Inside Huawei’s plan to make thousands of AI chips think like one computer
Imagine connecting thousands of highly effective AI chips scattered in dozens of server cupboards and making them work collectively as in the event that they have been a single, large computer. That is strictly what Huawei demonstrated at HUAWEI CONNECT 2025, the place the corporate unveiled a breakthrough in AI infrastructure structure that would reshape how the world builds and scales synthetic intelligence methods.
Instead of conventional approaches the place particular person servers work considerably independently, Huawei’s new SuperPoD expertise creates what the corporate’s executives describe as a single logical machine created from thousands of separate processing items, permitting them, or it, to “study, think, and cause as one.”
The implications prolong past spectacular technical specs, representing a shift in how AI computing energy may be organised, scaled, and deployed in industries.
The technical basis: UnifiedBus 2.0
At the core of Huawei’s infrastructure method is UnifiedBus (UB). Yang Chaobin, Huawei’s Director of the Board and CEO of the ICT Business Group, explained that “Huawei has developed the groundbreaking SuperPoD structure primarily based on our UnifiedBus interconnect protocol. The structure deeply interconnects bodily servers in order that they will study, think, and cause like a single logical server.”
The technical specs reveal the scope of this achievement. The UnifiedBus protocol addresses two challenges that, traditionally, have restricted large-scale AI computing: the reliability of long-range communications and bandwidth-latency. Traditional copper connections present excessive bandwidth however solely over quick distances, sometimes connecting maybe two cupboards.
Optical cables help longer vary however undergo from reliability points that grow to be extra problematic the better the gap and scale. Eric Xu, Huawei’s Deputy Chairman and Rotating Chairman, said that fixing these basic connectivity challenges was important to the corporate’s AI infrastructure technique.
Xu detailed the breakthrough options in phrases of the OSI model: “We have constructed reliability into each layer of our interconnect protocol, from the bodily layer and knowledge hyperlink layer, all the way in which up to the community and transmission layers. There is 100-ns-level fault detection and safety switching on optical paths, making any intermittent disconnections or faults of optical modules imperceptible on the software layer.”
SuperPoD structure: Scale and efficiency
The Atlas 950 SuperPoD represents the flagship implementation of this structure, comprising of up to 8,192 Ascend 950DT chips in a configuration that Xu described as delivering “8 EFLOPS in FP8 and 16 EFLOPS in FP4. Its interconnect bandwidth might be 16 PB/s. This signifies that a single Atlas 950 SuperPoD could have an interconnect bandwidth over 10 occasions greater than all the globe’s complete peak web bandwidth.”
The specs are greater than incremental enhancements. The Atlas 950 SuperPoD occupies 160 cupboards in 1,000m2, with 128 compute cupboards and 32 comms cupboards linked with all-optical interconnects. The system’s reminiscence capability reaches 1,152 TB and maintains what Huawei claims is 2.1-microsecond latency in all the system.
Later within the manufacturing pipeline would be the Atlas 960 SuperPoD, which is about to incorporate 15,488 Ascend 960 chips in 220 cupboards overlaying 2,200m2. Xu stated it would ship “30 EFLOPS in FP8 and 60 EFLOPS in FP4, and include 4,460 TB of reminiscence and 34 PB/s interconnect bandwidth.”
Beyond AI: General-purpose computing purposes
The SuperPoD idea extends past AI workloads into general-purpose computing via the TaiShan 950 SuperPoD. Built on Kunpeng 950 processors, this technique addresses enterprise challenges in changing legacy mainframes and mid-range computer systems.
Xu positioned this as notably related for the finance sector, the place “the TaiShan 950 SuperPoD, mixed with the distributed GaussDB, can function a perfect different, and change — as soon as and for all — mainframes, mid-range computer systems, and Oracle’s Exadata database servers.”
Open structure technique
Perhaps most importantly for the broader AI infrastructure market, Huawei introduced the discharge of UnifiedBus 2.0 technical specs as open requirements. The resolution displays each strategic positioning and sensible constraints.
Xu acknowledged that “the Chinese mainland will lag behind in semiconductor manufacturing course of nodes for a comparatively very long time” and emphasised that “sustainable computing energy can solely be achieved with course of nodes which are virtually out there.”
Yang framed the open method as ecosystem constructing: “We are dedicated to our open-hardware and open-source-software method that may assist extra companions develop their very own industry-scenario-based SuperPoD options. This will speed up developer innovation and foster a thriving ecosystem.”
The firm is to open-source {hardware} and software program parts, with {hardware} together with NPU modules, air-cooled and liquid-cooled blade servers, AI playing cards, CPU boards, and cascade playing cards. For software program, Huawei dedicated to totally open-sourcing CANN compiler instruments, Mind sequence software kits, and openPangu basis fashions by 31 December 2025.
Market deployment and ecosystem affect
Real-world deployment gives validation for these technical claims. Over 300 Atlas 900 A3 SuperPoD items have already been shipped in 2025, which have been deployed for greater than 20 prospects from a number of sectors, together with the Internet, finance, provider, electrical energy, and manufacturing sectors.
The implications for the event of China’s AI infrastructure are substantial. By creating an open ecosystem round home expertise, Huawei is addressing the challenges of constructing aggressive AI infrastructure inside parameters set by constrained semiconductor manufacturing and availability. Its method permits broader {industry} participation in growing AI infrastructure options while not having entry to essentially the most superior course of nodes.
For the worldwide AI infrastructure market, Huawei’s open structure technique introduces another to the tightly built-in, proprietary {hardware} and software program method dominant amongst Western opponents. Whether the ecosystem proposed by Huawei can obtain comparable efficiency and keep industrial viability stays to be demonstrated at scale.
Ultimately, the SuperPoD structure represents greater than an incremental advance for AI computing. Huawei is proposing a basic of how large computational sources are related, managed, and scaled. The open-source launch of its specs and components will check whether or not collaborative growth can speed up AI infrastructure innovation in an ecosystem of companions. That has the potential to reshape aggressive dynamics within the international AI infrastructure market.
See additionally: Huawei commits to training 30,000 Malaysian AI professionals as local tech ecosystem expands
Want to study extra about AI and large knowledge from {industry} leaders? Check out AI & Big Data Expo going down in Amsterdam, California, and London. The complete occasion is a component of TechEx and is co-located with different main expertise occasions, click on here for extra info.
AI News is powered by TechForge Media. Explore different upcoming enterprise expertise occasions and webinars here.
The submit Inside Huawei’s plan to make thousands of AI chips think like one computer appeared first on AI News.
