#China #Shanghai #Huawei #SuperPod #UnifiedBus

Huawei Unveils UnifiedBus Architecture for Advanced Computing

At the HUAWEI CONNECT 2026 event, Yang Chaobin, a member of Huawei’s board and the CEO of the ICT Business Group, revealed groundbreaking developments regarding the UnifiedBus architecture for computing. This innovative technology aims to address inefficiencies in traditional computing architectures, which face significant resource underutilization, especially in large clusters.

In conventional systems, as clusters expand (e.g., to 100,000 units), resource utilization plummets, leaving potentially 80% of processing power idle due to inefficiencies in data transmission. The architecture of UnifiedBus promises to transform this landscape, significantly enhancing performance and utility across computational systems.

Key Features of UnifiedBus Architecture

Yang detailed four primary features of this architecture that are designed to meet the increasing computational demands driven by new AI applications.

1. Unified Protocol and Memory Semantics

UnifiedBus integrates over ten existing connection protocols into a single framework, thus boosting throughput from 100 GB/s to multiple TB/s. Moreover, it notably reduces round-trip transmission delay (RTT) from seven microseconds to just two. This enhanced protocol also allows uniform memory addressing across SuperPoD systems, which streamlines data access.

2. Coordination of Heterogeneous Computing Resources

The architecture enables direct connections between CPUs, NPUs, memory systems, and SSDs, fostering decentralized peer-to-peer access. This innovation allows for better coordination and utilization of both CPU and NPU resources while also employing multi-level hardware acceleration for Transformer model computations.

3. Layered Storage with Shared Resource Pools

Another critical element is the usage of a common resource pool that harnesses various storage media types, facilitating the buffering of activations. This means that Double Data Rate (DDR) memory can act as supplementary memory for NPUs, cutting down latency during service searches and recommendations, while also doubling vector search efficiency in large data sets.

4. Optoelectronic Connections and Flexible Network Topology

The UnifiedBus functions as a global high-bandwidth data highway, allowing flexible scalability in computational resources without sacrificing performance or speed.

Latest Huawei Products Utilizing UnifiedBus

Yang presented an array of Huawei's latest products that leverage the UnifiedBus. These products aim to provide robust connectivity solutions within racks, between racks, and across clusters, allowing for scalable resource management from single racks to massive clusters featuring millions of NPUs.

For instance, the UnifiedBus LinkBlade, designed without cabling, mitigates losses associated with physical connections and is equipped to minimize copper wire lengths within a SuperPoD of 4096 NPUs, saving about 196 kilometers of wire.

At the inter-rack level, the UnifiedBus LinkDevice offers 176 ports, each with a bandwidth of 1.6 Tbit/s, resulting in a total optical connection capacity of 280 Tbit/s, putting Huawei ahead in terms of industry standards.

The inter-cluster switch, UnifiedBus UBG, boasts a radix of 1024, positioning it to support SuperClusters comprising million NPUs, and even accommodate future models scaling into tens of trillions of parameters.

The architecture also plays a vital role in the operation of Huawei's SuperCluster systems built for AI agents, integrating systems like TaiShan 950, Atlas 960 SuperPoD, OceanStor M900 storage systems, and Xinghe UBG switches. This allows diverse computing resources to operate in harmony.

Supporting Open-Source Ecosystem

Huawei’s commitment to collaborating with open-source communities is evident in their strategic partnerships to foster innovation. They are currently working on open plugins tailored for AI agents, improving smart management, inference acceleration, and security protocols.

As Yang emphasized, Huawei’s aim is to innovate across system levels, enhance the infrastructure of computing solutions, and promote open software and systems. The company is dedicated to working with clients, partners, and developers to cultivate a thriving ecosystem that presents a robust alternative in the realm of computational solutions.