to TB per second while subtracting the round trip delay from 7 microseconds to 2 microseconds downloading. The same fabric directly connects the CPU, NPU, memory and SSD units, allowing the components to establish peer-to-peer data access without the need for a central intermediary. Huawei also makes global memory addressing part of the same architecture in SuperPoD.
In the connection between cabinets, Huawei uses UnifiedBus LinkDevice 176 ports It offers 1.6 Tbit/sec bandwidth on each port. So the device total 280 Tbit/s Provides fully optical connection capacity. The UnifiedBus UBG switch, which works in larger clusters, takes the scale even higher with its fan-out capacity of up to 1,024 connections. This hardware allows Huawei to use the same connectivity approach, starting from single cabinets to clusters containing up to one million NPUs.
UnifiedBus is actually a critical piece of the broader Peerium Computing Architecture approach. With Peerium, Huawei aims to run multiple processors as a single large computer by using nested parallel processing, unified memory addressing and peer-to-peer connection methods. Here UnifiedBus provides the high-speed bus between the CPU, NPU, memory, SSD, network interfaces and switches. The company has begun using this approach in its Atlas 950 SuperPoD and SuperPoD-based SuperCluster systems, and is also testing the Atlas 960 system using NPO technology.
UnifiedBus focuses on bottleneck in AI infrastructure
Huawei’s new architecture does not only focus on increasing raw processing power, the company also connects the memory and storage side to the system via the same bus. One concrete example of this was the OceanStor M900. The system establishes single-hop access between NPUs and SSDs via UnifiedBus, while cluster-wide for KV cache. 64 PB capacity can provide. With this structure, Huawei aims to reduce the memory capacity problem faced by artificial intelligence agents during inference, especially those that use long contexts and execute many operations in a row.
We see the equivalent of this approach on the scale side in the Atlas and TaiShan systems. Huawei launches TaiShan 950 SuperPoD using all-optical UnifiedBus connectivity to 4,096 NPU It explains that it offers support for up to 256 TB and creates a combined memory pool of up to 256 TB. Atlas 960E SuperPoD uses NPO connection technology and the company positions this system to accelerate the training and inference processes of models at the 10 trillion parameter level. The Atlas 950 and Atlas 960 SuperPoD systems, announced by Huawei in 2025, reached 8,192 and 15,488 Ascend NPU scales, respectively.
Huawei brings its UnifiedBus ecosystem to the cloud
Huawei does not limit its UnifiedBus infrastructure to data center hardware. The company’s AI Cluster Service supports clusters with more than 100,000 accelerators and processes up to 5 million tokens per second in a 1,000-card configuration. Huawei Cloud also announced that it is making its new AI Cluster Service version available globally on September 18. Thus, the company places the UnifiedBus architecture at the center of a common artificial intelligence infrastructure that extends from SuperPoD hardware to cloud services.
In fact, the real importance of UnifiedBus lies in Huawei’s plan to scale processing, memory, storage and network resources under the same architecture, rather than offering a faster connection standard on its own. The company will launch Ascend 960DT in the coming period. In the first quarter of 2027Ascend 960PR In the third quarter of 2027 plans to present. Ascend 970 and Ascend 980 are also on the 2028 and 2029 roadmap, respectively. Therefore, beyond a connectivity technology that accelerates Huawei’s existing artificial intelligence systems, UnifiedBus becomes one of the fundamental parts of the infrastructure that will bring together future Ascend generations at the scale of hundreds of thousands and eventually one million NPUs.