Two days ago, we shared exciting news: the Chinese state media outlet, Securities Times, revealed that Huawei was set to unveil a groundbreaking technological advancement aimed at reducing China’s dependence on foreign HBM memory chips ( High Bandwidth Memory ). According to this source, Huawei was expected to make an official announcement during the Applications Forum and Development of Financial Reasoning 2025 in Shanghai.
Huawei has indeed fulfilled its promise, albeit not in the manner many anticipated. It is vital to recognize that Chinese chip manufacturers are currently unable to produce solutions that rival the cutting-edge memories developed by South Korean giants Samsung and SK Hynix, as well as American firm Micron Technology. Graphics Processing Units (GPUs) require robust HBM memory chips to function optimally, which implies that their performance is significantly influenced by these memory components.
According to the editors of SEMIANALYSIS, the total bandwidth of HBM3 memory chips used in conjunction with some of the most advanced GPUs from NVIDIA and AMD exceeds 819 GB/s . In contrast, DDR5 and GDDR6X memories only reach comparatively modest speeds of 70.4 GB/s and 96 GB/s , respectively. Future iterations, such as HBM3E and HBM4, promise even greater capabilities. Currently, Chinese manufacturers do not produce HBM memory of this caliber, but Huawei’s recent announcement aims to change this landscape.
An Algorithm Expressly Designed to Accelerate Inference in AI
Leaks surfaced just 48 hours before the announcement suggested that Huawei might reveal a cutting-edge packaging technology that could rival those employed by SK Hynix, Samsung, or Micron in the production of their HBM3 and HBM3E memories. The manufacturing process for these integrated circuits is complex, necessitating multiple stacked DRAM chips and an interface connecting the Extended Processing Unit (XPU) with the densely packed HBM memory chips. For instance, the HBM3E stack utilizes over 1,000 drivers to link the XPU and HBM memory.
<img alt="When the stars were formed, it has always been one of the greatest mysteries in the universe. And we are closer to solving it" width="375" height="142" src="https://i.blogs.es/32b524/estrellas-ap/375_142.jpeg"/>According to Huawei, the UCM algorithm is capable of drastically accelerating inference in the great AI models
However, Huawei ultimately introduced a different technology: an advanced algorithm known as UCM (Unified Cache Manager) . According to the company, this algorithm can drastically accelerate inference in large artificial intelligence models. Notably, inference refers to the computational process employed by language models to generate responses corresponding to the requests they receive.
To optimize its functionality, the UCM algorithm employs a clever strategy: it determines the appropriate memory type for data storage based on latency requirements. In practical terms, this algorithm functions as a massive cache, ensuring that each piece of data is directed to the suitable memory type, including HBM3, to minimize latency during inference. If a specific data point is frequently accessed, it will be stored in a high-speed memory format, such as HBM3. According to Huawei, this technology can reduce inference latency by an impressive 90% . Intriguingly, the company plans to release the UCM Source Open Algorithm in September.
For additional insights, refer to SCMP.
As the tech industry grapples with the implications of Huawei’s announcement, it remains to be seen how this algorithm will affect the balance of power in AI technology and memory manufacturing. The move signifies an ambitious attempt to lessen China’s reliance on foreign technology while enhancing its capability to innovate independently. This shift could reshape the future of AI development within the country and foster a new wave of competitive advancements in the memory chip market.