Huawei Unveils Revolutionary SuperPoD Architecture: A Single Logical Machine for AI Computing with 8 EFLOPS in FP8 and 16 PB/s Interconnect Bandwidth
At the cutting-edge of artificial intelligence (AI) innovation, Huawei has unveiled a revolutionary architecture that could redefine the way the world constructs and scales AI systems. This groundbreaking development was showcased at HUAWEI CONNECT 2025, where the company presented its SuperPoD technology – a game-changer in AI infrastructure design.
Instead of traditional setups that rely on individual servers operating somewhat independently, Huawei’s SuperPoD technology creates a unified logical machine from thousands of separate processing units, allowing them to learn, think, and reason collectively.
This advancement transcends impressive technical specifications, signifying a shift in how AI computing power can be organized, scaled, and deployed across various industries.
The core of this infrastructure revolves around UnifiedBus (UB), a protocol developed by Huawei that seamlessly connects physical servers to operate as a single entity. According to Yang Chaobin, Huawei’s Director of the Board and CEO of the ICT Business Group, “The SuperPoD architecture has been designed based on our UnifiedBus interconnect protocol, which deeply interconnects physical servers so that they can learn, think, and reason like a single logical server.”
To address challenges in large-scale AI computing – namely reliability of long-range communications and bandwidth-latency – the UnifiedBus protocol addresses both concerns. Traditional copper connections provide high bandwidth but are limited to short distances, typically connecting no more than two cabinets. On the other hand, optical cables support longer ranges but suffer from reliability issues that become more pronounced with distance and scale.
Huawei’s solution involves building reliability into every layer of their interconnect protocol, from the physical layer through to the network and transmission layers. This results in 100-ns-level fault detection and protection switching on optical paths, ensuring any intermittent disconnections or faults of optical modules are imperceptible at the application layer.
The Atlas 950 SuperPoD serves as the flagship implementation of this architecture, featuring up to 8,192 Ascend 950DT chips arranged in a configuration delivering “8 EFLOPS in FP8 and 16 EFLOPS in FP4.” Its interconnect bandwidth will be an astounding 16 PB/s, making it over ten times more powerful than the entire globe’s total peak internet bandwidth.
The Atlas 950 SuperPoD occupies 160 cabinets spanning 1,000m2, with 128 compute cabinets and 32 comms cabinets linked by all-optical interconnects. The system boasts a memory capacity of 1,152 TB and maintains a latency of just 2.1 microseconds across the entire system.
Later in the production pipeline will be the Atlas 960 SuperPoD, incorporating 15,488 Ascend 960 chips in 220 cabinets covering 2,200m2. This version promises “30 EFLOPS in FP8 and 60 EFLOPS in FP4,” along with 4,460 TB of memory and a 34 PB/s interconnect bandwidth.
The SuperPoD concept extends beyond AI workloads into general-purpose computing through the TaiShan 950 SuperPoD, which is built on Kunpeng 950 processors. This system addresses enterprise challenges in replacing legacy mainframes and mid-range computers, particularly relevant for the finance sector, where it can serve as an ideal alternative to replace mainframes, mid-range computers, and Oracle’s Exadata database servers.
In a move that could reshape global AI infrastructure competition dynamics, Huawei has announced the release of UnifiedBus 2.0 technical specifications as open standards. This decision reflects both strategic positioning and practical constraints, acknowledging China’s projected lag in semiconductor manufacturing process nodes for an extended period.
Huawei is committed to fostering an open-hardware and open-source-software approach that encourages partners to develop their industry-specific SuperPoD solutions. This strategy aims to accelerate developer innovation and create a thriving ecosystem. Huawei plans to open-source hardware and software components, including NPU modules, air-cooled and liquid-cooled blade servers, AI cards, CPU boards, and cascade cards. For software, Huawei will fully open-source CANN compiler tools, Mind series application kits, and openPangu foundation models by 31 December 2025.
Initial deployment of these technologies has already begun, with over 300 Atlas 900 A3 SuperPoD units shipped in 2025, serving more than 20 customers from multiple sectors such as the Internet, finance, carrier, electricity, and manufacturing.
This development represents a significant stride for China’s AI infrastructure, as Huawei addresses the challenges of building competitive AI infrastructure within constrained semiconductor manufacturing and availability parameters. By creating an open ecosystem around domestic technology, Huawei paves the way for broader industry participation in developing AI infrastructure solutions without requiring access to the most advanced process nodes.
For the global AI infrastructure market, Huawei’s open architecture strategy introduces a viable alternative to the tightly integrated, proprietary hardware and software approach prevalent among Western competitors. The success of this ecosystem remains to be seen at scale, but it has the potential to reshape competitive dynamics in the global AI infrastructure market.
