Huawei unveiled a new UnifiedBus computing architecture for SuperPods and clusters at HUAWEI CONNECT 2026.

At HUAWEI CONNECT 2026, Yang Chaobin, Executive Director of the Board and CEO of Huawei's ICT Business Group, unveiled the UnifiedBus interconnect architecture and a suite of computing products built on it. The announcement emphasized Huawei's collaboration with open-source communities to enable application innovation across customers, partners, and developers.

Traditional computing architectures suffer from declining resource utilization as clusters scale. A 100,000-NPU cluster achieves only 20% computing capacity utilization, with vast amounts of compute sitting idle during data communications. Training 10 trillion-parameter models compounds the problem: the massive intermediate data generated far exceeds the memory capacity of a single accelerator. Interconnection and communications between cluster components have become a critical performance bottleneck.

Huawei's solution is UnifiedBus, an interconnect technology designed to facilitate collaboration between clusters and SuperPoDs while supporting emerging demand from agentic applications. The architecture delivers four key capabilities:

*Unified protocol and memory semantics* combine more than ten interconnect protocols under a single UnifiedBus standard, increasing interconnect bandwidth from 100 GB/s to TB/s and reducing round-trip latency from 7 microseconds to 2 microseconds. The protocol enables unified global memory addressing within SuperPoDs.

*Heterogeneous compute collaboration* directly interconnects CPUs, NPUs, memory, and solid-state drives to enable decentralized peer-to-peer access and flexible CPU-NPU mixing. The architecture supports tiered hardware acceleration for Transformers through Attention-FFN disaggregation.

*Tiered storage with global pooling* uses hybrid-media resource pooling to cache activations. DDR memory serves as alternative memory for NPUs, reducing latency for search, recommendation, and advertising services, and doubling the performance of vector retrieval for 100 billion-level datasets with thousands of dimensions. This reduces per-NPU HBM capacity requirements for training 10 trillion-parameter models and improves cluster model FLOPs utilization.

*Optoelectronic interconnect and flexible networking* establishes a global ultra-high bandwidth, ultra-low latency data highway enabling elastic compute scaling.

Huawei introduced three UnifiedBus interconnect products covering within-cabinet, between-cabinet, and cross-cluster connections, enabling scaling from a single cabinet to million-NPU clusters. The LinkBlade eliminates circuit losses with a cable-free design, reducing copper cabling by approximately 196 kilometers within a 4,096-NPU SuperPoD. The LinkDevice, supporting 176 ports at 1.6 Tbit/s per port, delivers 280 Tbit/s all-optical interconnection with 2 microsecond RTT latency—the industry's highest-bandwidth, highest-port-count high-speed bus-protocol interconnect device. The UBG switch offers up to 1,024 radix fan-out capacity, supporting million-NPU SuperClusters and evolution toward tens of trillions of model parameters.

The UnifiedBus-powered agentic AI SuperCluster combines TaiShan 950 and Atlas 960 SuperPoDs, OceanStor M900 memory storage, and a Xinghe UBG switch, enabling diverse compute collaboration and tiered resource pooling. Beyond SuperPoDs, Huawei has extended UnifiedBus technology to compute appliances for small and medium-sized enterprises, offering one to eight NPU configurations with Kunpeng and Ascend modules for running trillion-parameter models locally.

Ascend is collaborating with open-source frameworks including DeepSeek Harness, OpenCode, and openJiuwen to develop fully open-source intelligent management, inference acceleration, and security framework agent plugins.

"Huawei is committed to innovating at the system level, building a computing-foundation product portfolio, and going open source and open system," Yang said. "We will continue working with customers, partners, and developers to help this ecosystem thrive and offer the world a new option for compute."