Redazione RHC : 14 August 2025 09:17
Huawei has unveiled a new software tool that speeds up large AI models by distributing data more efficiently across different types of memory. This could help China reduce its dependence on expensive, high-bandwidth memory chips, now almost entirely controlled by foreign suppliers.
The development is called Unified Cache Manager (UCM). This is an algorithm that places data by taking into account the latency of different memory levels, from high-speed HBM to standard RAM and SSD. According to Zhou Yuefeng, Huawei’s vice president of storage systems, the technology has reduced inference latency by 90% and increased system throughput by 22 times. These results were achieved during tests the company presented at the AI Industry Applications Forum in Shanghai.
With limited access to modern hardware, Chinese companies are increasingly turning to alternative software solutions. Huawei’s approach is largely similar to what was previously demonstrated by the startup DeepSeek: it trained large-scale AI models with limited computing resources, achieving competitive results.
Huawei promised to open source UCM as early as September, first to its online developer community and then to the entire industry. The initiative aims to reduce China’s dependence on foreign chips, particularly those from SK Hynix, Samsung, and Micron, which currently dominate the HBM market.
HBM technology is a multilayer memory with high bandwidth and low latency, crucial for modern AI chips. According to consulting firm Yole Group, the global HBM market will reach $34 billion in 2025 and could grow to $98 billion by 2030. This rapid growth is explained by the boom in generative artificial intelligence.
Since last year, the United States has restricted exports of advanced HBM chips to China in an attempt to slow the development of the country’s semiconductor industry. This decision is part of a broader strategy pursued by the Biden administration. In response, Beijing has accelerated its push for technological independence and begun actively developing its own memory production. Major players include Yangtze Memory, Changxin Memory, and Tongfu Microelectronics.
However, Chinese manufacturers are still struggling to keep up with foreign competitors. While the country is still preparing the infrastructure for the production of the second-generation HBM (HBM2), SK Hynix is already marketing the HBM4, a chip with even higher speeds.
Restrictions on the export of equipment required for the production of modern microcircuits also continue to pose a significant obstacle, seriously hampering the development of Chinese microelectronics.
Against this backdrop, the fate of the H20, a simplified Nvidia AI chip designed specifically for the Chinese market, has become crucial. According to the Financial Times, Beijing has offered to ease export barriers for HBM as part of a potential trade deal with the United States, but there are no signs of a rapprochement between the two sides.
Despite sanctions, Huawei remains a key player in the Chinese AI chip market thanks to its Ascend line. Earlier this year, the company unveiled the CloudMatrix 384, a “supernode” computing system capable of competing with Nvidia’s GB200 NVL72 systems.
In early August, Huawei also announced its intention to open source its CANN (Compute Architecture for Neural Networks) framework, centered around Ascend processors. This represents a direct challenge to Nvidia’s CUDA ecosystem, which has long been the de facto standard for training and running AI models.
Redazione