AnswersPublished21 sources
Huawei’s Lingqu UnifiedBus: The Interconnect Plan for Agentic SuperPoDs
Huawei presented Lingqu UnifiedBus as the shared interconnect and memory fabric for Agentic SuperPoDs and future clusters, claiming TB/s class bandwidth, roughly 2 µs round trip latency, and a path to million NPU syst... The architecture is meant to make CPUs, NPUs, memory, SSDs, storage, and network components work...
Published byEdited with GPT-5.6 TerraImages generated with GPT Image 2
Research answer
Key takeaways
- Huawei presented Lingqu UnifiedBus as the shared interconnect and memory fabric for Agentic SuperPoDs and future clusters, claiming TB/s class bandwidth, roughly 2 µs round trip latency, and a path to million NPU syst...
- The architecture is meant to make CPUs, NPUs, memory, SSDs, storage, and network components work as peers rather than as isolated infrastructure layers.
- At the smaller end of the portfolio, Huawei says the Atlas 650E pairs two eight NPU servers into a switchless 16 NPU full mesh configuration using UnifiedBus Link 2.0.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Huawei present at Huawei Connect 2026 about Lingqu UnifiedBus as the interconnect foundation for an Agentic SuperPoD and cluster co. Article summary: Huawei positioned Lingqu UnifiedBus as the common interconnect and memory fabric for an “Agentic SuperPoD” and a co-designed SuperPod/cluster architecture: one intended to scale compute, memory, and storage from cabinet . Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
Huawei’s central message at Huawei Connect 2026 was that AI infrastructure increasingly depends on the connections between processors—not just the processors themselves. Lingqu UnifiedBus was presented as a common interconnect and memory fabric for an “Agentic SuperPoD” and for a co-designed SuperPoD/cluster architecture that could extend from a cabinet to a future million-NPU SuperCluster. 1
7
The important qualification is that the bandwidth, latency, scale, and performance figures below are Huawei specifications, targets, or conference claims. The cited material does not establish independent end-to-end measurements at full SuperPoD or multi-cluster scale. 1
6
7
The bottlenecks Huawei says it is addressing
Huawei argued that scaling a conventional AI cluster can leave compute resources waiting on communications and data movement. It cited about 20% effective model-compute utilization for a 100,000-NPU cluster, while also identifying memory limits for very large-model training, changing CPU/NPU requirements in agentic workloads, and large KV-cache and storage demands. 1
Its proposed answer is to treat compute, memory, storage, and networking as parts of a more tightly coupled system. In that model, a SuperPoD is not simply a collection of servers connected by a conventional network; UnifiedBus is intended to provide the fabric through which system resources collaborate. 1
7
One protocol model and memory semantics
Huawei says UnifiedBus consolidates more than 10 interconnect protocols under a common protocol and memory-semantic model, with unified global memory addressing inside a SuperPoD. The company claims this changes interconnect bandwidth from roughly the 100-GB/s class to the TB/s class and reduces round-trip latency from about 7 microseconds to 2 microseconds. 1
The design is intended to directly connect CPUs, NPUs, memory, and SSDs, enabling peer-to-peer access rather than relying solely on centralized, host-mediated paths. Huawei also frames the architecture as a way to mix CPU and NPU resources more flexibly. 1
19
For data-heavy AI workloads, Huawei described pooled hybrid-media storage for activation caching. DDR can act as alternative NPU memory, according to the company, with the goal of lowering latency in search, recommendation, and advertising workloads while reducing per-NPU HBM requirements for very-large-model training. The platform also supports tiered Transformer acceleration and Attention–FFN disaggregation, which Huawei calls AF-separated deployment. 1
How the fabric is designed to scale
Huawei presented products for three levels of interconnection: within cabinets, between cabinets, and across clusters. The company says this family provides elastic scaling from a single cabinet to a million-NPU cluster. 7
- Within a cabinet: UnifiedBus LinkBlade is a cable-free component intended to reduce circuit losses. Huawei says a 4,096-NPU SuperPoD could eliminate approximately 196 km of copper cabling with this approach.
