Huawei is bringing processor, memory, storage and networking components together in AI infrastructures with the UnifiedBus architecture it introduced at HUAWEI CONNECT 2026. The system aims to reduce data transfer bottlenecks in large-model training and inference, using a protocol and shared memory semantics as the core connection for SuperPoD. According to Huawei, bandwidth rises from hundreds of GB per second to TB per second, while latency drops from 7 microseconds to 2 microseconds; CPU, NPU, memory and SSD are connected directly in a peer-to-peer configuration. LinkDevice offers speeds of 1.6 Tbit/s across 176 ports and an optical capacity of 280 Tbit/s; UBG scales to up to 1,024 connections. The architecture supports SuperClusters with 1 million NPUs and the Peerium Computing Architecture. It is used in Atlas 950 SuperPoD and SuperCluster; Atlas 960 with NPO technology is being tested.
Why it matters
The importance of UnifiedBus stems from bringing processor, memory, storage and networking resources in AI infrastructure together around a common connection and memory model, rather than managing them as separate components. This structure directly affects infrastructure scale because it aims to reduce the bottleneck caused by moving data between components, particularly in large-model training and inference. The fact that UBG can scale to up to 1,024 connections and supports SuperClusters with 1 million NPUs shows that the architecture is designed for use cases extending from a single system to larger clusters. However, the fact that Atlas 960 is still in the testing phase indicates that the results of NPO technology within this architecture are still being evaluated.
Background
Huawei is not a new name in the FikirPilot archive: we have published 7 reports mentioning the name in the past 90 days; the latest is dated September 19, 2026.