Inside Huawei’s Breakthrough in AI Software
In an innovative leap set to improve China’s position in the semiconductor market, Huawei introduces a software solution in high-bandwidth memory (HBM) technology.
During the 2025 Financial AI Reasoning Application Landing and Development Forum in Shanghai, Huawei unveiled Unified Cache Manager (UCM), a software advancement poised to drastically enhance AI inference capabilities by expertly managing memory resources.
This development is part of the company’s effort to reduce reliance on high-end foreign memory chips.
AI efficiency through advanced memory management
The UCM software functions by allocating data according to the latency requirements of various memory types, such as HBM, standard dynamic random access memory (DRAM) and solid-state drives (SSDs).
The hierarchical strategy reallocates some of the more memory-intensive tasks away from traditional HBM consumption, thus improving the entire system's efficiency.
Zhou Yuefeng, the Vice-President and Head of Huawei’s data storage product line, notes that UCM has successfully reduced AI inference latency by up to 90% and increased system throughput by up to 22 times.
The strategic impact on China’s chip industry amid US sanctions
The domestic chip production capabilities in China are under strain due to US-imposed sanctions limiting access to advanced HBM chips, which are vital for AI workloads.
These sanctions specifically bar exports of advanced HBM versions like HBM2e, HBM3, HBM3e and HBM4 to Chinese firms, affecting third-party manufacturers depending on US technology.
This regulatory framework poses challenges for China in securing supplies from leading global vendors such as SK Hynix, Samsung Electronics and Micron Technology.
The role of collaboration and practical deployment
The UCM is not just a theoretical innovation, but has been tested in real-world scenarios, including customer voice analysis, marketing planning and office assistance applications with China UnionPay, a major Chinese financial services entity.
These trials underline the software’s efficacy in speeding up AI inference processes and lowering operational expenses for large-scale applications.
Crucially, Huawei plans to open-source the UCM platform in September 2025, which is expected to foster widespread adoption among developers and industry partners, promoting a collaborative ecosystem that goes beyond Huawei’s proprietary influence.
Huawei’s future prospects and industry importance
While UCM offers a substantial boost in efficiency and provides a partial solution to current hardware limitations, experts agree it cannot completely replace access to high-end HBM hardware.
However, it’s a testament to how Chinese technology companies are utilising software strength to navigate hardware shortages caused by geopolitical dynamics.
The demand for HBM is growing fast, driven by increased AI application demands, with global revenue from HBM chips expected to nearly double by 2025, reaching a projected US$98bn by 2030.
Huawei’s software innovation could strategically position both itself and China to better engage with this expanding market despite ongoing supply constraints.
Huawei’s Unified Cache Manager is an instrumental development that could alleviate China’s dependency on imported high-end HBM chips amidst tightening US sanctions.
By optimising latency and throughput across different memory types, UCM enables marked improvements in AI inference performance using less advanced hardware.
The forthcoming open-source release aims to bolster China’s quest for self-reliance in AI hardware and software networks while promoting cooperative innovation across the telecommunications and tech sectors.