1
- Between cabinets: UnifiedBus LinkDevice is specified by Huawei with 176 ports at 1.6 Tbit/s per port, or 280 Tbit/s aggregate all-optical interconnect capacity. Huawei also claims round-trip latency as low as 2 microseconds.
- Across clusters: Xinghe UBG serves as the UnifiedBus gateway or switching layer for scale-out beyond a SuperPoD.
The optical element is central to the strategy. Huawei introduced the Atlas 960E SuperPoD with near-packaged optics (NPO), a configuration the company describes as the first NPO-based SuperPoD. 1
6
12
The proposed Agentic SuperPoD stack
Huawei described an Agentic SuperPoD that brings together compute, acceleration, context storage, and the cluster interconnect layer:
- Kunpeng 950 provides CPU-side and general-purpose processing in the unified system.
- Ascend 960 is the next-generation AI accelerator in the UnifiedBus portfolio. Huawei has said the chip delivers twice the performance of existing products; reporting on the announcement places the Ascend 960DT training version in Q1 2027 and the 960PR inference version in Q3 2027.
3
4
- OceanStor M900 is the memory-context storage component intended to retain and serve agent and model context.
- Xinghe UBG is the cluster-level switching or gateway layer for extending the fabric across clusters.
These components sit within Huawei’s broader 11-chip UnifiedBus-oriented portfolio spanning computing, interconnect, storage, and management functions. 1
5
Atlas 650E: the 16-NPU example
The Atlas 650E illustrates the same interconnect approach at a smaller deployment size. Huawei says each server uses eight Ascend 950DT NPUs and that two servers can form a 16-NPU full-mesh configuration through UnifiedBus Link 2.0, without a switch in that two-server design. 17
18
Huawei lists up to 12.4 PFLOPS at mxFP4 and up to 4.0 TB/s on-chip-memory read/write bandwidth for the server. Its Chinese product specification also lists 13.4 TB/s aggregate NPU interconnect bandwidth for the two-server scenario. These are vendor specifications, and workload-level results will depend on software, model architecture, precision, communication patterns, and system configuration. 17
18
What to watch next
Lingqu UnifiedBus is best understood as Huawei’s attempt to make the interconnect, memory model, storage tiering, and optical fabric a single AI-system design problem. The ambition is substantial: make resources act more like parts of one computer, then extend that model from a cabinet through a SuperPoD and eventually across a million-NPU SuperCluster. 7
23
The practical question is not whether the roadmap is bold—it is whether production systems can deliver the claimed latency, throughput, resource pooling, reliability, and usable model-training efficiency at scale. Huawei’s announcements establish the architecture and its stated targets; independent measurements will be needed to assess real-world results. 1
6
7
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Huawei’s Lingqu UnifiedBus: The Interconnect Plan for Agentic SuperPoDs"?
Huawei presented Lingqu UnifiedBus as the shared interconnect and memory fabric for Agentic SuperPoDs and future clusters, claiming TB/s class bandwidth, roughly 2 µs round trip latency, and a path to million NPU syst...
What are the key points to validate first?
Huawei presented Lingqu UnifiedBus as the shared interconnect and memory fabric for Agentic SuperPoDs and future clusters, claiming TB/s class bandwidth, roughly 2 µs round trip latency, and a path to million NPU syst... The architecture is meant to make CPUs, NPUs, memory, SSDs, storage, and network components work as peers rather than as isolated infrastructure layers.
What should I do next in practice?
At the smaller end of the portfolio, Huawei says the Atlas 650E pairs two eight NPU servers into a switchless 16 NPU full mesh configuration using UnifiedBus Link 2.0.
Sources
scmp.comHuawei unveils latest tech to boost AI power in push to break ...
thesun.myHuawei Pioneers a New Computing Architecture for the AI Era