How Huawei’s Lingqu UnifiedBus Connects AI SuperPoDs and SuperClusters

Huawei presented Lingqu UnifiedBus at Huawei Connect 2026 as the connective layer of its Agentic AI SuperPoD and SuperCluster architecture. Rather than scaling only by adding accelerators, its proposal links compute, memory and storage across physical servers so they can work as a larger system. That makes the interconnect central to two problems: moving data quickly enough between processors and keeping enough context memory available for AI inference. 81921

One fabric for communication and memory access

Huawei says UnifiedBus brings more than ten interconnect protocols under a common protocol with memory semantics. Its SuperPoD design emphasizes peer-to-peer connections and memory access across physical servers, allowing CPUs, NPUs and other resources to exchange data without treating each server as an isolated compute island. The company reports a move from roughly 100-GB/s-class to TB/s-class interconnect bandwidth and a reduction in round-trip latency from about 7 µs to 2 µs. These are reported fabric figures, not measurements of application-level speedups. 271923

Fast communication also matters to Huawei’s proposed separation of Transformer Attention and feed-forward-network work: placing different operations on suitable resources is useful only if moving intermediate data between them does not erase the benefit. The available material describes that architectural intent but does not quantify its end-to-end gain. 78

How compute and context storage fit together

In Huawei’s broader stack, Kunpeng 950 provides general-purpose compute, Ascend 960 provides AI compute, and OceanStor M900 supplies context memory for inference. M900 uses UnifiedBus to pool KV-cache data across a tiered hierarchy extending from on-chip memory and DRAM to SSDs. Huawei presents this as a way to expand the working context available to SuperPoDs rather than relying solely on accelerator-local memory. 421

Those roles should not be conflated: faster processor-to-processor communication addresses one bottleneck, while a shared, tiered KV cache addresses capacity. Accessing storage-backed context still depends on the performance of the complete path, not just a headline interconnect latency. 721

From cabinets to optical SuperPoDs

Huawei describes UnifiedBus-powered equipment for connections within cabinets, between cabinets and across clusters. One reported interconnect device has 176 ports rated at 1.6 Tbit/s each, or approximately 280 Tbit/s of aggregate port capacity. That switch specification is distinct from the bandwidth available to an individual processor or an AI workload. 719

For larger systems, Huawei pairs the fabric with optical networking. Its Atlas 960E SuperPoD incorporates near-packaged optics, while Huawei says an all-optical UnifiedBus configuration of the TaiShan 950 SuperPoD supports up to 4,096 NPUs. Its stated longer-term architecture extends toward million-NPU clusters; that is a scaling goal, not a published benchmark of a million-NPU deployment. 2519

What remains unproven

Huawei positions interconnect equipment, including the UBG layer described in the conference coverage, as part of the cabinet-to-cluster design. The supplied evidence does not establish a precise role or deployment status for the Atlas 650E, or substantiate a numerical copper-saving claim. It also provides no independent full-cluster benchmark demonstrating that the cited latency, bandwidth and scale figures translate into sustained training or inference gains. For now, UnifiedBus is best understood as Huawei’s integrated architecture and a set of vendor-reported specifications—not a verified end-to-end performance result. 7819