Huawei Introduces Groundbreaking UnifiedBus Architecture for Next-Gen SuperPoDs and Clusters

Huawei's UnifiedBus Architecture Revolutionizes Computing

In a dynamic keynote at HUAWEI CONNECT 2026 held in Shanghai, Yang Chaobin, a board member and CEO of Huawei's ICT Business Group, introduced major innovations in computing architecture, showcasing the new UnifiedBus connection technology. This infrastructure aims to address significant limitations that arise from traditional computing setups, especially as clusters expand. The shortcomings of previous architectures become apparent when attempting to scale; notably, the computing capacity of large clusters like those utilizing 100,000 NPUs often sees only 20% utilization. As noted by Yang, much of the processing power remains underused, largely due to inefficiencies in data communication.

The Need for Advanced Computing

The rise of model training requirements, especially with monumental datasets and complex models demanding massive interim storage, has emphasized the need for improved architectural solutions. In traditional setups, the storage capacity of single accelerators pales in comparison to what is needed for training models with trillions of parameters. This has made inter-component communication a bottleneck for efficient data processing and system performance.

UnifiedBus: The Solution to Connectivity Challenges

To tackle these limitations, Huawei's UnifiedBus architecture facilitates seamless connections across clusters and SuperPoDs. The architecture is particularly geared towards addressing the impending surge in computing demand anticipated with the emergence of new agent-based applications. The characteristics of UnifiedBus include:

1. Unified Protocols and Memory Semantics: By consolidating over ten connection protocols, UnifiedBus increases connection bandwidth substantially—boosting from 100 GB/s to terabits per second. Additionally, it reduces round-trip time latency from 7 to 2 microseconds, while allowing a unified global memory addressing scheme within SuperPoDs.

2. Heterogeneous Computing Collaboration: UnifiedBus enables direct connections among CPUs, NPUs, memory, and SSDs, facilitating a decentralized peer-to-peer access model that supports flexible combinations of CPU and NPU. Notably, it enhances hardware acceleration for transformers, allowing for more efficient processing.

3. Multilevel Storage with Global Pooling: The architecture utilizes hybrid media resource pooling to enhance activation caching. For instance, DDR memory acts as an alternative storage for NPUs, effectively cutting down latency for search and recommendation services. This aspect greatly improves the efficiency of model training, particularly for large-scale models and resource utilization.

4. Optoelectronic Connectivity and Flexible Networking: With its exceptional bandwidth and low latency, UnifiedBus serves as a 'data highway' that allows for flexible scalability of computing power.

After introducing the architectural framework, Yang unveiled several UnifiedBus-based connection products designed to improve intercommunication within cabinets and across clusters. Notably, the UnifiedBus LinkBlade minimizes wiring losses, reducing the copper cabling required in a SuperPoD with 4,096 NPUs by roughly 196 km.

With UnifiedBus LinkDevice, you gain access to 176 ports, each delivering a bandwidth of 1.6 Tbit/s, equating to an impressive aggregate capacity of 280 Tbit/s. This state-of-the-art connection device boasts the highest number of ports and bandwidth in the industry, with a latency of merely 2 microseconds.

Further innovations include the UnifiedBus-UBG-Switch, which offers significant scalability potential—supporting superclusters with one million NPUs and paving the path for systems that can handle dozens of trillions of model parameters.

Advancements for Smaller Enterprises

In addition, Huawei has integrated the UnifiedBus technology into its computing appliances, developing a new range of devices that facilitate local execution of billion-parameter models. This initiative particularly benefits small and medium enterprises that require advanced computational capabilities without hefty investments in infrastructure.

Commitment to an Open Ecosystem

Yang emphasized Huawei's commitment to collaborating with open-source communities, ensuring that their technology aligns with frameworks such as DeepSeek Harness and OpenCode. The ultimate goal is to cultivate a thriving ecosystem that enhances innovation across the board.

In closing, Yang stated, "Huawei is dedicated to system-level innovation and aims to build a comprehensive product portfolio covering foundational computing while transitioning to open-source systems. We will continue forging partnerships with our customers, collaborators, and developers to ensure this ecosystem thrives and introduces new computing solutions worldwide."