Press Launch
In a keynote speech on computing structure innovation at HUAWEI CONNECT 2026, Yang Chaobin (pictured above), Huawei’s Govt Director of the Board and CEO of the ICT Enterprise Group, unveiled new particulars concerning the UnifiedBus interconnect structure, introduced a sequence of computing merchandise the corporate has not too long ago constructed based mostly on this structure, and defined how the corporate has collaborated with open-source communities to allow utility innovation by clients, companions, and builders.
In a conventional computing structure, useful resource utilization declines as a cluster scales. As little as 20% of a 100,000-NPU cluster’s computing capability is definitely utilized by fashions, with an unlimited quantity of compute sitting idle throughout knowledge communications. A ten trillion-parameter mannequin, which requires huge quantities of intermediate knowledge to coach, far exceeds the capability of the reminiscence on a single accelerator beneath conventional architectures. Moreover, AI brokers can have lots of of interactions per hour with AI fashions, which implies computing methods want to have the ability to flexibly make the most of altering quantities of central processing models (CPUs) and neural processing models (NPUs). The ensuing interactions between these processors additionally generate KV cache that far exceeds reminiscence capability, which is driving up demand for knowledge storage throughout compute methods and tiers. All of this implies interconnection and communications between parts inside a cluster have grow to be a key efficiency bottleneck for computing methods.
In accordance with Yang, Huawei’s answer to this bottleneck is its new computing structure, powered by the interconnect expertise known as UnifiedBus, which facilitates collaboration between clusters and SuperPoDs. This structure has been designed to help the growth in compute demand that’s rising as new agentic purposes hit the market. UnifiedBus-powered computing methods have 4 key options:
- Unified protocol and reminiscence semantics: Greater than ten interconnect protocols have been mixed beneath the UnifiedBus protocol, which will increase interconnect bandwidth from 100 GB/s stage to TB/s stage and reduces round-trip time (RTT) latency from 7 microseconds to 2 microseconds. This protocol additionally allows unified world reminiscence addressing inside SuperPoDs.
- Heterogeneous compute collaboration: UnifiedBus instantly interconnects CPUs, NPUs, reminiscence, and solid-state drives (SSDs) to allow decentralized, peer-to-peer entry between them and versatile CPU-NPU mixing. As well as, tiered {hardware} acceleration for Transformer allows Consideration-FFN disaggregation (AFD).
- Tiered storage with world pooling: UnifiedBus makes use of hybrid-media useful resource pooling to help the caching of activations. Double Knowledge Price (DDR) reminiscence serves as the choice reminiscence for NPUs, lowering latency for search, suggestion, and promoting companies, and doubling the efficiency of vector retrieval for 100 billion-level knowledge objects with hundreds of dimensions. This additionally reduces the per-NPU HBM capability necessities for the coaching of 10 trillion-parameter fashions, and improves cluster mannequin FLOPs utilization (MFU).
- Optoelectronic interconnect and versatile networking: UnifiedBus serves as a worldwide ultra-high bandwidth, ultra-low latency “knowledge freeway”, enabling versatile scaling of compute.
Yang then moved on to unveil Huawei’s latest UnifiedBus interconnect merchandise that cowl interconnection inside cupboards, between cupboards, and throughout clusters. This permits elastic scaling from a single cupboard to a million-NPU cluster.
- Inside cupboards, the UnifiedBus LinkBlade eliminates circuit losses with a cable-free design, which reduces copper cabling inside a 4,096-NPU SuperPoD by about 196 kilometers.
- Between cupboards, the UnifiedBus LinkDevice helps 176 ports, with 1.6 Tbit/s bandwidth per port, delivering 280 Tbit/s all-optical interconnection. That is the {industry}’s highest-bandwidth, highest-port-count high-speed bus-protocol interconnect gadget, with an RTT latency as little as 2 microseconds.
- Throughout clusters, the UnifiedBus UBG swap offers an enormous radix fan-out capability of as much as 1,024, which means it will probably help a million-NPU SuperCluster and evolution in direction of supporting tens of trillions of mannequin parameters.
Yang additionally launched Huawei’s UnifiedBus-powered agentic AI SuperCluster, which is comprised of TaiShan 950 and Atlas 960 SuperPoDs, OceanStor M900 reminiscence storage, and a Xinghe UBG swap. This cluster allows numerous compute collaboration and tiered useful resource pooling.
Past SuperPoDs, Huawei has additionally utilized the UnifiedBus expertise to its compute home equipment. The corporate has developed a brand new sequence of UnifiedBus-based home equipment (supporting one to eight NPUs) and Kunpeng and Ascend modules. Two Atlas 650E air-cooled servers might be instantly linked by UnifiedBus to allow switch-free interconnect throughout 16 NPUs, in order that they work seamlessly as a single unit. They will additionally easily scale to type a small-scale SuperPoD that permits small and medium-sized enterprises to run trillion-parameter fashions regionally.
On the associated ecosystem entrance, Yang mentioned that Ascend is collaborating with open-source frameworks like DeepSeek Harness, OpenCode, and openJiuwen to develop absolutely open-source clever administration, inference acceleration, and safety framework agent plugins. Ascend has additionally labored with greater than 50 clients, companions, and universities to create 26 industry-specific operator libraries, and has collaborated with over 90 mainstream open-source communities to assist companions rapidly develop industry-specific brokers and combine them with IT methods.
Rising AI purposes are making compute the cornerstone of commercial clever transformation. Yang promised, “Huawei is dedicated to innovating on the system stage, constructing a computing-foundation product portfolio, and going open supply and open system. We are going to proceed working with clients, companions, and builders to assist this ecosystem thrive and supply the world a brand new choice for compute.”
Huawei Join 2026 occurred on September 17 to 19 on the Shanghai World Expo Exhibition & Conference Heart and Shanghai Expo Heart. For extra info, please go to HUAWEI CONNECT 2026 on-line at www.huawei.com/en/occasions/huaweiconnect
Additionally within the information
Shared Rural Community rollout extends to UK nationwide parks with over 150 4G masts dwell
Vodafone touts newest try and bridge the UK’s digital divide
Virgin Media O2 cuts 5,200 tonnes of carbon dioxide emissions
