Huawei and Hubei Mobile complete China's first AI inference acceleration test, boosting throughput by 372%

At the 2026 MWC Shanghai event, Huawei and China Mobile Hubei announced the completion of the first operator AI inference acceleration test in China. The test uses Huawei's OceanStor A800 storage and Ascend A3 super node architecture with UCM technology.

For long-sequence AI inference, UCM provides PB-level KV Cache via external storage, breaking the limits of traditional high-bandwidth memory. This boosts token throughput by up to 372%. Tests covered models like MiniMax M2.5 and GLM-5.1, simulating sequences from 8K to 190K.

Results show for MiniMax M2.5, first-token delay improved 26%-62%, and output efficiency increased 78% at 128K sequences. For GLM-5.1, first-token delay improved 51%-93%, and output efficiency increased 372% at 128K sequences. Huawei says the advantage grows with longer contexts, solving KV Cache bottlenecks.