Author | Li Shuiqing Editor | Mo Ying
Reporting on September 19, Zhidx noted that at the Huawei Connect 2026 held from September 17 to 19, Huawei launched a series of new AI infrastructure products centered around the UnifiedBus architecture and super node clusters, including the Ascend 960 chip, Ascend 960 super node, and OceanStor M900 AI memory storage, attracting widespread attention from the industry. Among them, "UnifiedBus" has become a high-frequency keyword running through all the releases, arousing much curiosity.
What is UnifiedBus? What makes it stand out?
Yang Chaobin, Managing Director of Huawei and CEO of ICT BG, provided the complete answer in his speech themed "UnifiedBus Interconnection, Architectural Innovation, and Building the Computing Power Foundation for Agentic AI." Simply put, UnifiedBus is Huawei's unified interconnection bus protocol for its entire computing power foundation. It uniformly interconnects various computing and storage units such as CPUs, NPUs, DPUs, DDRs, and SSDs, and further penetrates single machines, super nodes, and ultra-large-scale clusters.
Under traditional interconnection, the actual computing power utilization rate of a 100,000-card cluster is only about 20%, which is difficult to meet the computing demands of the AI Agent era. Data released by Huawei shows that a single cluster based on UnifiedBus can support 1 million AI processors, and the Model Flops Utilization (MFU) of a 100,000-card cluster has increased from 20% to 35%. Why did Huawei pin the key point of this architectural innovation on "communication"? This starts with the changes in computing demands in the AI Agent era.
01.Communication Becomes the Decisive Factor for AI Agent Computing Power Three Challenges "Force" the UnifiedBus Architectural Answer
In the AI Agent era, communication is becoming the decisive factor for computing power. Currently, the scale of model parameters has moved from the trillion level to the 10-trillion level, and may further reach 20 trillion or even 50 trillion in the future; the context length has also grown from the K-level to the M-level, and may reach tens of Ms in the future. Meanwhile, the frequency of human-Agent interaction is also much higher than that of models. As the underlying computing system for AI, it needs to simultaneously support ultra-large model inference, high-frequency calls, multi-tool usage, and multi-level information transmission and storage, facing numerous challenges.
1. Communication Traffic Surges Quadratically, 100,000-Card MFU Falls Below 20%
The first challenge comes from the rapid expansion of the communication scope. The communication traffic within the data parallel domain has an approximately quadratic relationship with the number of parameters. The proportion of unavoidable communication time increases exponentially, causing the MFU to drop rapidly. The MFU of a typical MoE model on a 100,000-card cluster is less than 20%. How to break the trend of MFU dropping rapidly with the increase of cluster scale has become the first major hurdle for computing systems.
2. One Type of Computing Power Cannot Satisfy All AI Workloads; Heterogeneous Collaboration and Memory Pooling Become Imperative
The second challenge stems from the diversification of AI business workloads. Prefill and Decode exhibit different computing and memory access characteristics; therefore, P and D are separated and satisfied with different computing powers. For future inference with larger models and lower latency, the Attention and FFN in Decode, due to differences in computing and memory access characteristics, can be accelerated by dedicated computing power to improve efficiency. Meanwhile, as Agent applications become popular, the ratio of CPU to NPU changes. Computing systems require heterogeneous computing power collaboration and memory resource pooling to achieve dynamic and efficient adjustments.
3. Ever-Lengthening Contexts; Legacy Interconnects Cannot Support Multi-Level Caching
The third challenge is the storage pressure brought by ever-lengthening contexts. Whether for training or inference, it is imperative to build a hierarchical multi-level caching architecture, including tiered HBM, DDR, and SSD. The existing PCIe plus RoCE networks are difficult to meet the latency and bandwidth requirements needed for caching. Future PB-level KV Caches also require larger bandwidth and lower latency to match. It can be said that communication has begun to become the key to computing system performance. It requires ultra-strong interconnection capabilities to support efficient collaboration and expansion of heterogeneous computing power; it also needs to break physical limits to realize the pooling of resources such as computing power and memory; and it also requires ultra-large bandwidth and lower latency to ensure the communication efficiency of the computing system. UnifiedBus is Huawei's architectural answer to these changes.
02.One Protocol Connects Chips to Clusters Four Core Competencies of UnifiedBus Reload the Computing Power Foundation
The core of UnifiedBus is to use a single protocol to run through the entire system from chips to clusters, unify complex protocols, unify memory semantics, and eliminate protocol conversion overhead. Yang Chaobin mentioned that its competitiveness can be summarized in the following four dimensions:
1. Protocol Unification and Unified Memory Semantics. Traditional computing systems adopt a master-slave architecture. CPUs and memory within a server can be directly accessed via load/store instructions, but cross-node access usually relies on RDMA's Send/Receive message passing. Data needs to be repeatedly packed and unpacked, and protocol conversion is also required. What UnifiedBus does is to reorganize this process from the protocol layer. Huawei uniformly interconnects core components such as CPUs, NPUs, DPUs, memory, storage, and network cards based on the UnifiedBus protocol, achieving the unification of over a dozen protocols; at the same time, it unifies memory semantics within the super node to achieve unified global memory addressing. As a result, its interconnection bandwidth has increased from the hundreds of GB level to the TB level, and the RTT communication latency has been compressed from 7 microseconds to 2 microseconds, a reduction of about 70%.
