Huawei AI Chips Architecture Innovations
At HUAWEI CONNECT 2025, Huawei unveiled a groundbreaking development in its AI infrastructure, showcasing a revolutionary SuperPoD technology that connects thousands of AI chips scattered across server cabinets. This innovative approach allows the chips to operate together as if they were one massive computer, a significant departure from traditional architectures where servers work independently.
UnifiedBus 2.0: The Technical Backbone
Central to this system is Huawei’s UnifiedBus (UB) interconnect protocol, which creates a cohesive network of physical servers. As Yang Chaobin, Huawei’s ICT Business Group CEO, describes, the SuperPoD architecture enables these units to learn, think, and reason as a single logical machine. This represents a major leap in organizing and deploying AI computing power within various industries.
The UnifiedBus protocol is designed to overcome the long-standing challenges of reliability in long-range communications and bandwidth-latency issues that have hindered large-scale AI computing. Traditional copper connections, while providing high bandwidth, are limited to short distances. In contrast, optical cables, supporting longer ranges, often suffer from reliability issues. Eric Xu, Huawei’s Deputy Chairman, emphasized the necessity of resolving these connectivity challenges as integral to their strategy.
SuperPoD Architecture: Scale and Performance
The Atlas 950 SuperPoD is the flagship of this new architecture, incorporating up to 8,192 Ascend 950DT chips. Xu noted that it can achieve an astounding performance level, delivering 8 EFLOPS in FP8 and 16 EFLOPS in FP4, with interconnect bandwidth reaching 16 PB/s, surpassing the total peak internet bandwidth globally.
The footprint of the Atlas 950 SuperPoD spans 160 cabinets over 1,000 m², evidencing its significant capacity with 1,152 TB of memory and a system latency of only 2.1 microseconds. The upcoming Atlas 960 SuperPoD is set to push this boundary even further, featuring 15,488 chips and showcasing enhanced processing power and memory capacity.
Extending Beyond AI: General-Purpose Computing
Interestingly, the SuperPoD technology is not limited to AI applications. The TaiShan 950 SuperPoD, built on Kunpeng 950 processors, aims to tackle enterprise challenges, particularly for the finance sector. Xu suggests that this system could replace traditional mainframes and mid-range computers, representing a major advantage for finance and data management.
Open Architecture Strategy
In a pivotal move for the AI infrastructure market, Huawei has made the technical specifications of UnifiedBus 2.0 open standards. This decision is grounded both in strategic foresight and practical necessity due to the current limitations in China’s semiconductor manufacturing capabilities. Yang highlighted the importance of fostering an ecosystem that encourages industry partners to develop tailored SuperPoD solutions, thereby driving innovation.
To facilitate this, Huawei plans to open-source several components, including both hardware and software, by the end of 2025. This initiative could potentially accelerate the development of diverse AI infrastructure solutions and broaden market participation, mitigating dependence on cutting-edge semiconductor processes.
Market Deployment and Its Implications
Real-world implementation has already begun with over 300 Atlas 900 A3 SuperPoD units shipped in 2025 to various sectors, underscoring Huawei’s ambitious scaling. The development of an open ecosystem around domestic technology may enable China to strengthen its AI infrastructure amid the challenges posed by limited semiconductor manufacturing capacity.
Globally, Huawei’s approach introduces a unique alternative to the proprietary systems prevalent in Western markets. As the SuperPoD architecture redefines how computation resources are managed and scaled, the success of Huawei’s open-source strategy will be crucial in determining its impact on competitive dynamics within the AI infrastructure sector.