Huawei Unveils New Global Compute Options

The core of Huawei's announcement is the UnifiedBus interconnect protocol, which underpins the SuperPoD architecture. This technology enables high-bandwidth, low-latency connections between compute and storage, allowing multiple physical servers to function as a single logical unit. Huawei is also opening up its SuperPoD hardware and the UnifiedBus protocol to partners to foster broader industry collaboration. The Atlas 950 SuperPoD is engineered for ultra-large-scale AI tasks, connecting up to 8,192 of Huawei's Ascend NPUs. It delivers 8 exaFLOPS of FP8 and 16 exaFLOPS of FP4 computing power, with an interconnect bandwidth of 16.3 PB/s. A single Atlas 950 SuperPoD requires 160 cabinets and covers approximately 1,000 square meters. For more general-purpose computing, the TaiShan 950 SuperPoD is the first of its kind in the industry. It is built on the Kunpeng 950 processor and is designed to replace traditional mainframes and database servers, showing a 2.9x performance improvement for GaussDB. The system also improves memory utilization in virtualized settings by 20% and reduces real-time data processing time in Spark workloads by 30%. Huawei's strategy extends to massive cluster deployments, with the Atlas 950 SuperCluster integrating 64 Atlas 950 SuperPoDs. This configuration comprises over 520,000 Ascend 950DT chips across more than 10,000 cabinets, delivering a total of 524 FP8 exaFLOPS. The SuperCluster is positioned to compete with upcoming platforms from Nvidia. This aggressive hardware push places Huawei in direct competition with established players like Nvidia, which currently dominates the AI chip market. The industry is seeing a broader trend of companies like Google, AMD, and Intel developing their own specialized AI hardware to create comprehensive ecosystems beyond just chip manufacturing. For supply chain and retail operations, Huawei offers cloud solutions that leverage this new computing power. Their Smart Logistics and Warehousing Solution uses a smart operations cloud connected to intelligent systems in logistics parks, aiming to digitize and automate the entire process from warehousing to distribution. These solutions have been deployed in over 100 ports and 200 logistics enterprises. To ensure the stability of these large-scale clusters, Huawei also announced its MindOps Intelligent Computing O&M Solution. This platform uses a 7-layer digital twin and AI models to predict and proactively repair faults, aiming to increase the availability of computing clusters from the industry average of 90% to 99.9%. The company has a roadmap extending to the Atlas 960 SuperPoD and SuperCluster, expected in the fourth quarter of 2027. The Atlas 960 SuperPoD will feature 15,488 Ascend 960 chips, and the SuperCluster will scale to over one million NPUs.