2. Diverse Computing Power and Heterogeneous Collaboration. Computing and storage units such as CPUs, NPUs, memory, and SSDs are directly peer-to-peer interconnected through UnifiedBus, forming a decentralized and equal access method. In this way, the computing system can flexibly allocate CPUs and NPUs according to different business needs, while possessing scalability, without being constrained by fixed node structures.
3. Tiered Storage and Global Pooling. As AI models continue to grow larger, the memory capacity provided by a single chip or a single server is increasingly difficult to meet the demand. Through tiered storage and global pooling, UnifiedBus incorporates multiple mixed media into a unified resource system, improving memory utilization.
4. Optoelectronic Interconnection and Flexible Networking. When super nodes scale from thousands of cards to tens of thousands of cards, electrical interconnection begins to be limited by physical distance, bandwidth, and power consumption. Huawei builds an ultra-high bandwidth and ultra-low latency "data highway" through optoelectronic interconnection, allowing computing power to extend from within the cabinet to across cabinets.
Taking the 4096-card super node as an example, it requires 56 computing cabinets and 14 UnifiedBus interconnection cabinets, with the maximum transmission distance of UnifiedBus exceeding 200 meters. Through optical path protection, high-speed optical interconnection, and UnifiedBus switching equipment with 176 ports and 280Tb high-speed interconnection, Huawei has expanded the scale of a single super node from 1024 cards to 4096 cards. Further expanding towards Scale-Out, the 4096-card super node can serve as a basic expansion unit. Through the 1024 ultra-large Radix fan-out capability, two-tier CLOS, and multi-plane, multi-rail topology, a single cluster can ultimately support 1 million AI processors.
An industry chain insider told Zhidx that in the past few years, the industry has mostly optimized the existing PCIe plus RoCE system. However, Agent workloads simultaneously amplify the shortcomings of communication, memory, and storage. Continuing to patch old protocols will yield diminishing returns. UnifiedBus is equivalent to reloading the underlying logic of computing power for AI Agents to meet diverse computing power demands.
03.From Single Cabinet to Millions of Cards, From Super Nodes to All-in-One Machines Meeting Various Agent Demands
Based on UnifiedBus, Huawei has further built an AI Agent super node cluster featuring optoelectronic interconnection, diverse computing power, and tiered pooling. Computing, storage, and connectivity products fully support UnifiedBus, featuring modular design and flexible configuration.
1. Supporting the Super Node "Full Suite", Ascend 960 Makes an Early Debut, Pushing Computing Efficiency to the Limit
The newly released Agentic AI super node cluster at the conference consists of the Kunpeng 950 super node, Ascend 960 super node, OceanStor M900 memory storage, and Xinghe UBG switch. The entire super node cluster can achieve UnifiedBus unified protocol, flexible expansion of diverse computing power, and tiered resource pooling, optimizing computing efficiency. Among them, based on UnifiedBus + NPO optical engine Hi-ONE, a single Ascend 960 super node can achieve a scale of 4096 cards, providing a maximum computing power of 8 EFLOPS FP8, achieving up to 1PB of HBM capacity, doubling the mean time between failures, and reaching a system availability of 99.8%. The Ascend 960 air-cooled super node and liquid-cooled super node will be launched in Q2 and Q3 2027, respectively.
2. From 172T per Single Cabinet to Millions of Cards: UnifiedBus Unification, Tiered Elastic Expansion on Demand
In the Scale-Up direction, the in-cabinet UnifiedBus switching blade supports ultra-large bandwidth of 172T, enabling the interconnection of 72 Ascend processors and 18 Kunpeng processors. Huawei also reduces in-cabinet cables through optical interconnection. A single cabinet within a 4096 super node can save 196 kilometers of copper cables, a distance greater than that from Shanghai to Hangzhou.
Across cabinets, the UnifiedBus switching equipment has 176 ports, with each port reaching 1.6T, achieving 280T all-optical interconnection in a single machine. Two-tier CLOS networking can build a 4096-card super node, with RTT latency as low as 2 microseconds. Expanding further towards Scale-Out, the UBG UnifiedBus network switch has a 1024 ultra-large Radix fan-out capability, supporting million-card super node clusters and supporting the evolution of model parameters to tens of trillions. Thus, UnifiedBus has formed a tiered and hierarchical interconnection system from a single cabinet, thousands of cards, tens of thousands of cards, to millions of cards, capable of elastic expansion on demand.
