Huawei Unveils UnifiedBus Architecture to Revolutionize SuperPoD and Cluster Systems

Huawei's Groundbreaking UnifiedBus Architecture

At the HUAWEI CONNECT 2026 conference, Yang Chaobin, CEO of Huawei's IKT division, introduced the UnifiedBus, a revolutionary computing architecture aimed at transforming SuperPoD and cluster systems. This innovation is set to significantly enhance resource utilization and application performance within data centers.

Traditionally, computing architectures faced diminishing resource usage as clusters expanded. For instance, models often leveraged only 20% of the total computing capacity available in a cluster equipped with 100,000 NPUs (Neural Processing Units), leaving a vast amount of computational power idle during data communications. The growing demand from models with trillions of parameters further complicates the situation, taxing the memory capacities of single accelerators within conventional architectures. Thus, inter-component communication and connectivity have emerged as pivotal bottlenecks.

Huawei’s UnifiedBus architecture directly addresses these challenges by facilitating seamless collaboration among clusters and SuperPoD systems. It is designed to support the surging demand for computing power driven by new application requirements in today’s tech landscape. The key attributes of this architecture include:

1. Unified Protocol and Memory Semantics: UnifiedBus combines over ten connectivity protocols to enhance the connection bandwidth from 100 GB/s to terabyte speeds and reduce round-trip latency from 7 microseconds to 2 microseconds. This results in unified global memory addressing across SuperPoD systems.

2. Heterogeneous Computing Collaboration: The architecture allows direct connections between CPUs, NPUs, memory, and SSDs, enabling decentralized peer-to-peer access. This flexibility aids in efficiently leveraging both CPU and NPU resources. Additionally, a multi-level hardware acceleration framework for Transformer architecture enhances the disaggregation of Attention and Feed Forward Networks (FFN).

3. Multi-Level Storage with Global Aggregation: Utilizing hybrid media aggregation, UnifiedBus optimizes activation storage in cache. The Double Data Rate (DDR) memory serves as auxiliary storage for NPUs, significantly reducing latency for search queries and recommendation systems. It also doubles vector search performance for high-dimensional data, alleviating the need for high-bandwidth memory (HBM) for every NPU during training of large models.

4. Optoelectronic Connectivity and Flexible Networks: Cost-efficient, high-bandwidth, and low-latency global data highways allow for scalable computing power expansion.

During the conference, Yang also showcased the latest UnifiedBus products which facilitate connections within racks, between racks, and across clusters. This facilitates elastic scalability, allowing setups with as many as one million NPUs, while the UnifiedBus LinkBlade dramatically reduces circuit losses through its wireless design by about 196 kilometers in copper wiring across systems with 4,096 NPU units.

The UnifiedBus LinkDevice, supporting 176 ports each with a bandwidth of 1.6 Tbit/s, ensures fully optical connectivity with a throughput of 280 Tbit/s, making it the industry's high-speed connectivity device with minimal 2 microsecond round-trip latency.

In terms of clustering infrastructure, the UnifiedBus UBG switch offers massive branching capacity, supporting superclusters of up to one million NPU units and enabling models with trillions of parameters.

Additionally, Huawei introduced its agent-based AI supercluster utilizing UnifiedBus technology, which incorporates TaiShan 950 and Atlas 960 devices in the SuperPoD system, along with OceanStor M900 storage and Xinghe UBG switch. This architecture facilitates the collaborative operation of diverse computational systems while aggregating resources.

Moreover, Huawei has expanded the application of its UnifiedBus technology beyond SuperPoD by developing a new range of devices tailored for smaller businesses, enabling local execution of trillion-parameter models. Yang emphasized Huawei's commitment to fostering system-level innovations, building a comprehensive product portfolio, and promoting open source technologies. He reiterated Huawei's dedication to collaborating with clients, partners, and developers to enhance the ecosystem and provide new computational capabilities to the world.

In conclusion, Huawei’s innovations in UnifiedBus architectures promise to redefine computational efficiency and interoperability across various platforms, paving the way for enhanced AI and data processing capabilities across industries.