China's largest DRAM manufacturer, CXMT (Changxin Memory Technologies), plans to start mass production of high-bandwidth memory (HBM) for artificial intelligence (AI) applications, despite current yield rates of only 20% to 30%.
According to Reuters on October 28, citing multiple sources familiar with the matter, CXMT has delivered HBM3E samples to several AI semiconductor design firms in China and intends to begin limited production within the next two months. HBM3E is advanced memory used for training and inference in AI accelerators, featuring a structure that vertically stacks multiple memory chips to enhance data processing speeds, playing a crucial role in powering large-scale AI models.
Reuters reported that CXMT's HBM3E yield is between 20% and 30%, a significant gap compared to SK Hynix's yield of approximately 70%. The lower yield results in higher production costs, and CXMT's products are expected to be priced higher than those of overseas competitors. Nevertheless, CXMT is accelerating its production efforts due to a shortage of advanced memory among Chinese AI semiconductor companies. The United States has tightened export controls on advanced HBM products to China, effective December 2024. Chinese firms are reportedly sourcing memory through unofficial channels, leading to increased price pressures.
In fact, Chinese AI chip companies such as Huawei, Cambrian, MetaX, and Illuvatar CoreX have raised prices for existing and next-generation AI chips by 20% to 50% in response to rising memory costs and supply shortages.
Additionally, CXMT's HBM3E is said to perform closer to the previous generation, HBM3, rather than SK Hynix's HBM3E in performance tests. CXMT has also reportedly delivered modified designs to clients to improve yield rates.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.