3. Computing and Storage Trio, UnifiedBus Integration Yields 1+1+1>3
Looking at computing power and storage again, the computing blade of the Kunpeng 950 super node provides 384 CPU cores and 6TB of memory. A single cabinet can provide 12,288 CPU cores and support 192TB of memory. Based on UnifiedBus, the Kunpeng super node can achieve unified scheduling of over 1.5 million cores, as well as global pooling, unified addressing, and unified access of 24,000TB of memory.
The Ascend 960 super node computing blade supports 8 Ascend 960 NPUs, providing 32 PFLOPS FP4 computing power, 4.5TB of unified memory, and 17.9TB/s UnifiedBus interconnection bandwidth. The OceanStor M900 is responsible for memory storage. Through a brand-new storage controller and ASU module, it achieves the "three-chip integration" of network, CPU, and disk control chips. A single ASU module has a capacity density of 64TB, and the single-node access bandwidth reaches 480GB/s.
The three types of hardware are flexibly configured through the UnifiedBus unified protocol to achieve coordinated collaboration and global access. Huawei summarizes this integration effect as "1+1+1>3".
4. UnifiedBus Integrated into All-in-One Machines: Atlas 650E Direct Dual-Machine Connection, Running Trillion-Parameter Models Locally
Beyond larger-scale clusters, small and medium-sized scenarios are also within the coverage of UnifiedBus. The actual deployment of most enterprises still requires air-cooled and lightweight solutions, and desktop-level AI applications are also rapidly gaining popularity. Therefore, based on UnifiedBus, Huawei has launched a serialized, multi-level all-in-one machine product line ranging from 1 card to 8 cards, as well as Kunpeng and Ascend modules.
The Atlas 650E can achieve direct dual-machine UnifiedBus interconnection, allowing 16 NPUs to form a Full-Mesh network without the need for switches. The two machines can run as one, supporting trillion-parameter models. Through EP16 parallel partitioning, inference throughput is increased by more than 40%, and latency is as low as 10 milliseconds, enabling small and medium-sized enterprises to run trillion-parameter large models locally while ensuring user experience and data security.
In the future, when expanded to more than 16 cards, it can continue to scale smoothly through UnifiedBus interconnection to form small super nodes, while supporting PD separation deployment to achieve efficient collaboration of heterogeneous AI computing power. In general, with million-card clusters for large scenarios and all-in-one machines for small and medium scenarios, Huawei aims to make "flexible computing power" an industry imperative.
04.Open-Source Components Plus Industry Operator Libraries Enabling Ecosystem Partners to "Use Out of the Box"
For UnifiedBus to become a foundation, it cannot rely solely on hardware. Yang Chaobin recalled in his speech that on August 5 last year, Huawei's CANN was fully open-sourced, and has since risen to become the most active open-source project in China. Currently, the CANN community has reached over 5,200 monthly active developers, with continuous participation from developers from enterprises, universities, and other groups. More than 50 leading industry customers, partners, and university research teams have collaborated with Ascend to integrate industry knowledge and business experience into Ascend, creating 26 industry-specific operator libraries.
Together with China Southern Power Grid, universities, and other partners, Huawei has built a series of exclusive open-source operator libraries in the community, including power simulation solving, power load forecasting, and power equipment inspection, making transmission line inspection more accurate and power grid load forecasting more efficient, truly bringing AI to the front line of power production. iFLYTEK has deeply cultivated AI large models and application technology innovation. Based on Ascend, it has created leading AI models and developed general solutions such as intelligent customer service and office applications, as well as industry-scenario solutions for finance, education, and healthcare, providing customers with high-quality AI solutions.
On the other side is the Agent development ecosystem. Huawei is collaborating with open-source frameworks such as DeepSeek Harness, OpenCode, and openJiuwen to create Agent plug-in components such as intelligent management systems, inference acceleration libraries, and security frameworks, and is fully open-sourcing them. To further lower the threshold for development and deployment, Huawei has also cooperated with more than 90 mainstream open-source communities, covering the full-link scenarios of operator programming, basic library construction, model training and fine-tuning, inference deployment, and reinforcement learning.
It can be seen that what Huawei ultimately hopes to achieve is "out-of-the-box" usability. This is also the core logic of the so-called "silicon-based black soil" strategy.
05.Conclusion: The Battle for the Computing Power Foundation Huawei Offers a New Choice of "System-Level Plus Open Ecosystem"
The AI era is surging forward, and computing power is the core cornerstone of this transformation. At the end of his speech, Yang Chaobin stated that in the future, Huawei will adhere to system-level technological innovation, build a serialized computing power foundation for all walks of life; adhere to open source and openness, unite with customers, partners, and developers to co-build the ecosystem, and provide new choices for the world.
From UnifiedBus breaking the communication wall, to 11 key chips supporting the super node, to the million-card cluster and open-source ecosystem crossing the inflection point, Huawei is building a computing power foundation for the 10-trillion-parameter large model era with its vertical integration capability of "computing plus communication." When the supply of computing power moves from "hardware monetization" to "ecosystem win-win," the competition in AI infrastructure has been elevated to a contest of systematic capabilities. Huawei's silicon-based black soil is trying to provide another solid choice for the intelligent world.