Imagine orchestrating thousands of advanced AI chips, dispersed across multiple server cabinets, to function seamlessly as a single, colossal computing entity. This is the revolutionary vision Huawei presented at HUAWEI CONNECT 2025, unveiling its innovative SuperPoD infrastructure architecture—a transformative leap in how artificial intelligence systems are constructed, scaled, and deployed globally.
The Innovation of SuperPoD Technology
Unlike conventional AI infrastructures where individual servers operate relatively independently, Huawei’s SuperPoD technology integrates thousands of processing units into one cohesive logical machine. This architecture enables these chips to learn, think, and reason collectively, mimicking the capabilities of a unified system rather than disparate components.
This paradigm promises not just an upgrade in technical specifications but signals a fundamental shift in organizing computing power for AI workloads, enhancing scalability, efficiency, and reliability across diverse industry applications.
Technical Backbone: UnifiedBus 2.0 Protocol
At the heart of Huawei’s SuperPoD is the UnifiedBus 2.0 interconnect protocol. According to Yang Chaobin, Huawei’s CEO of the ICT Business Group, this protocol enables a tightly coupled network of physical servers to operate as a single logical entity.
Solving Connectivity and Latency Challenges
UnifiedBus 2.0 addresses two historically limiting factors in large-scale AI computing:
- Reliable long-distance communication: Traditional copper cables offer high bandwidth but only across short distances—usually between two cabinets. Optical cables enable longer distances but face reliability and latency issues that escalate with scale.
- Bandwidth-latency trade-offs: Maintaining low latency and high bandwidth simultaneously across thousands of chips is a complex challenge.
Huawei’s Deputy Chairman Eric Xu emphasized a multi-layered solution incorporating OSI model reliability features, enabling fault detection and seamless protection switching within 100 nanoseconds. This innovation makes any interruptions at the optical connection level imperceptible at the application layer, ensuring uninterrupted AI workloads.
SuperPoD Architecture: Unprecedented Scale and Performance
The Atlas 950 SuperPoD is Huawei’s flagship showcase, integrating up to 8,192 Ascend 950DT AI chips across 160 cabinets spanning 1,000 square meters. This assembly delivers approximately 8 EFLOPS (Exa Floating Point Operations per Second) in FP8 precision and 16 EFLOPS in FP4 precision.
- Interconnect Bandwidth: 16 petabytes per second, surpassing global peak internet bandwidth by over tenfold.
- Memory Capacity: 1,152 terabytes with ultra-low latency of 2.1 microseconds across the system.
Looking ahead, the Atlas 960 SuperPoD will push the frontier further, featuring 15,488 Ascend 960 chips over 220 cabinets (2,200 square meters), promising 30 EFLOPS in FP8, 60 EFLOPS in FP4, and a staggering 34 PB/s interconnect bandwidth, supported by 4,460 terabytes of memory.
Extending Beyond AI: General-Purpose SuperPoD Systems
Huawei’s SuperPoD vision expands with the TaiShan 950 SuperPoD, leveraging Kunpeng 950 CPUs for broader, general-purpose computing. This system aims to modernize critical enterprise applications, providing alternatives to legacy mainframes and servers, especially in finance. For instance, integrated with distributed GaussDB, it is positioned as a viable replacement for mainframes and Oracle’s Exadata database clusters, promising improved scalability and cost-effectiveness.
Embracing Open Architecture for a Collaborative Ecosystem
Significantly, Huawei has released the UnifiedBus 2.0 protocol and supporting hardware/software specifications as open standards. This open approach acknowledges the existing constraints in semiconductor manufacturing within China and the global market.
Huawei’s executives explained that future sustainable computing power depends on widely available process nodes, advocating an open hardware and software ecosystem to stimulate innovation across industry participants. Partners can develop scenario-specific SuperPoD solutions, strengthening the AI infrastructure ecosystem collaboratively.
- Hardware Releases Include: NPU modules, air-cooled/liquid-cooled blade servers, AI accelerator cards, CPU boards, and cascade cards.
- Software Components: Full open source of CANN compiler toolkits, Mind series application kits, and OpenPangu foundational models expected by end-2025.
Real-World Deployment and Market Impact
In 2025 alone, Huawei has shipped over 300 Atlas 900 A3 SuperPoD units to more than 20 customers across sectors including internet service providers, finance, telecommunications carriers, electric utilities, and manufacturing industries. This traction validates the technical viability and multi-industry demand for large-scale, unified AI computing systems.
Huawei’s open architecture challenges the proprietary hardware-software stacks dominant in Western AI infrastructure providers, offering an alternative pathway rooted in collaborative innovation rather than closed ecosystems.
Conclusion: Transforming the Future of AI Infrastructure
Huawei’s SuperPoD represents a fundamental reimagining of AI hardware architecture. By enabling thousands of AI chips to operate as one unified machine, it addresses major barriers to AI scalability, connectivity, and efficiency.
The open-source release of UnifiedBus 2.0 protocol and associated hardware/software components invites the global developer and industrial community to accelerate AI infrastructure innovation together, potentially reshaping competitive dynamics in the global AI hardware market.
As AI workloads continue escalating in complexity and scale—projected to grow over 10x in compute demand every few years per OpenAI’s reported trend data—such breakthroughs will be critical. (Source: OpenAI Research)
Huawei’s SuperPoD initiative aligns with the broader industry push towards modular, interoperable, and high-performance AI computing solutions, supporting innovations from autonomous driving to advanced natural language processing and scientific research.
Key Takeaways
- SuperPoD unifies thousands of AI chips to function as one logical system.
- UnifiedBus 2.0 protocol solves long-range communication and latency at scale.
- Leading-edge performance with up to 30-60 EFLOPS and multi-petabyte per second interconnect.
- Open hardware and software standards accelerate ecosystem development.
- Real-world deployments across industries demonstrate broad applicability and market readiness.