#Shanghai #USA #Huawei #SuperPod #UnifiedBus

Huawei's Revolutionary UnifiedBus Computing Architecture

In a groundbreaking announcement at the HUAWEI CONNECT 2026 conference, Yang Chaobin, Executive Director and CEO of Huawei ICT Business Group, presented a transformative new computing architecture named UnifiedBus. The architecture aims to address critical issues associated with traditional computing models, particularly as data demands continue to surge across various sectors.

The Challenge of Traditional Computing Architectures

Traditional computing systems face significant operational inefficiencies, particularly when scaling up clusters. For instance, within a cluster of 100,000 NPUs, the actual utilization of computing power dips dangerously low, with only about 20% of available resources engaged during data transmission processes. This inefficiency stems from the overwhelming demand for intermediate data during training processes, which often surpasses the memory capacity available on traditional accelerators. The inter-component communication within these clusters has emerged as a critical performance bottleneck for computing systems.

Introducing UnifiedBus

Huawei's UnifiedBus architecture presents a solution to these challenges with a focus on enhanced connectivity and resource utilization. With the rapidly growing demand for computation driven by new agent applications, UnifiedBus was specifically designed to support this expansion. The architecture boasts four main features:

1. Unified Protocol and Memory Semantics: UnifiedBus combines over ten interconnection protocols into a singular protocol, enhancing interconnect bandwidth from 100 GB/s to a staggering TB/s, while reducing round-trip time (RTT) from 7 microseconds to just 2. This feature also standardizes memory addressing within SuperPoD.

2. Heterogeneous Computing Collaboration: The UnifiedBus connects processors, NPUs, memory, and SSDs directly, facilitating decentralized peer-to-peer access among components and a more flexible CPU-NPU mix. Moreover, it allows for multilevel hardware acceleration for transformers, enabling disaggregation of Attention and Feed Forward Network layers.

3. Multilevel Storage with Global Pool: Utilizing hybrid resource pooling, UnifiedBus facilitates activation caching, allowing DDR memory to serve as alternative storage for NPUs. This significantly minimizes search latencies in recommendation and advertising services and doubles vector search performance for massive datasets.

4. Optoelectronic Communication and Flexible Networking: Acting as a global high-bandwidth data highway, UnifiedBus enables flexible scaling of computation with ultra-high bandwidth and minimal latency.

Yang then introduced the latest Huawei UnifiedBus products that include interconnect solutions for within racks, between racks, and across clusters, enabling elastic scaling from individual racks to million-NPU clusters.

Products That Redefine the Landscape

• - UnifiedBus LinkBlade: This innovative product eliminates circuit losses with a cable-free design, cutting down copper cabling by about 196 kilometers within a 4096-NPU SuperPoD setting.
• - UnifiedBus LinkDevice: Supporting 176 ports with a throughput of 1.6 Tbps per port, this device represents the industry’s highest-capacity connectivity, featuring fully optical interconnections reaching 280 Tbps, all with an RTT of 2 microseconds.
• - UnifiedBus Switch UBG: This switch is engineered for heavy throughput, supporting a wide bit-width of up to 1024, thus allowing for the management of superclusters with a million NPUs and adapting to support models with tens of trillions of parameters.

Furthermore, Huawei has launched the AI SuperCluster based on UnifiedBus, which integrates TaiShan 950, Atlas 960 SuperPoD, OceanStor M900 for storage, and the Xinghe UBG switch. This configuration enables diverse computational collaboration and multilevel resource pooling.

A Commitment to Open Innovation

Looking towards the ecosystem, Yang emphasized the importance of collaboration with open-source frameworks like DeepSeek Harness and OpenCode. These partnerships aim to foster fully intelligent, open-source management systems and accelerate output alongside safety framework plugins.

In conclusion, Yang reiterated Huawei’s commitment to innovative system-level advancements, signaling a dedication to product portfolios for computing while transitioning toward open-source solutions. "Our goal is to nourish this ecosystem alongside clients, partners, and developers, delivering new computational opportunities to the global market," he stated.

With the introduction of UnifiedBus, Huawei aims to redefine the performance landscape of computing systems and cultivate an environment conducive to innovation.