(Corrects media packaging keyword)
Oct 28 (Reuters) - China's CXMT plans to start production of HBM3E, an advanced type of high-bandwidth memory chip used with AI processors, despite as little as 20% of the chips it makes being usable, two people familiar with the matter said.
The decision highlights China's intense shortage of the AI memory chips, which are essential to a government-backed push to develop domestic alternatives to Nvidia's AI processors at scale amid US curbs on access to foreign-made
technology.
ChangXin Memory Technologies — China's largest maker of dynamic random-access memory (DRAM), the conventional memory used in computers and servers — has sent HBM3E samples to several domestic AI chip designers, the people said. It expects to begin limited production over the next two months, one person said.
Yields, or the share of chips coming off a production line that meet specifications, are only about 20% to 30%, compared with yields of around 70% at South Korea's SK Hynix, the market leader, they said.
The people declined to be identified because the plans are not public.
CXMT did not respond to a request for comment. SK Hynix did not immediately respond to a request for comment on the yield estimate.
The Information first reported CXMT's HBM3E plans. Reuters is reporting details about the product yields and delivery timeline for the first time.
High-bandwidth memory (HBM) is made by vertically stacking memory chips so they can exchange data rapidly with AI processors. It is a crucial component in the specialized chips used to train and run large language models.
Washington tightened export controls on certain advanced HBM products bound for China in December 2024. Chinese chipmakers have increasingly sought the memory through unofficial channels, where it can cost several times more, the people said.
CXMT is pressing ahead despite low yields because Chinese AI chipmakers have struggled to obtain enough advanced memory to increase output, the people said. Low yields make production expensive because a large share of chips must be discarded or reworked.
Reuters reported last month that Huawei, Cambricon, MetaX and Iluvatar CoreX had raised prices for current and next-generation AI chips by 20% to 50% as memory costs complicated efforts to ramp up production.
CXMT's HBM3E is expected to cost more than comparable products from overseas rivals, one of the people said. The company began developing the technology internally as early as 2024, the people added.
GENERATIONS BEHIND
In testing, CXMT's HBM3E performance was closer to SK Hynix's previous-generation HBM3 than to the South Korean company's HBM3E, one of the people said.
CXMT has sent customers two HBM3E versions, the other person said: a standard design and an updated version that includes changes intended to improve yields.
The difficulty of producing HBM illustrates the gap between CXMT and the industry's established leaders — SK Hynix, Samsung Electronics and Micron Technology — which dominate global supply.
SK Hynix held 50% of global HBM revenue in the second quarter, according to Counterpoint Research, followed by Samsung with 33% and Micron with 18%. Most revenue came from HBM3E, the research firm said.
CXMT held a 10% share of global DRAM revenue in the quarter, making it the world's fourth-largest supplier of conventional memory chips, Counterpoint said. But US export controls limit its access to some of the most advanced chipmaking equipment, making it harder to close the gap with global rivals.
Samsung said in July that it had expanded HBM4 sales and shipped HBM4E samples to major customers, putting CXMT at least one generation behind in commercial HBM production.
(Reporting by Reuters Staff; Editing by Eduardo Baptista and Kevin Buckland)
