Huawei has introduced an innovative architectural approach at HUAWEI CONNECT 2025, showcasing its SuperPoD technology that fundamentally transforms how AI chips interact. The core idea is to knit together thousands of individual AI processing units, distributed across numerous server racks, enabling them to function as a unified, colossal computer. This development represents a significant leap from conventional server architectures, where components typically operate with a degree of independence.
At the heart of this advancement is Huawei's proprietary UnifiedBus (UB) interconnect protocol, now in its 2.0 iteration. This protocol is specifically engineered to overcome the persistent challenges of long-distance communication reliability and bandwidth-latency constraints that have historically hampered large-scale AI systems. Traditional methods, such as copper connections, offer high bandwidth but are limited to short distances, while optical cables, though capable of longer ranges, often suffer from reliability issues that escalate with distance and scale. Huawei's solution embeds reliability into every layer of its interconnect protocol, ensuring fault detection and protection switching at a 100-nanosecond level, rendering any transient disconnections virtually undetectable at the application layer.
The SuperPoD architecture is exemplified by the Atlas 950 SuperPoD, which integrates up to 8,192 Ascend 950DT chips, delivering an astonishing 8 EFLOPS in FP8 and 16 EFLOPS in FP4, with an interconnect bandwidth of 16 PB/s. This system, housed within 160 cabinets spanning 1,000 square meters, boasts a massive 1,152 TB memory capacity and maintains a remarkably low 2.1-microsecond latency. Huawei also plans to launch the Atlas 960 SuperPoD, which will further scale these capabilities. Beyond AI, the SuperPoD concept extends to general-purpose computing with the TaiShan 950 SuperPoD, designed to offer a robust alternative to legacy mainframe systems, particularly for the finance industry. In a strategic move to foster broader innovation and adoption, Huawei has committed to open-sourcing the technical specifications of UnifiedBus 2.0, along with various hardware and software components, by the end of 2025. This open-architecture strategy aims to build a thriving ecosystem and address the limitations in semiconductor manufacturing, especially in regions like mainland China, by promoting collaborative development and deployment of AI infrastructure solutions.
The successful deployment of over 300 Atlas 900 A3 SuperPoD units across more than 20 diverse customers underscores the immediate impact and potential of this technology. By championing an open ecosystem built around its domestic advancements, Huawei is actively shaping the future of AI infrastructure, particularly in contexts where access to cutting-edge semiconductor process nodes is restricted. This approach not only provides a viable alternative to the proprietary systems prevalent in the global AI market but also empowers a wider array of partners to contribute to AI innovation. The SuperPoD architecture symbolizes more than just an incremental improvement; it signifies a fundamental rethinking of how computational resources can be synergistically connected, managed, and expanded to unlock unprecedented capabilities in artificial intelligence. This vision, rooted in openness and collaboration, promises to propel the global AI infrastructure market into a new era of innovation and shared progress.
