{
    "success": true,
    "data": {
        "id": 1995515,
        "msgid": "huawei-launches-newest-unifiedbus-computing-architecture-for-superpod-and-clusters-1790085590",
        "date": "2026-09-22 19:55:12",
        "title": "Huawei Launches Newest UnifiedBus Computing Architecture for SuperPoD and Clusters",
        "author": "",
        "source": "ANTARA_ID",
        "tags": "",
        "topic": "Technology",
        "summary": "Huawei has unveiled its latest UnifiedBus interconnection architecture designed to optimise large-scale AI computing clusters. The new technology aims to overcome bottlenecks in data communication, significantly increasing bandwidth and reducing latency for massive AI models.",
        "content": "<p>Shanghai, (ANTARA\/PRNewswire) - Huawei has revealed the latest\ndetails of its UnifiedBus interconnection architecture during a\npresentation on computing architecture innovation at the HUAWEI CONNECT\n2026 event. Yang Chaobin, Executive Director of the Board and CEO of\nHuawei\u2019s ICT Business Group, introduced a range of UnifiedBus-based\ncomputing products and explained the company\u2019s collaboration with the\nopen-source community to drive application innovation among customers,\npartners, and developers.<\/p>\n<p>In conventional computing architectures, resource utilisation\nefficiency tends to decrease as cluster scales increase. In a cluster\nwith 100,000 NPUs, for example, AI models only utilise approximately 20%\nof computing capacity, with much of the remaining capacity idling during\ndata communication. Meanwhile, models with 10 trillion parameters\nrequire massive amounts of intermediate data during the training\nprocess, far exceeding the memory capacity of a single accelerator in\nstandard architectures. Consequently, interconnection and\ninter-component communication within clusters have become a primary\nbottleneck in computing system performance.<\/p>\n<p>According to Yang, Huawei is addressing this challenge with a new\ncomputing architecture based on UnifiedBus interconnection technology,\nallowing clusters and SuperPoDs to operate in an integrated manner. This\narchitecture meets the surging demand for computing power driven by the\nincreasing availability of AI agent applications in the market. The\nUnifiedBus-based computing system offers four key advantages:<\/p>\n<ul>\n<li><p>Unified protocol and memory semantics: UnifiedBus merges over 10\ninterconnection protocols into a single protocol. This approach\nincreases interconnection bandwidth from the 100 GB\/s level to the TB\/s\nlevel and reduces round-trip time (RTT) latency from 7 microseconds to 2\nmicroseconds. UnifiedBus also supports unified global memory addressing\nwithin the SuperPoD.<\/p><\/li>\n<li><p>Heterogeneous computing collaboration: UnifiedBus directly\nconnects CPUs, NPUs, memory, and solid-state drives (SSDs), enabling\npeer-to-peer access between components. The architecture also supports\nflexible combinations of CPUs and NPUs. Furthermore, it features\nmulti-level hardware acceleration for Transformers, supporting\nAttention-FFN disaggregation (AFD).<\/p><\/li>\n<li><p>Tiered storage with global pooling: UnifiedBus utilises hybrid\nstorage-based resource pooling to support activation caching. Double\nData Rate (DDR) memory can serve as alternative memory for NPUs,\nreducing latency in search, recommendation, and advertising services.\nThis technology also doubles vector retrieval performance for 100\nbillion data items with thousands of dimensions. Simultaneously, the HBM\ncapacity requirement per NPU for training 10-trillion-parameter models\ncan be reduced, thereby increasing Model FLOPs Utilisation (MFU) in\nclusters.<\/p><\/li>\n<li><p>Optoelectronic interconnection and flexible networking:\nUnifiedBus functions as a global \u2018data highway\u2019 with extremely high\nbandwidth and very low latency, allowing computing capacity to be\nexpanded flexibly.<\/p><\/li>\n<\/ul>\n<p>Yang subsequently introduced Huawei\u2019s latest range of UnifiedBus\ninterconnection products, covering connections within cabinets, between\ncabinets, and between clusters. This portfolio facilitates the elastic\naddition of computing capacity, ranging from a single cabinet to a\ncluster housing one million NPUs.<\/p>\n<ul>\n<li><p>Within the cabinet, UnifiedBus LinkBlade uses a wireless design\nto eliminate circuit losses. This technology reduces the requirement for\ncopper cabling in a SuperPoD with 4,096 NPUs by approximately 196\nkilometres.<\/p><\/li>\n<li><p>Between cabinets, UnifiedBus LinkDevice features 176 ports with a\nbandwidth of 1.6 Tbit\/s per port and provides full optical\ninterconnection up to 280 Tbit\/s. The device is claimed to be the\nindustry\u2019s high-speed bus-protocol-based interconnection device with the\nhighest bandwidth and port count, featuring an RTT latency as low as 2\nmicroseconds.<\/p><\/li>\n<li><p>Between clusters, the UnifiedBus UBG switch offers a fan-out\nradix capacity of up to 1,024. This capacity supports a SuperCluster\nwith one million NPUs, paving the way for systems capable of handling\nmodels with tens of trillions of parameters.<\/p><\/li>\n<\/ul>\n<p>Yang also introduced Huawei\u2019s UnifiedBus-powered agentic AI\nSuperCluster, consisting of TaiShan 950 and Atlas 960 SuperPoDs,\nOceanStor M900 memory storage, and Xinghe UBG switches. The cluster\nsupports various types of computing collaboration alongside multi-level\nresource pooling.<\/p>\n<p>Beyond SuperPoDs, Huawei is applying UnifiedBus technology to\nindividual computing devices. The company is developing a new series of\nUnifiedBus-based devices supporting one to eight NPUs, as well as\nKunpeng and Ascend modules. This solution assists small and medium-sized\nenterprises in running trillion-parameter AI models locally.<\/p>\n<p>Regarding the ecosystem, Yang stated that Ascend is collaborating\nwith open-source frameworks such as DeepSeek Harness, OpenCode, and\nopenJiuwen to develop agent plugins for fully open-source intelligent\nmanagement, inference acceleration, and security frameworks.<\/p>\n<p>Yang concluded, \u201cHuawei is committed to innovating at the system\nlevel, building a product portfolio as a computing foundation, and\nencouraging active participation through open source and open systems.\nWe will continue to work with customers, partners, and developers to\ndevelop this ecosystem and provide new options for the world to build\ncomputing capacity.\u201d<\/p>",
        "url": "https:\/\/jawawa.id\/newsitem\/huawei-launches-newest-unifiedbus-computing-architecture-for-superpod-and-clusters-1790085590",
        "image": ""
    },
    "sponsor": "Okusi Associates",
    "sponsor_url": "https:\/\/okusiassociates.com"
}