Huawei Debuts UnifiedBus Architecture to Scale AI SuperClusters

Traditional computing clusters often struggle as they scale, with nearly 80% of capacity lost to idle time during data transfers. Huawei’s UnifiedBus addresses this by consolidating over ten interconnect protocols into a single framework. This integration boosts bandwidth into the terabyte-per-second range while slashing round-trip latency from seven microseconds to just two, enabling unified global memory addressing within SuperPoDs.

Beyond raw speed, the architecture enables heterogeneous collaboration by directly linking CPUs, NPUs, memory, and SSDs. This peer-to-peer access facilitates tiered hardware acceleration, specifically for Transformer-based models, and reduces the HBM capacity requirement per NPU. Hardware components—including the LinkBlade for cable-free cabinet interconnects and the UBG switch—allow for elastic scaling from single-cabinet appliances to million-NPU SuperClusters.

These advancements are packaged into the new agentic AI SuperCluster, which combines TaiShan 950 and Atlas 960 hardware with OceanStor M900 storage. Huawei is also extending this technology to smaller compute appliances, providing enterprises with the ability to run massive models locally. To support the wider ecosystem, the company is integrating its Ascend platform with open-source frameworks like DeepSeek Harness and OpenCode to streamline intelligent management and inference acceleration.